# auditing

Published articles for auditing.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits

DevFeed: [LLM GerryMandering: Separating Interactions from Additive Effects in LLM Audits](<https://devfeed.tech/articles/llm-gerrymandering-separating-interactions-from-additive-effects-in-llm-audits-145681.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26538>)

Author: George Rahul, Ronit Jahilal, Bhaskara Hanuma Vedula, Ponnurangam Kumaraguru, Pushkar Shukla

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [audit](<https://devfeed.tech/topics/audit.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [arxiv](<https://devfeed.tech/tags/arxiv.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [audits](<https://devfeed.tech/tags/audits.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bias](<https://devfeed.tech/tags/bias.md>), [challenge](<https://devfeed.tech/tags/challenge.md>), [class](<https://devfeed.tech/tags/class.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [effects](<https://devfeed.tech/tags/effects.md>), [gender](<https://devfeed.tech/tags/gender.md>), [gerrymandering](<https://devfeed.tech/tags/gerrymandering.md>), [identity](<https://devfeed.tech/tags/identity.md>), [independent](<https://devfeed.tech/tags/independent.md>), [indian](<https://devfeed.tech/tags/indian.md>), [interaction](<https://devfeed.tech/tags/interaction.md>), [language](<https://devfeed.tech/tags/language.md>), [llm](<https://devfeed.tech/tags/llm.md>)

### AI overview

The standard subgroup criterion can label independent attribute effects as intersectional bias. Across 150 synthetic datasets with no interactions, it produced 66.1 false subgroup flags per dataset on average; a repaired version recovered 32% of planted interactions. ADDIS fits an additive model across attributes, measures each subgroup's Additive Residual Distance from its prediction, and tests it with a Freedman-Lane null. On the same benchmark it averaged 0.09 false flags per dataset and recovered 84% of planted interactions. In audits of 13 open language models using 10,000 synthetic Indian personas, the standard criterion flagged 7.1% of 738 tested subgroups on average, compared with 3.0% for ADDIS. Gender identity crossed with sexual orientation was the largest pairing in every model. Counterfactual personas also revealed a caste-class interaction in all but one model, which the observational test lacked data to detect.

### Source excerpt

arXiv:2611.26538v1 Announce Type: new Abstract: Intersectional bias arises when a model treats people who hold several marginalised identities at once differently from what their separate identities would predict. Auditing such intersectional subgroups remains a challenge: their number grows quadratically, and there is no agreed baseline for what independent attributes would predict. Without such a baseline, many existing audits flag a subgroup when it exceeds each attribute alone, a condition that attributes acting independently already satisfy. We test this standard criterion on 150 synthetic datasets built with no interaction. With the shuffle null, it flags 66.1 false subgroups per dataset on average, and at least one in every dataset. When we plant interactions of known size into these datasets, a repaired null lets the criterion recover only 32% of them. We introduce ADDIS (Additive Intersectional Screen), which fits an additive model over all attributes, scores each subgroup by its Additive Residual Distance (ARD) from that model's prediction, and tests it against a Freedman-Lane null. On our synthetic benchmark, ADDIS flags 0.09 false subgroups per dataset and recovers 84% of the same planted interactions. We then build IndiPersona, 10,000 synthetic Indian personas with six identity attributes, and audit 13 open language models. The standard criterion flags 7.1% of the 738 tested subgroups on average; ADDIS flags 3.0%. Gender identity crossed with sexual orientation is the largest pairing in every model, and counterfactual personas confirm it and additionally reveal a caste x class interaction in all but one model that the observational test lacked the data to detect. Much of what audits report as intersectional bias can thus arise from independent single-attribute effects alone, yet a small core of interactions survives.

## A Statistical Framework for Training Data Auditing of Large Language Models

DevFeed: [A Statistical Framework for Training Data Auditing of Large Language Models](<https://devfeed.tech/articles/a-statistical-framework-for-training-data-auditing-of-large-language-models-145158.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26264>)

Author: Shengjie Niu, Yeheng Ge, Jian Huang

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark-contamination](<https://devfeed.tech/tags/benchmark-contamination.md>), [books](<https://devfeed.tech/tags/books.md>), [case-studies](<https://devfeed.tech/tags/case-studies.md>), [copyright](<https://devfeed.tech/tags/copyright.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data-collection](<https://devfeed.tech/tags/data-collection.md>), [error-control](<https://devfeed.tech/tags/error-control.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [framework](<https://devfeed.tech/tags/framework.md>), [global](<https://devfeed.tech/tags/global.md>), [inference](<https://devfeed.tech/tags/inference.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [medical](<https://devfeed.tech/tags/medical.md>), [multiple-testing](<https://devfeed.tech/tags/multiple-testing.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [stat-ap](<https://devfeed.tech/tags/stat-ap.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [training-data](<https://devfeed.tech/tags/training-data.md>)

### AI overview

The framework audits whether documents and collections were used to train large language models, while accounting for membership signals concentrated in only a few passages. It combines complementary chunk-level evidence with Cauchy combination to score documents, then uses calibrated p-values for collection-level auditing and multiple-testing procedures to identify suspicious subsets. Under stated exchangeability conditions, the authors establish finite-sample validity of document-level p-values, type-I error control for global audits, and false discovery rate control for subset identification. Case studies cover BookTection, Pile-arXiv, Cosmopedia, and HealthCareMagic, spanning pre-training and fine-tuning data.

### Source excerpt

arXiv:2611.26264v1 Announce Type: cross Abstract: Training-data auditing asks whether data of interest were used to train a large language model (LLM), a question central to privacy auditing, copyright protection, and benchmark-contamination detection. These concerns are particularly relevant to the use of copyrighted books and scientific papers in pre-training, as well as synthetic text and medical question-answer records in fine-tuning. In practice, training membership signals may be concentrated in only a few passages of a long document, making it important to aggregate evidence within each document without diluting these localized signals. Beyond individual documents, auditing a target collection requires an overall assessment based on the document-level evidence. In this paper, we propose a unified statistical framework for document-level membership inference, target-collection global auditing, and suspicious-subset identification. We integrate complementary chunk-level membership evidence with Cauchy combination to construct document scores for document-level membership inference. The document scores and their calibrated p-values further enable an overall assessment of training-data collection use and the identification of suspicious documents through multiple testing. Under the stated exchangeability conditions, we establish finite-sample validity of document-level p-values, type-I error control for global auditing, and false discovery rate control for subset identification. Case studies on BookTection, Pile-arXiv, Cosmopedia, and HealthCareMagic demonstrate the effectiveness and reliability of the proposed framework in auditing pre-training and fine-tuning data.

## Behavioral Integrity Verification for AI Agent Skills

DevFeed: [Behavioral Integrity Verification for AI Agent Skills](<https://devfeed.tech/articles/behavioral-integrity-verification-for-ai-agent-skills-145331.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.11770>)

Author: Yuhao Wu, Tung-Ling Li, Hongliang Liu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Agent Skills](<https://devfeed.tech/topics/agent-skills.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [threat detection](<https://devfeed.tech/topics/threat-detection.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-skills](<https://devfeed.tech/tags/agent-skills.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-skills](<https://devfeed.tech/tags/ai-agent-skills.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-chains](<https://devfeed.tech/tags/attack-chains.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [capability](<https://devfeed.tech/tags/capability.md>), [categories](<https://devfeed.tech/tags/categories.md>), [cause](<https://devfeed.tech/tags/cause.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-sy](<https://devfeed.tech/tags/cs-sy.md>), [eess-sy](<https://devfeed.tech/tags/eess-sy.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [skills](<https://devfeed.tech/tags/skills.md>), [verification](<https://devfeed.tech/tags/verification.md>)

### AI overview

Behavioral integrity verification (BIV) checks whether an AI agent skill's declared capabilities match what its code and instructions enable. It pairs deterministic code analysis with LLM-assisted capability extraction over a shared taxonomy, producing evidence for deviation analysis, root-cause classification, and malicious-skill detection. In 49,943 OpenClaw skills, 80.0% had at least one classified deviation; the authors attributed 81.1% of deviations to developer oversight and 18.9% to adversarial intent, and predicted multi-stage attack chains in 5.0% of skills. On a separate 906-skill benchmark, BIV achieved an F1 of 0.949, outperforming the reported rule-based and single-pass LLM baselines. These are results reported in the supplied abstract.

### Source excerpt

arXiv:2605.11770v2 Announce Type: replace-cross Abstract: Agent skills extend LLM agents with privileged third-party capabilities such as filesystem access, credentials, network calls, and shell execution. Existing safety work catches malicious prompts and risky runtime actions, but the skill artifact itself goes unverified. We formalize this as the behavioral integrity verification (BIV) problem: a typed set comparison between declared and actual capabilities over a shared taxonomy that bridges code, instructions, and metadata. The BIV framework instantiates this comparison by pairing deterministic code analysis with LLM-assisted capability extraction. The resulting structured evidence supports three downstream analyses: deviation taxonomy, root-cause classification, and malicious-skill detection. On 49,943 skills from the OpenClaw registry, the deviation taxonomy reveals a pervasive description-implementation gap: 80.0% of skills have at least one deviation in the taxonomy, with four compound-threat categories surfaced. Root-cause classification finds that deviations are mostly oversight, not malice: 81.1% trace to developer oversight and 18.9% to adversarial intent, with 5.0% of skills carrying predicted multi-stage attack chains. On a 906-skill malicious-skill detection benchmark, BIV reaches an F1 of 0.949, outperforming state-of-the-art rule-based and single-pass LLM baselines. These results demonstrate behavioral integrity auditing for agent skills at scale.

## Auditing Learned Dynamics in Weather Models

DevFeed: [Auditing Learned Dynamics in Weather Models](<https://devfeed.tech/articles/auditing-learned-dynamics-in-weather-models-145011.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25119>)

Author: Skylar Gu, Christina Last

Published: 2026-11-20T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Climate / Weather / Ocean Modeling](<https://devfeed.tech/topics/climate-weather-ocean-modeling.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Feature Engineering](<https://devfeed.tech/topics/feature-engineering.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [accessible](<https://devfeed.tech/tags/accessible.md>), [ai](<https://devfeed.tech/tags/ai.md>), [also](<https://devfeed.tech/tags/also.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [association](<https://devfeed.tech/tags/association.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [can](<https://devfeed.tech/tags/can.md>), [central](<https://devfeed.tech/tags/central.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deepmind](<https://devfeed.tech/tags/deepmind.md>), [error](<https://devfeed.tech/tags/error.md>), [event](<https://devfeed.tech/tags/event.md>), [experimental](<https://devfeed.tech/tags/experimental.md>), [fall](<https://devfeed.tech/tags/fall.md>), [feature](<https://devfeed.tech/tags/feature.md>), [features](<https://devfeed.tech/tags/features.md>), [models](<https://devfeed.tech/tags/models.md>), [weather](<https://devfeed.tech/tags/weather.md>)

### AI overview

Sparse autoencoder features in GraphCast can have causal effects even without conventional physical meaning, and physically associated features can be steered to alter forecast outcomes. Ablating a feature tied to the model's computational mesh increased global geopotential height error. Suppressing convection- and low-level-vorticity-associated features weakened predicted tropical-cyclone deepening; amplifying them reduced intensity error. Scaling convection features yielded a median 51% error reduction across seven storms. The abstract presents these features as experimental coordinates for auditing and intervening on learned weather dynamics; the evidence provided is an abstract, not the full paper.

### Source excerpt

arXiv:2611.25119v1 Announce Type: new Abstract: Learned weather models are experimentally accessible computational systems whose internal representations can be inspected and manipulated. Using sparse autoencoder features from GraphCast, DeepMind's autoregressive weather forecasting model, we test whether causal relevance requires physical semantics, a question that matters for auditing whether AI forecasters used in high-assurance settings rely on reliable mechanisms. We also look at whether physically associated latent directions can be steered to improve extreme-event forecasts, where interventions are applied to the model's internal activations during the rollout. We find that ablating a feature locked to GraphCast's computational mesh increases global geopotential height error despite lacking conventional physical semantics. Second, suppressing convection- and low-level-vorticity-associated features weakens predicted tropical-cyclone deepening (the fall in central pressure during intensification), while amplifying them reduces intensity error, with amplification of convection features through feature scaling achieving a median of 51\% error reduction over seven storms. Together, these results separate interpretable, physical association from causal utility and establish sparse features as experimental coordinates for auditing and intervening on learned weather dynamics.

## Visual Attribution Limits and Graded Provenance Auditing in Commercial Video-Generation APIs

DevFeed: [Visual Attribution Limits and Graded Provenance Auditing in Commercial Video-Generation APIs](<https://devfeed.tech/articles/visual-attribution-limits-and-graded-provenance-auditing-in-commercial-video-generation-apis-146060.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26508>)

Author: Haocheng Mei, Yuchuan Tian, Yingte Shu, Chao Xu, Kai Han, Hanting Chen, Yunhe Wang

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [AI software supply chain security](<https://devfeed.tech/topics/ai-software-supply-chain-security.md>), [software artifact signing](<https://devfeed.tech/topics/software-artifact-signing.md>), [API Monetization](<https://devfeed.tech/topics/api-monetization.md>)

Tags: [apis](<https://devfeed.tech/tags/apis.md>), [attribution](<https://devfeed.tech/tags/attribution.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [generation](<https://devfeed.tech/tags/generation.md>), [limits](<https://devfeed.tech/tags/limits.md>), [provenance](<https://devfeed.tech/tags/provenance.md>), [video](<https://devfeed.tech/tags/video.md>), [video-generation](<https://devfeed.tech/tags/video-generation.md>)

### AI overview

The audit finds that delivered video files alone provide weak evidence of origin: six-model container and bitstream accuracy dropped from 83.9% on raw files to 45.3% after uniform re-encoding, while frame-rate subsampling from 24 to 3 fps raised pooled D3 AUROC from 0.845 to 0.988. The authors reserve accusations for authenticated provenance contradictions or duplicate signed instances under fresh-generation contracts, excluding retries and idempotent returns. Unverifiable deliveries from signed endpoints prompt requests for original files under pass-through clauses; archive-reuse and delivery-parameter screens support review only. Across 19 endpoints, only BytePlus authenticated tier claims under the pinned policy, covering 187 of 187 audited deliveries across three Seedance tiers. Processed Seedance files became unverifiable after tested remuxing and re-encoding. The audit flagged all 100 tested raw 720p signed tier swaps, with no observed false flags in the tested honest pools; these results are limited to the tested cases.

### Source excerpt

arXiv:2611.26508v1 Announce Type: new Abstract: Commercial video-generation APIs bill by model and tier, while intermediaries could substitute cheaper models, lower tiers, or cached clips. What can a delivered file establish about its origin? We combine measurements of passive attribution and cryptographic provenance into a graded audit. We test attribution probes and zero-shot detectors under re-encoding, downscaling, and frame-rate changes, and survey C2PA and GB 45438 metadata across 19 commercial endpoints. Only authenticated provenance supports accusations: tier contradictions or duplicate signed instances under fresh-generation contracts, excluding retries and idempotent returns. Unverifiable signed-endpoint deliveries trigger original-file demands under pass-through clauses; archive-reuse and delivery-parameter screens support review only. Six-model container and bitstream accuracy falls from 83.9% on raw files to 45.3% after uniform re-encoding, using different feature sets. Subsampling real references from 24 to 3 fps raises pooled D3 AUROC from 0.845 to 0.988. Only BytePlus authenticates tier claims under our pinned policy, including 187/187 audited deliveries across three Seedance tiers; tested remuxing and re-encoding leave processed Seedance files unverifiable. The audit identifies authenticated contradictions in 100/100 raw 720p signed tier-swap cases, with zero observed false flags on the tested honest pools.

## Diagnosis is not a policy: a state-transition audit of caoability-conditioned test-time inference

DevFeed: [Diagnosis is not a policy: a state-transition audit of caoability-conditioned test-time inference](<https://devfeed.tech/articles/diagnosis-is-not-a-policy-a-state-transition-audit-of-caoability-conditioned-test-time-inference-145537.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25566>)

Author: Yuchong Hu, Zhao Zhang, Kainan Liu, Haoyu Wang, Siyan Zhang, Jing Xiao

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benefit](<https://devfeed.tech/tags/benefit.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [configuration](<https://devfeed.tech/tags/configuration.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [context](<https://devfeed.tech/tags/context.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [diagnosis](<https://devfeed.tech/tags/diagnosis.md>), [diagnostic](<https://devfeed.tech/tags/diagnostic.md>), [error](<https://devfeed.tech/tags/error.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [financial](<https://devfeed.tech/tags/financial.md>), [headroom](<https://devfeed.tech/tags/headroom.md>), [inference](<https://devfeed.tech/tags/inference.md>), [information](<https://devfeed.tech/tags/information.md>), [policy](<https://devfeed.tech/tags/policy.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

TERRA is an evaluation and auditing protocol, not an inference algorithm: it examines whether test-time interventions apply, realize a target state, translate into a correct endpoint, cause harm, and incur resource or inference costs. Its paired comparisons hold the incoming state and evaluation conditions fixed, while cross-model analyses vary the receiving configuration. In executable financial reasoning experiments, the same context expansion harmed one configuration and benefited another; sufficient evidence did not ensure correctness because configurations translated evidence into answers differently; and diagnostic headroom plus local consistency did not guarantee practical gains. The authors conclude that an action's value depends on the input and model configuration, so diagnosis alone does not specify a deployable policy.

### Source excerpt

arXiv:2611.25566v1 Announce Type: new Abstract: Test-time actions can satisfy retrieval, state-realization, program-consistency, or rendering objectives without correcting an answer. Test-time Error Response and Realizability Audit (TERRA) is an evaluation/auditing protocol, not a new inference algorithm. It uses a unified decomposition of intervention outcomes into applicability, target-state realization, endpoint translation, harm, and resource/inference cost. Paired comparisons fix the incoming state and evaluation conditions; cross-model analyses vary the receiving configuration. Executable financial reasoning experiments yield three findings: the same context expansion can harm one configuration and benefit another; sufficient evidence does not ensure correctness because configurations differ in evidence-to-answer translation; and diagnostic headroom---accuracy attainable with idealized diagnostic information over the measured baseline---plus local consistency does not ensure practical gains. Action value is $V(a\mid x,m)$, not diagnosis alone; diagnosis is not a deployable policy $\pi(a\mid x,m)$.

## Your AI Agent Is Only Seeing Half The Answer

DevFeed: [Your AI Agent Is Only Seeing Half The Answer](<https://devfeed.tech/articles/your-ai-agent-is-only-seeing-half-the-answer-145830.md>)

Original publisher: [Read original article](<https://www.starburst.io/blog/your-ai-agent-is-only-seeing-half-the-answer>)

Published: 2026-11-19T18:00:05Z

Content type: article

Language: en

Sources: [Starburst Data](<https://devfeed.tech/sources/starburst-data.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Apache Iceberg](<https://devfeed.tech/topics/apache-iceberg.md>)

Tags: [access-control](<https://devfeed.tech/tags/access-control.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-data](<https://devfeed.tech/tags/ai-data.md>), [ai-data-engine](<https://devfeed.tech/tags/ai-data-engine.md>), [analytics](<https://devfeed.tech/tags/analytics.md>), [api](<https://devfeed.tech/tags/api.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [business](<https://devfeed.tech/tags/business.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [connectors](<https://devfeed.tech/tags/connectors.md>), [console](<https://devfeed.tech/tags/console.md>), [content](<https://devfeed.tech/tags/content.md>), [context](<https://devfeed.tech/tags/context.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [control](<https://devfeed.tech/tags/control.md>)

### AI overview

Starburst AIDA and NetApp AIDE are presented as a way to answer questions that require both database values and document context, such as matching contract renewal dates and spend with notice clauses. AIDA plans the question, sends structured queries through Starburst's federated SQL engine and document searches through AIDE's MCP server, then joins the results and cites source versions. Starburst is described as querying structured and semi-structured data in place across 50+ connectors, with row- and column-level controls, masking, and auditing. AIDE makes files on NetApp storage searchable, preserves file permissions, and exposes content through MCP; its Metadata Services catalog reads metadata without copying files or using GPUs. The stated tradeoff is that this approach avoids a separate external index, but its document capabilities depend on files being accessible through AIDE and NetApp storage interfaces. The excerpt recommends starting with one cross-source question and a joint architecture workshop.

### Source excerpt

Most teams building AI agents hit the same problem. Everything isn't in the same place. ...

## Sovereign AI is about control, not just location

DevFeed: [Sovereign AI is about control, not just location](<https://devfeed.tech/articles/sovereign-ai-is-about-control-not-just-location-143168.md>)

Original publisher: [Read original article](<https://www.networkworld.com/article/4250239/sovereign-ai-is-about-control-not-just-location.html>)

Author: Zeus Kerravala

Published: 2026-11-19T08:00:00Z

Content type: news

Language: en

Sources: [Network World](<https://devfeed.tech/sources/network-world.md>)

Topics: [Sovereign AI](<https://devfeed.tech/topics/sovereign-ai.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [digital sovereignty](<https://devfeed.tech/topics/digital-sovereignty.md>), [observability](<https://devfeed.tech/topics/observability.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-factory](<https://devfeed.tech/tags/ai-factory.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-services](<https://devfeed.tech/tags/ai-services.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [big-ip](<https://devfeed.tech/tags/big-ip.md>), [big-ip-next-for-kubernetes](<https://devfeed.tech/tags/big-ip-next-for-kubernetes.md>), [bluefield-3](<https://devfeed.tech/tags/bluefield-3.md>), [control](<https://devfeed.tech/tags/control.md>), [deal](<https://devfeed.tech/tags/deal.md>), [edition-us](<https://devfeed.tech/tags/edition-us.md>), [f5](<https://devfeed.tech/tags/f5.md>), [governance](<https://devfeed.tech/tags/governance.md>), [gpus](<https://devfeed.tech/tags/gpus.md>), [high-performance](<https://devfeed.tech/tags/high-performance.md>), [industry](<https://devfeed.tech/tags/industry.md>), [industry-markets-network-security-networking-security-telecommunications-industry](<https://devfeed.tech/tags/industry-markets-network-security-networking-security-telecommunications-industry.md>), [isolation](<https://devfeed.tech/tags/isolation.md>), [kubernetes](<https://devfeed.tech/tags/kubernetes.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [location](<https://devfeed.tech/tags/location.md>), [markets](<https://devfeed.tech/tags/markets.md>), [network-security](<https://devfeed.tech/tags/network-security.md>), [networking](<https://devfeed.tech/tags/networking.md>), [norwegian](<https://devfeed.tech/tags/norwegian.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [security](<https://devfeed.tech/tags/security.md>), [sovereign-ai](<https://devfeed.tech/tags/sovereign-ai.md>), [storytype-analysis](<https://devfeed.tech/tags/storytype-analysis.md>), [telecommunications-industry](<https://devfeed.tech/tags/telecommunications-industry.md>)

### AI overview

The Telenor AI Factory deployment pairs F5 BIG-IP Next for Kubernetes with Nvidia BlueField-3 DPUs to provide an AI service plane for sovereign AI services. Its current controls include tenant isolation, policy-based traffic control, and telemetry for audits and compliance; token-aware governance, circuit breaking, and inference-aware traffic optimization are planned as services scale. The article frames sovereignty as control and auditability over data and workloads, beyond keeping GPUs in-country. It argues telcos could monetize AI by metering and governing tokens, while an AI-aware data path can help allocate GPU capacity efficiently. F5 supplies granular records for billing systems but not billing itself. The Telenor deployment is described as small, and much of its token governance remains planned.

### Source excerpt

Every year at MWC and other telco-related shows, I hear the same line. A new technology wave will finally let telcos create new revenue-generating services. Cloud, virtualization, software, and a number of other trends were the magic elixir that would help telcos be more than just pipes. Every year, most of them use the technology to cut costs while continuing to sell connectivity. The industry has spent two decades trying to avoid becoming a "dumb pipe," and for the most part, it has failed. AI could be the thing that breaks the pattern, but only for operators who understand what they're selling. A recent announcement from F5 and Telenor ASA offers a glimpse of what that could look like. The news Recently, F5 announced that Telenor AI Factory selected BIG-IP Next for Kubernetes, running on Nvidia BlueField-3 data processing units, as the AI service plane for its Nvidia-powered sovereign AI infrastructure. The platform is designed for Norwegian and European organizations seeking high-performance AI services while maintaining control over where data is processed, stored, and accessed. From the outset, F5 provides multitenancy and isolation, policy-based traffic control, and real-time telemetry for auditing and compliance. Telenor also plans to add token-aware governance, circuit breaking, and inference-aware traffic optimization as its commercial AI services scale. I recently spoke with Ahmed Guetari, senior vice president and general manager of F5's service provider business, about the deal and its implications for telcos more broadly. Sovereignty is a control problem Most of the conversation about sovereign AI centers on GPUs located on national soil. That's obviously important, but it's table stakes. A rack of GPUs in Oslo doesn't make an AI service sovereign if the operator can't prove which tenant accessed what data, which model was called, and where the traffic went. Guetari was open about the term itself being fuzzy. "Sovereignty is very vague and not always c

## Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems

DevFeed: [Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems](<https://devfeed.tech/articles/colosseum-auditing-collusion-in-cooperative-multi-agent-systems-142508.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2602.15198>)

Author: Mason Nakamura, Abhinav Kumar, Saswat Das, Sahar Abdelnabi, Saaduddin Mahmud, Ferdinando Fioretto, Shlomo Zilberstein, Eugene Bagdasarian

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [agent to agent communication](<https://devfeed.tech/topics/agent-to-agent-communication.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [coalition](<https://devfeed.tech/tags/coalition.md>), [communication](<https://devfeed.tech/tags/communication.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [decision-making](<https://devfeed.tech/tags/decision-making.md>), [framework](<https://devfeed.tech/tags/framework.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [multi-agent-systems](<https://devfeed.tech/tags/multi-agent-systems.md>), [network](<https://devfeed.tech/tags/network.md>), [plan](<https://devfeed.tech/tags/plan.md>), [safety](<https://devfeed.tech/tags/safety.md>), [secret](<https://devfeed.tech/tags/secret.md>), [strategies](<https://devfeed.tech/tags/strategies.md>), [tactics](<https://devfeed.tech/tags/tactics.md>), [tasks](<https://devfeed.tech/tags/tasks.md>)

### AI overview

Colosseum audits collusion among cooperative LLM agents by grounding cooperation in a formal multi-agent decision framework and measuring action-based behavior through regret against the cooperative optimum, alongside communication-based behavior. It supports audits across benign settings, coalition objectives, persuasion tactics, and network topologies. The abstract reports that secret communication channels reveal a propensity to collude in most out-of-the-box models, termed "emergent collusion"; it also finds "collusion on paper," where agents discuss plans to collude but often choose non-collusive actions. The supplied abstract says the framework studies factors affecting collusion and possible mitigation strategies, but does not specify those strategies.

### Source excerpt

arXiv:2602.15198v3 Announce Type: replace-cross Abstract: Multi-agent systems, where LLM agents communicate through free-form language, enable sophisticated coordination for solving complex cooperative tasks. This surfaces a unique safety problem when a group of agents forms a coalition and colludes to pursue secondary goals and degrade the joint objective. In this paper, we present Colosseum, a framework for auditing LLM agents' collusive behavior in multi-agent settings. We ground how agents cooperate through a formal multi-agent decision-making framework and measure action-based collusive behavior in actions via regret relative to the cooperative optimum and compare it with communication-based collusive behavior. Colosseum enables audits of LLM agents for collusion under benign settings, different coalition objectives, persuasion tactics, and network topologies. We then introduce a new behavioral probe by creating secret communication channels between agents, showing that most out-of-the-box models exhibit a propensity to collude under this probe, which we term emergent collusion. Furthermore, we discover ``collusion on paper'' when agents plan to collude in text but often pick non-collusive actions. Colosseum provides a new way to audit collusion in cooperative multi-agent systems while presenting observations about how collusion emerges, what affects collusion efficacy, and which strategies may mitigate it.

## Overlooked and Underutilized: Prototypes are the Missing Piece in Tabular XAI

DevFeed: [Overlooked and Underutilized: Prototypes are the Missing Piece in Tabular XAI](<https://devfeed.tech/articles/overlooked-and-underutilized-prototypes-are-the-missing-piece-in-tabular-xai-143567.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24831>)

Author: Jacek Karolczak, Jerzy Stefanowski

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [cognitive-load](<https://devfeed.tech/tags/cognitive-load.md>), [concepts](<https://devfeed.tech/tags/concepts.md>), [cs-hc](<https://devfeed.tech/tags/cs-hc.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data](<https://devfeed.tech/tags/data.md>), [explainable-ai](<https://devfeed.tech/tags/explainable-ai.md>), [features](<https://devfeed.tech/tags/features.md>), [future](<https://devfeed.tech/tags/future.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [machine-learning-models](<https://devfeed.tech/tags/machine-learning-models.md>), [parts](<https://devfeed.tech/tags/parts.md>), [pipelines](<https://devfeed.tech/tags/pipelines.md>), [prototype](<https://devfeed.tech/tags/prototype.md>), [prototypes](<https://devfeed.tech/tags/prototypes.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [xai](<https://devfeed.tech/tags/xai.md>)

### AI overview

The paper argues that prototypes can make tabular AI explanations easier for non-experts to use alongside feature-level auditing. It highlights prototypical parts--groups of related features bounded as subspaces--as cohesive, human-understandable concepts that may reduce cognitive effort compared with inspecting many isolated feature attributions. It also introduces a Geometric Validity Triad to frame how prototypes can validate the data manifold, a dimension the abstract says existing methods often neglect. The proposed direction is a composite explanation pipeline combining similarity-based reasoning with rigorous feature-level auditing; open challenges and potential solutions are outlined, but not detailed in the supplied abstract.

### Source excerpt

arXiv:2611.24831v1 Announce Type: new Abstract: Explainable AI (XAI) is frequently applied to high-stakes tabular data, yet the field relies heavily on feature attribution methods. While these approaches offer technical transparency, evaluating dozens of isolated features demands excessive cognitive effort from non-expert users. In this paper, we argue that prototypes remain overlooked and underutilized, and should be integrated as a complementary, intuitive gateway to auditing complex machine learning models. Although entire exemplars function as an intuitive baseline for prototype-based explanations, we highlight prototypical parts as particularly valuable for tabular data. By grouping related features into bounded subspaces, prototypical parts technically embody human-understandable concepts, replacing isolated features with cohesive semantic units. We support this integration through two primary lenses. Cognitively, these parts align with Rosch's categorization theory and Kahneman's System 1 thinking to naturally reduce cognitive load for humans. Then, we introduce a Geometric Validity Triad to conceptualize why prototypes validate the data manifold - a critical dimension often neglected by existing methods. We contend that making tabular XAI genuinely comprehensible requires shifting toward composite explanation pipelines that pair intuitive similarity-based reasoning with rigorous feature-level auditing. The paper concludes by outlining open challenges of prototype-based methods and discussing potential solutions to guide future research.

## Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets

DevFeed: [Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets](<https://devfeed.tech/articles/are-you-learning-biological-signal-or-shortcuts-auditing-and-mitigating-bias-in-protein-protein-interaction-datasets-143307.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.10193>)

Author: Judith Bernett, Anton Spannagl, Joel {\\AA}s, Markus List, David B. Blumenthal

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Bioinformatics](<https://devfeed.tech/topics/bioinformatics.md>), [Dataset development](<https://devfeed.tech/topics/dataset-development.md>)

Tags: [3d](<https://devfeed.tech/tags/3d.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [also](<https://devfeed.tech/tags/also.md>), [amplify](<https://devfeed.tech/tags/amplify.md>), [analyze](<https://devfeed.tech/tags/analyze.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [bank](<https://devfeed.tech/tags/bank.md>), [based](<https://devfeed.tech/tags/based.md>), [bias](<https://devfeed.tech/tags/bias.md>), [can](<https://devfeed.tech/tags/can.md>), [choice](<https://devfeed.tech/tags/choice.md>), [combines](<https://devfeed.tech/tags/combines.md>), [construction](<https://devfeed.tech/tags/construction.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [interaction](<https://devfeed.tech/tags/interaction.md>), [learning](<https://devfeed.tech/tags/learning.md>), [protein](<https://devfeed.tech/tags/protein.md>), [q-bio-mn](<https://devfeed.tech/tags/q-bio-mn.md>), [shortcuts](<https://devfeed.tech/tags/shortcuts.md>), [signal](<https://devfeed.tech/tags/signal.md>)

### AI overview

The study finds that PPI prediction datasets can encode shortcuts that models exploit instead of biological signal. Random splitting creates strong topological shortcuts; even after removing train-test protein overlap, shortcuts related to self-interactions, taxonomic identity, and functional relatedness remain, with prevalence varying by data source. Sampling negatives from highly replicated non-interactions can amplify functional-relatedness shortcuts. The authors provide an open Nextflow pipeline that uses integer linear programs for similarity-aware, data-loss-minimizing splits and bias-minimizing negative sampling. They suggest the negative-sampling approach may extend to other machine-learning tasks with many more negative candidates than positives.

### Source excerpt

arXiv:2609.10193v2 Announce Type: replace-cross Abstract: Protein-protein interaction (PPI) databases do not faithfully reflect biological realities. Instead, they are influenced by study and technical biases that distort certain protein and interaction attributes. Machine learning models can exploit these as learning shortcuts if the negative dataset is not constructed with care. So far, the shortcuts introduced during PPI dataset construction have only been examined in isolation. Here, we systematically characterize both reported and, to our knowledge, previously unreported biases in PPI datasets that lead machine learning models to learn shortcuts instead of biological signal. We analyze HIPPIE, IntAct, and STRING, dedicated PPI databases, as well as two datasets derived from 3D-structural information in the Protein Data Bank (PDB). We show that random data splitting introduces strong topological shortcuts. When train-test protein overlap is removed, the resulting datasets still retain usable shortcuts stemming from self-interactions, taxonomic identity, and functional relatedness, whose prevalence interestingly depends on the data source. We further show that sampling negatives from a set of highly replicated non-interactions, an intuitively appealing choice, can amplify the shortcut stemming from functional relatedness. To detect and mitigate these biases, we provide an open Nextflow pipeline that combines similarity-aware, data-loss-minimizing dataset splitting with bias-minimizing negative sampling, both formulated as integer linear programs. Its key concept of quantifying biases to minimize them through optimization-based negative sampling can, in principle, be extended to any machine learning problem where the pool of negative candidates is much larger than the positives and is thus of interest also beyond PPI prediction.

## Auditing Path-Specific Fairness Across Causal Discovery Algorithms: A Heart Failure Case Study

DevFeed: [Auditing Path-Specific Fairness Across Causal Discovery Algorithms: A Heart Failure Case Study](<https://devfeed.tech/articles/auditing-path-specific-fairness-across-causal-discovery-algorithms-a-heart-failure-case-study-142512.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2603.15926>)

Author: Nitish Nagesh, Elahe Khatibi, Thomas Hughes, Mahdi Bagheri, Pratik Gajane, Amir M. Rahmani

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [responsible-ai](<https://devfeed.tech/topics/responsible-ai.md>), [causality](<https://devfeed.tech/topics/causality.md>), [Graph Mining](<https://devfeed.tech/topics/graph-mining.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [allows](<https://devfeed.tech/tags/allows.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [assessment](<https://devfeed.tech/tags/assessment.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bias](<https://devfeed.tech/tags/bias.md>), [can](<https://devfeed.tech/tags/can.md>), [case](<https://devfeed.tech/tags/case.md>), [case-study](<https://devfeed.tech/tags/case-study.md>), [choice](<https://devfeed.tech/tags/choice.md>), [contribution](<https://devfeed.tech/tags/contribution.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [discovery](<https://devfeed.tech/tags/discovery.md>), [failure](<https://devfeed.tech/tags/failure.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [study](<https://devfeed.tech/tags/study.md>)

### AI overview

The study finds that causal discovery algorithm choice can change path-specific fairness estimates and the fairness-utility trade-off for heart failure risk models. Using an expert-elicited graph benchmark, it reports that ICA-LiNGAM had the lowest structural Hamming distance and highest recall, while GES had the highest precision. The mediated effect on ejection fraction differed by 1.2 percentage points between the original and discovered graphs, contributing to variation in the fairness-utility ratio. The authors argue these findings support fine-grained, discovery-aware fairness audits; the supplied abstract does not detail the methods or broader limitations.

### Source excerpt

arXiv:2603.15926v2 Announce Type: replace-cross Abstract: Causal discovery is increasingly applied to healthcare data, where the recovered relationships can encode biased pathways that propagate to downstream prediction. Path-specific fairness analysis, through direct, indirect, and spurious effect decompositions along with fairness-utility ratio assessment allows bias quantification and associated prediction cost. However, existing evaluations of causal discovery under fairness constraints have largely used non-clinical datasets and have not separated the contribution of confounders from that of mediators, leaving open how discovery-algorithm choice propagates through path-specific fairness to the utility trade-off in clinical settings. This work addresses that gap by establishing an expert-elicited causal graph benchmark for the heart failure clinical records dataset and tracing discovery algorithms through composite, per-variable fairness decompositions and fairness-utility ratio. ICA-LiNGAM achieves the lowest structural Hamming distance and highest recall, while GES achieves the highest precision. For path-specific effects, there was a 1.2 percentage-point difference in the mediated effect on ejection fraction on overall causal fairness between the original and discovered graphs. This ultimately led to variations in the fairness-utility ratio highlighting need for fine-grained evaluation. These results indicate that discovery-aware, path-specific fairness evaluation is consequential in clinical settings and warrants inclusion in routine auditing of clinical risk models.

## Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness

DevFeed: [Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness](<https://devfeed.tech/articles/securing-and-auditing-long-horizon-llm-agents-state-transfer-auditing-and-controls-in-the-agent-harness-143119.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23346>)

Author: Deep Shah, Nehal Kathrotia

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-harness](<https://devfeed.tech/tags/agent-harness.md>), [agents](<https://devfeed.tech/tags/agents.md>), [analyze](<https://devfeed.tech/tags/analyze.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [attribution](<https://devfeed.tech/tags/attribution.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [authorization](<https://devfeed.tech/tags/authorization.md>), [boundary](<https://devfeed.tech/tags/boundary.md>), [browsers](<https://devfeed.tech/tags/browsers.md>), [cache](<https://devfeed.tech/tags/cache.md>), [controls](<https://devfeed.tech/tags/controls.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [harness](<https://devfeed.tech/tags/harness.md>), [horizon](<https://devfeed.tech/tags/horizon.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [state](<https://devfeed.tech/tags/state.md>), [state-transfer](<https://devfeed.tech/tags/state-transfer.md>)

### AI overview

The paper argues that important failures in long-horizon LLM agents occur when state moves between working context, session history, persistent memory, and external systems--transfers conventional prompt-and-tool logs may miss. It proposes hash-chained audit records at each transfer boundary, combining deterministic checks with model-based semantic checks and linking internal context, memory, and inference-server cache state to external actions. The records are also intended to support controls such as least-privilege capabilities, replay-resistant authorization outside the prompt, information flow control, obligation monitoring, and transactional recovery. A reference middleware evaluated on 240 multi-step fault-injection trajectories reportedly localized causes and detected violations missed by lifecycle and tool-only logs, with microsecond deterministic overhead. This summary is based on the abstract.

### Source excerpt

arXiv:2611.23346v1 Announce Type: new Abstract: Long-horizon LLM agents across codebases, web browsers, and enterprise infrastructure rely on an execution harness to move state between the working context, session history, persistent memory, and external environment. While conventional execution traces log model prompts and tool calls, critical security and reliability failures occur during state transfers: a transfer can admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily (context compaction dropping user constraints, memory consolidation stripping source authority), or restore state incompletely (workspace rollbacks leaving aborted KV state active on the inference server). Because current traces omit what crosses these boundaries, failure attribution and runtime monitors fail on multi-step trajectories. We present a state-transfer audit architecture for the agent harness. At each transfer boundary, we define an audit point specifying the minimal hash-chained record the harness emits, separating low-overhead deterministic checks from model-based semantic checks and binding internal context, memory, and inference-server cache state to external actions. We show that the same transfer records required for causal failure attribution supply the durable state needed by runtime security controls, including least-privilege capabilities, replay-resistant authorization outside the prompt, information flow control, obligation monitors, and transactional recovery across multi-agent and infrastructure pipelines. Finally, we analyze multimodal GUI adaptations and implement and evaluate a reference harness middleware across 240 multi-step fault-injection trajectories, showing that transfer records localize root causes and catch cross-transfer violations missed by lifecycle and tool-only logs at microsecond deterministic overhead.

## Prioritize, Don't Decide: What We Learned Auditing a Deployed Human-in-the-Loop CV Sorter

DevFeed: [Prioritize, Don't Decide: What We Learned Auditing a Deployed Human-in-the-Loop CV Sorter](<https://devfeed.tech/articles/prioritize-don-t-decide-what-we-learned-auditing-a-deployed-human-in-the-loop-cv-sorter-143383.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23359>)

Author: Hadi Mohammadi, Jeroen Sparla, Alexandros Paliouras, Georg Krempl, Robert A. Bagheri

Published: 2026-11-19T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [experience report](<https://devfeed.tech/topics/experience-report.md>)

Tags: [academic](<https://devfeed.tech/tags/academic.md>), [applications](<https://devfeed.tech/tags/applications.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [budget](<https://devfeed.tech/tags/budget.md>), [classifier](<https://devfeed.tech/tags/classifier.md>), [cold-start](<https://devfeed.tech/tags/cold-start.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [control](<https://devfeed.tech/tags/control.md>), [cross-encoder](<https://devfeed.tech/tags/cross-encoder.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-ir](<https://devfeed.tech/tags/cs-ir.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cv](<https://devfeed.tech/tags/cv.md>), [data](<https://devfeed.tech/tags/data.md>), [database](<https://devfeed.tech/tags/database.md>), [discrimination](<https://devfeed.tech/tags/discrimination.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [embedding-model](<https://devfeed.tech/tags/embedding-model.md>), [encoder](<https://devfeed.tech/tags/encoder.md>), [human-in-the-loop](<https://devfeed.tech/tags/human-in-the-loop.md>), [llm](<https://devfeed.tech/tags/llm.md>), [load](<https://devfeed.tech/tags/load.md>), [local](<https://devfeed.tech/tags/local.md>)

### AI overview

The CV Priority Sorter proposes a reading order while leaving hiring decisions to the committee; its frozen multilingual encoder feeds a session-specific linear model that retrains after each label and is discarded at session end. In a 300-CV synthetic pool, the order surfaced 95% of relevant candidates after half the pile, but document length alone made precision-style metrics uninformative. Regenerating the pool with a shared length budget removed that shortcut while preserving the reading-depth result. The reported four-level agreement score is a thresholded binary probability, so it reflects classifier confidence rather than ranking and declines as labels accumulate. Uncertainty sampling preserved model quality but increased total committee reading above the no-label baseline by querying CVs that otherwise would be skipped. Under length control, hosted LLM embeddings improved zero-label ordinal discrimination from 0.66 to 0.80, but the gain faded after about two dozen labels; a local cross-encoder matched the cold-start gain without sending data out.

### Source excerpt

arXiv:2611.23359v1 Announce Type: cross Abstract: A single academic vacancy can draw hundreds of applications, and the first screening round consumes most of a committee's time. We describe a CV Priority Sorter deployed at a national platform for academic vacancies and integrated with its vacancy database. It proposes a reading order for a pile of CVs and leaves every decision to the committee, pairing a frozen multilingual transformer encoder with a per-session linear model that retrains after each label and is discarded when the session ends. It works: on a pool of 300 synthetic CVs, a committee reading in the tool's order reaches 95% of the relevant candidates after half the pile, where an arbitrary order needs almost all of it. This paper is about what we found when we audited that claim, rather than about the system itself. Three findings changed the product. The pool turned out to be sortable by document length alone, which makes precision-style metrics on it uninformative; regenerating it under a shared length budget removes that shortcut and leaves the reading-depth result standing. The four-level agreement score such systems report, ours included, is a binary probability cut at fixed thresholds, so it follows the classifier's confidence rather than its ranking and falls as labels accumulate. And uncertainty sampling, the obvious way to implement a "suggest the next CV" button, leaves model quality intact but raises the committee's total reading load above the no-label baseline, because the CVs it asks about are the ones that would otherwise be skipped. A fourth result shows where to look next: a hosted LLM embedding model lifts zero-label ordinal discrimination from 0.66 to 0.80 under length control, then stops mattering once the session model has two dozen labels; a local cross-encoder matches that cold-start gain without sending data out. We report the fixes each finding forced.

## Auditing black-box clinical prediction models with a large language model through queries alone: a blinded test with planted flaws

DevFeed: [Auditing black-box clinical prediction models with a large language model through queries alone: a blinded test with planted flaws](<https://devfeed.tech/articles/auditing-black-box-clinical-prediction-models-with-a-large-language-model-through-queries-alone-a-blinded-test-with-planted-flaws-142842.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24877>)

Author: Zhiyu Liu, Xinyu Yang

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [queries](<https://devfeed.tech/tags/queries.md>)

### AI overview

In a blinded test on International Stroke Trial data, GPT-6 Astra audited clinical prediction models using prediction queries alone. Its surrogate models reproduced model log-odds with mean R² 0.976, and it identified planted flaws in 69 of 72 runs; a purpose-written script found all 24 flaws. However, many reported problems did not match the models' average behavior, and 15% of checkable reports were not found in the model. Of unexpected relationships extracted from clean models, 11% held in later patients, compared with 6% for random pairs and 28% for the strongest candidates from a regression scan. The results support query-only auditing while showing that reported problems require model checks and discovered relationships require validation on data.

### Source excerpt

arXiv:2611.24877v1 Announce Type: cross Abstract: Large language models (LLMs) are used to explain clinical prediction models, on the promise that they can say what a model computes, find its faults and even reveal new knowledge in it. We tested the three promises where the answers could be checked. From the International Stroke Trial we trained four kinds of model (logistic regression, explainable boosting machine, gradient-boosted trees, neural network) for two outcomes on patients randomised earlier, planted one flaw in each of 24 copies (an under-predicted subgroup, a reversed risk factor or an implausible step) and mixed them with the 8 clean models. GPT-6 Astra audited each model three times, blind, through a tool that only returned the model's predictions; the judging rules were committed before any run and applied by program. Its surrogate models, with a median of 128 terms, reproduced the audited models' log-odds with a mean $R^2$ of 0.976 (0.959 for an additive surrogate of 42 terms fitted by an analyst). It found the planted flaw in 69 of 72 runs (96%; 95% CI 80% to 99%); a script written for the purpose found all 24. It also reported many more problems: 60% of 1,089 checkable reports failed the pre-specified check against the model's average behaviour; post hoc, most of these were real but small steps, and 15% of the checkable reports were not found in the model. Of 601 unexpected relationships it read out of the clean models, 11% held in patients randomised later (random pairs 6%; a regression scan's shorter list of its strongest candidates, 28%); after a check on the training data, 33% against 30%. With query access alone, an LLM can describe a black-box model faithfully and find nearly every planted flaw, as a purpose-written script also did; its reports of problems must be checked on the model and its discoveries on data.

## What's new in PHP 8.6 and what that means for WordPress?

DevFeed: [What's new in PHP 8.6 and what that means for WordPress?](<https://devfeed.tech/articles/what-s-new-in-php-8-6-and-what-that-means-for-wordpress-144055.md>)

Original publisher: [Read original article](<https://developer.wordpress.org/news/2026/11/whats-new-in-php-8-6-and-what-that-means-for-wordpress/>)

Author: Milana Cap

Published: 2026-11-18T19:53:24Z

Content type: article

Language: en

Sources: [WordPress](<https://devfeed.tech/sources/wordpress-developer-blog.md>)

Topics: [WordPress](<https://devfeed.tech/topics/wordpress.md>), [deprecated](<https://devfeed.tech/topics/deprecated.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>)

Tags: [7-4](<https://devfeed.tech/tags/7-4.md>), [adoption](<https://devfeed.tech/tags/adoption.md>), [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [caching](<https://devfeed.tech/tags/caching.md>), [clamp](<https://devfeed.tech/tags/clamp.md>), [closure](<https://devfeed.tech/tags/closure.md>), [compatibility](<https://devfeed.tech/tags/compatibility.md>), [core](<https://devfeed.tech/tags/core.md>), [css](<https://devfeed.tech/tags/css.md>), [dashboard](<https://devfeed.tech/tags/dashboard.md>), [dependencies](<https://devfeed.tech/tags/dependencies.md>), [developers](<https://devfeed.tech/tags/developers.md>), [features](<https://devfeed.tech/tags/features.md>), [performance](<https://devfeed.tech/tags/performance.md>), [php](<https://devfeed.tech/tags/php.md>), [php-8-6](<https://devfeed.tech/tags/php-8-6.md>), [php-8-x](<https://devfeed.tech/tags/php-8-x.md>), [security](<https://devfeed.tech/tags/security.md>), [updates](<https://devfeed.tech/tags/updates.md>), [wordpress](<https://devfeed.tech/tags/wordpress.md>)

### AI overview

PHP 8.6 compatibility work is underway in WordPress 7.2, but adopting its new language features in core is constrained by WordPress's PHP 7.4 minimum. The article highlights deprecation audits--including mbstring's Oniguruma regex functions and other deprecated calls--and areas needing attention such as tests, bundled libraries, formatting, and HTTP/URI handling. PHP 8.6 adds partial function application and `clamp()`, while its closure optimizations can benefit WordPress at runtime once compatible. The author sees native URI handling and `Time\Duration` as possible future improvements, and argues that feature gating and dashboard upgrade prompts could make newer PHP versions more compelling. The article reports PHP 7.4 use at about 17% at October's end, a factor in minimum-version decisions.

### Source excerpt

Having a reputation for long backward compatibility, as stable as it sounds, has an ugly face of dragging performance and security down, as well as keeping the shiny new functionalities away from WordPress core developers. Which PHP 8.6 features we can look forward to?

## Auditing AI Agents in Banks: What Changes for Internal Audit

DevFeed: [Auditing AI Agents in Banks: What Changes for Internal Audit](<https://devfeed.tech/articles/auditing-ai-agents-in-banks-what-changes-for-internal-audit-141048.md>)

Original publisher: [Read original article](<https://www.lyzr.ai/blog/auditing-ai-agents-in-banks/>)

Author: Lyzr Team

Published: 2026-11-18T10:17:35Z

Content type: article

Language: en

Sources: [Lyzr.ai](<https://devfeed.tech/sources/lyzr-ai.md>)

Topics: [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [accountability](<https://devfeed.tech/tags/accountability.md>), [actions](<https://devfeed.tech/tags/actions.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-behavior](<https://devfeed.tech/tags/agent-behavior.md>), [agent-control](<https://devfeed.tech/tags/agent-control.md>), [agent-identity](<https://devfeed.tech/tags/agent-identity.md>), [agent-observability](<https://devfeed.tech/tags/agent-observability.md>), [agent-safety](<https://devfeed.tech/tags/agent-safety.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [banks](<https://devfeed.tech/tags/banks.md>), [changes](<https://devfeed.tech/tags/changes.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [defense](<https://devfeed.tech/tags/defense.md>), [testing](<https://devfeed.tech/tags/testing.md>)

### AI overview

Bank internal audit of agentic AI depends on an action trail captured when decisions occur: it should link agent identity and permissions to decisions, approvals, escalations, overrides, tests, and changes. KPMG and IIA Singapore favor real-time monitoring over retrospective sampling, while Sphere notes auditors may still sample when the full population is available. The key change is that sampling becomes a choice, provided the trail is complete. Auditors should verify the agent inventory against live systems, prioritize work by risk, replay real decisions from records, test approval gates for meaningful review, and retest after model, tool, data, or rule changes. Agents can retrieve evidence and run defined tests, but auditors retain judgment over materiality, context, and scope. The excerpt says US supervisors had not issued audit-specific agent rules among those reviewed, and the sources did not address auditor independence when using agents. Madison is described as a first and second line system whose records auditors should test independently.

### Source excerpt

Agentic AI changes the third line of defense in three ways: audit moves from periodic samples toward continuous testing, evidence has to come from a...

## RagHarnessDebugger: Jev-Guided Agentic RAG in Contract Inference Repair Auditing

DevFeed: [RagHarnessDebugger: Jev-Guided Agentic RAG in Contract Inference Repair Auditing](<https://devfeed.tech/articles/ragharnessdebugger-jev-guided-agentic-rag-in-contract-inference-repair-auditing-139615.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22991>)

Author: Moriya Dechtiar, Daniel Martin Katz, Michael J. Bommarito II, Hongming Wang, Sylvain Jaume

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [Agentic AI](<https://devfeed.tech/topics/what-is-agentic-ai.md>), [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-rag](<https://devfeed.tech/tags/agentic-rag.md>), [audit](<https://devfeed.tech/tags/audit.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [audits](<https://devfeed.tech/tags/audits.md>), [continuous-improvement](<https://devfeed.tech/tags/continuous-improvement.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [debugger](<https://devfeed.tech/tags/debugger.md>), [debugging](<https://devfeed.tech/tags/debugging.md>), [decision](<https://devfeed.tech/tags/decision.md>), [explainability](<https://devfeed.tech/tags/explainability.md>), [harness](<https://devfeed.tech/tags/harness.md>), [inference](<https://devfeed.tech/tags/inference.md>), [jev](<https://devfeed.tech/tags/jev.md>), [rag](<https://devfeed.tech/tags/rag.md>), [repair](<https://devfeed.tech/tags/repair.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [retrieval-augmented-generation](<https://devfeed.tech/tags/retrieval-augmented-generation.md>), [retrieval-augmented-generation-rag](<https://devfeed.tech/tags/retrieval-augmented-generation-rag.md>), [root-cause-analysis](<https://devfeed.tech/tags/root-cause-analysis.md>)

### AI overview

RagHarnessDebugger audits agentic RAG repair decisions by checking trace integrity, whether validations apply to the candidates they are credited to, and which answers can satisfy the acceptance rule. Tested on ContractNLI nondisclosure agreements with two reader models and a separate-contract replication set, its audit, labeled error analysis, and replay exposed missed repairs and tradeoffs obscured by aggregate metrics. The reported acceptance gap covered 77% of available but unrealized repairs on development data and 76% on replication data. A revised harness improved accuracy over the BM25-only baseline by 3.6 and 3.7 points, respectively, and improved the full-contract-access reference score by 1.1 and 2.4 points. The abstract presents the work as decision root-cause analysis and visibility for ongoing RAG improvement; broader validation is not described.

### Source excerpt

arXiv:2611.22991v1 Announce Type: cross Abstract: Agentic systems using retrieval-augmented generation (RAG) must repair their own answers by gathering more evidence, testing alternative policies and deciding if, when, and what repairs are needed. Yet how do we know if a system's repairs are working when metrics like accuracy can be misleading? We present RagHarnessDebugger, a framework that records every repair decision in a Jev-guided harness for contract inference and audits the record to facilitate complete debugging of all decisions taken during a RAG repair journey. RagHarnessDebugger checks that each trace is intact, that each validation applies to the candidate it is counted for, and which answers the acceptance rule can reach. We tested our approach on ContractNLI non-disclosure agreements with two reader models and a replication set of separate contracts. Our debugger's audit layer exposed failure patterns that were hidden under aggregate metrics, policies acceptance cost of repairs vs regressions and policy acceptance gap of 77% and 76% of available but unrealised repairs on development and replication sets. Our audit, together with a labeled error analysis and replay mechanism, allowed following the execution path all the way to the terminal decision - uncovering where missing repairs became lost. Our revised harness improved accuracy over the initial BM25-only baseline by 3.6 points on development contracts and 3.7 points on the replication set, and even improved our full contract access reference score by 1.1 on development and 2.4 on replication set. While our work here demonstrates how the debugger can conduct decision root-cause analysis for missed repairs, the true value lies with providing this debugger for judgment visibility serving continuous improvement of RAG decisions and their explainability. Keywords: agentic RAG, AI evidence repair, harness explainability, counterfactual replay, Jev, ContractNLI

## Simplicity Goes Far: Auditing Prompt Optimizers with Demonstration-Conditioned Prompt Search

DevFeed: [Simplicity Goes Far: Auditing Prompt Optimizers with Demonstration-Conditioned Prompt Search](<https://devfeed.tech/articles/simplicity-goes-far-auditing-prompt-optimizers-with-demonstration-conditioned-prompt-search-140056.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22605>)

Author: Bohe Chen

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [DSPy](<https://devfeed.tech/topics/dspy.md>), [audit](<https://devfeed.tech/topics/audit.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [classification](<https://devfeed.tech/tags/classification.md>), [classification-accuracy](<https://devfeed.tech/tags/classification-accuracy.md>), [control](<https://devfeed.tech/tags/control.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [dspy](<https://devfeed.tech/tags/dspy.md>), [generate](<https://devfeed.tech/tags/generate.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reflection](<https://devfeed.tech/tags/reflection.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [search](<https://devfeed.tech/tags/search.md>), [select](<https://devfeed.tech/tags/select.md>), [selection](<https://devfeed.tech/tags/selection.md>), [structured](<https://devfeed.tech/tags/structured.md>), [systems](<https://devfeed.tech/tags/systems.md>), [token-usage](<https://devfeed.tech/tags/token-usage.md>)

### AI overview

Demonstration-Conditioned Prompt Search (DCPS) isolates prompt generation and candidate selection from heavier optimizer mechanisms. It freezes the generator, conditions proposals on randomly sampled demonstrations, scores them on validation data, and retains the best candidate, without learned updates, reflection, or evolution. In StablePrompt, DCPS matches PPO classification accuracy within 0.5 percentage points at roughly one-third the measured cost. In DSPy-based systems on GPT-4.1-mini, it matches GEPA's macro accuracy under the reported budgets with 55% fewer rollouts and 32% lower average measured API cost. Results vary by task: proposal and selection recover much of heavier methods' performance on structured and instruction-following tasks, while added mechanisms retain advantages on some open-ended and reasoning-heavy tasks. The authors recommend comparing optimizers with generate-and-select controls and reporting rollouts, token use, and dollar costs alongside accuracy.

### Source excerpt

arXiv:2611.22605v1 Announce Type: new Abstract: Automatic prompt optimization is growing more elaborate. Recent methods stack policy-gradient updates, trace reflection, and evolutionary search. These methods share a generate-and-select step; without controlling for it, reported gains may conflate the contributions of optimizer mechanisms and larger sampling budgets. We isolate this step as Demonstration-Conditioned Prompt Search (DCPS), repurposing the Automatic Prompt Engineer (APE) recipe as a mechanism-attribution control: it freezes the generator, conditions it on randomly sampled demonstrations, scores candidates on validation data, and keeps the best---no learned update, reflection, or evolution. Across two optimizer families, DCPS replaces Proximal Policy Optimization (PPO) in StablePrompt, matching classification accuracy within 0.5 pp at roughly one-third the measured cost; in DSPy-based compound AI systems, under the reported budgets, it matches GEPA's macro accuracy on GPT-4.1-mini while using 55% fewer rollouts at 32% lower measured API cost on average. The audit reveals a task-dependent pattern: proposal and selection recover much of heavier optimizers' performance on structured and instruction-following tasks, while additional mechanisms retain advantages in some open-ended and reasoning-heavy settings. We recommend benchmarking prompt optimizers against generate-and-select controls and reporting rollout counts, token usage, and dollar costs alongside accuracy. Code and artifacts: https://github.com/bhc6/DCPS.

## Right Diagnosis, Unsafe Route: Auditing Clinical AI Decision Trajectories and Corrective Oversight for Validity and Potential Harm

DevFeed: [Right Diagnosis, Unsafe Route: Auditing Clinical AI Decision Trajectories and Corrective Oversight for Validity and Potential Harm](<https://devfeed.tech/articles/right-diagnosis-unsafe-route-auditing-clinical-ai-decision-trajectories-and-corrective-oversight-for-validity-and-potential-harm-140059.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22612>)

Author: Guanqiao Chen, Wei Jiang, Shu Zhang, Qian Lin, Guanxu Yu, Siran Li, Chunhui Wang, Tianyu Shi

Published: 2026-11-18T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [audit](<https://devfeed.tech/topics/audit.md>), [ai-governance](<https://devfeed.tech/topics/ai-governance.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [clinical-ai](<https://devfeed.tech/tags/clinical-ai.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [decision](<https://devfeed.tech/tags/decision.md>), [diagnosis](<https://devfeed.tech/tags/diagnosis.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [hospital](<https://devfeed.tech/tags/hospital.md>), [llm](<https://devfeed.tech/tags/llm.md>), [oversight](<https://devfeed.tech/tags/oversight.md>), [safe](<https://devfeed.tech/tags/safe.md>), [safety](<https://devfeed.tech/tags/safety.md>)

### AI overview

The study finds that correct diagnoses and successful corrections can still involve unsafe clinical AI decisions. Its trajectory audit replays 300 cases from a hospital cardiology cohort, MIMIC-IV, and ASCENT, rating each test, treatment, and disposition for potential harm and clinical validity using only evidence available at that point. An LLM judge scores decisions, calibrated against physician review. Among hospital trajectories ending in an acceptable diagnosis, 55.2% included a decision with moderate or severe potential harm, and 24.8% included a clinically unacceptable decision; reported harm rates were 76.2% on MIMIC-IV and 39.6% on ASCENT. In an oversight audit, structured feedback resolved 73.6% of first errors, compared with 24.8% after bare rejection, but was followed by new potentially unsafe decisions in 30.2% of cases versus 19.7% after rejection. The findings support evaluating safety throughout the decision path and recovery, not only final diagnoses.

### Source excerpt

arXiv:2611.22612v1 Announce Type: new Abstract: A clinical AI system can reach a correct final diagnosis through intermediate decisions that could have harmed the patient, and can repair an error while introducing new risks. Most evaluations score final diagnoses; even stepwise benchmarks, which release a case's evidence over time, score whether systems reach the right diagnosis or gather the right information, not whether the decisions a patient passes through are safe. We make patient safety the target of stepwise evaluation. A trajectory audit replays real admissions checkpoint by checkpoint, lets each system decide with only the evidence released so far, and rates every test, treatment and disposition for its potential harm to the patient and, separately, for clinical validity, against the documented final diagnosis and physician-written disease rubrics yet on the evidence available when the decision was made. An LLM judge scores every decision, and a physician reference standard (triple annotation and senior adjudication of a stratified sample) calibrates the rates. An oversight audit branches each trajectory at its first actionable error and measures, on the same branches, whether corrective feedback repairs the error and whether it introduces new harm. Across 300 cases from a hospital cardiology cohort, MIMIC-IV and the public ASCENT benchmark, with up to seven AI systems per source, 55.2% of hospital trajectories reaching an acceptable final diagnosis passed through a decision with moderate or severe potential harm and 24.8% through a clinically unacceptable one; the harm rates were 76.2% on MIMIC-IV and 39.6% on ASCENT. Structured corrections resolved 73.6% of first errors versus 24.8% after bare rejection, yet new potentially unsafe decisions followed 30.2% of corrections versus 19.7% of rejections. A correct final diagnosis does not certify a safe trajectory, and a successful correction does not certify a safe recovery.

## How Many Training Labels Are Wrong? Auditing a Fitted Model through Gradient Allocation

DevFeed: [How Many Training Labels Are Wrong? Auditing a Fitted Model through Gradient Allocation](<https://devfeed.tech/articles/how-many-training-labels-are-wrong-auditing-a-fitted-model-through-gradient-allocation-141777.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22449>)

Author: Yuchen Li, Zongqi Fan, Xintong Li, Nguyen H. Tran, Ken-Tye Yong

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [machine learning overfitting](<https://devfeed.tech/topics/machine-learning-overfitting.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [cifar-10](<https://devfeed.tech/topics/cifar-10.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>)

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [cifar-10](<https://devfeed.tech/tags/cifar-10.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [gradient](<https://devfeed.tech/tags/gradient.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [model](<https://devfeed.tech/tags/model.md>), [noise](<https://devfeed.tech/tags/noise.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

The paper estimates training-label noise from one trained model by normalizing each example's per-example gradient magnitudes into a gradient allocation and checking whether relabeling it as its nearest other class better matches that class's mean allocation. The threshold-free audit needs no clean data, held-out set, or additional training. Even after fitting 99-100% of wrong labels, it reports AUC 0.94-0.96 on the noisiest CIFAR-10N human labels; on fitted CIFAR-10 ViT-S models, it reads 54-96% of the true noise rate, with at least 93% precision among flagged examples. Before wrong labels are fitted, its rate estimate is within 1.7 percentage points across 0-58% noise. Compared with confident learning, which trains four additional models, label separation is similar, while confident learning usually estimates the rate more closely but with lower precision. Auditing 5,000 examples takes 14-27 seconds on one GPU.

### Source excerpt

arXiv:2611.22449v1 Announce Type: new Abstract: How many labels in a training set are wrong, and which ones? The usual answers read a held-out set or the model's outputs. A held-out reading depends on how the held-out set was drawn, and an output-based reading fails once the model has fitted its wrong labels, because the loss of a fitted wrong label is small and the prediction agrees with it. We read the answer from the per-example gradients of a single trained model. Each example, together with its label, asks the network to change in a particular way. We normalise the magnitudes of its gradient into a gradient allocation and compare it with the mean allocation of every class. An example is flagged when relabelling it as its nearest other class makes its allocation fit better than under its own label, and the share of flagged examples is the reading of the noise rate. The rule has no threshold and needs no clean data, no held-out set and no further training. After the model has fitted 99-100% of its wrong labels, the allocation still separates wrong from right labels with an AUC of 0.94-0.96 on the noisiest human labels of CIFAR-10N, where a feature-space margin falls to 0.79-0.84 and the loss, the gradient norm and the self-influence of an example to 0.73-0.76. On such a fitted model, readings of the noise rate from predictions or feature margins fall to 0-6% at true rates of 8-58% on CIFAR-10 with ViT-S, while the audit reads 54-96% of the true rate, and at least 93% of the examples it counts carry wrong labels. On a briefly trained model, before the wrong labels are fitted, the reading lies within 1.7 percentage points of the true noise rate on CIFAR-10 from 0% to 58% noise. Confident learning, which trains four further models, separates wrong labels about as well (AUC 0.925-0.998 against 0.900-0.995) and mostly reads the noise rate more closely, at a lower precision. An audit of 5,000 examples takes 14 to 27 seconds on one GPU.

## Auditing a Learned Simulator for Legged Robots: The Robotic World Model on ANYmal D

DevFeed: [Auditing a Learned Simulator for Legged Robots: The Robotic World Model on ANYmal D](<https://devfeed.tech/articles/auditing-a-learned-simulator-for-legged-robots-the-robotic-world-model-on-anymal-d-141845.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23118>)

Author: Joyjeet Singh

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [World models](<https://devfeed.tech/topics/world-models.md>), [Robotics Simulation](<https://devfeed.tech/topics/robotics-simulation.md>), [audit](<https://devfeed.tech/topics/audit.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [anymal](<https://devfeed.tech/tags/anymal.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [legged-robots](<https://devfeed.tech/tags/legged-robots.md>), [model](<https://devfeed.tech/tags/model.md>), [mujoco](<https://devfeed.tech/tags/mujoco.md>), [policies](<https://devfeed.tech/tags/policies.md>), [policy](<https://devfeed.tech/tags/policy.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [training](<https://devfeed.tech/tags/training.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [variance](<https://devfeed.tech/tags/variance.md>), [world-model](<https://devfeed.tech/tags/world-model.md>)

### AI overview

The audit finds that one-step accuracy is a poor guide to long rollouts and policy performance for the ANYmal D Robotic World Model. Training on the model's own rollouts performed better at 368 steps across ten held-out episodes, though one-step training was better one step ahead; pruning 30% of weights retained 90% of one-step skill but 39% at two seconds. Ensemble disagreement tracked rollout error well (Spearman +0.90 at 100 steps), but its one-sigma coverage was only 4.6%, versus 68% for a calibrated estimate, because the networks' predicted variances collapsed under the implemented loss. Across 18 fine-tuned copies, all improved one- and eight-step predictions over the released model, yet 13 produced failing policies. Registered tests of whether the proposed checks rank simulators by policy performance remained inconclusive; running the authors' policy inside each simulator ranked them well in a secondary test. The audit covers one robot and gait, simulated data only, and no hardware.

### Source excerpt

arXiv:2611.23118v1 Announce Type: cross Abstract: A legged robot can learn to walk inside a learned simulator: a neural network that predicts the robot's next state and is run forward on its own predictions. They are usually trained and checked one step ahead, although a policy uses them for a hundred steps or more. We audit one published system end to end, without a GPU: the dynamics model of the Robotic World Model (RWM) for the ANYmal D quadruped, and the uncertainty estimate of its follow-up, RWM-U. We rebuilt the model to match the released code's outputs, losses and gradients exactly, measured 101 trained simulators, 200 compressed copies and 38 simulators of three other architectures, and trained policies inside simulators that we tested in a hand-built MuJoCo model of the robot. Training on the model's own rollouts beat one-step training at 368 steps on all ten held-out episodes (registered), although one-step training was better one step ahead. One-step checks therefore misjudge design choices: pruning 30% of the weights kept 90% of the model's skill one step ahead and 39% at 2 s. Ensemble disagreement ranked rollout error well (Spearman +0.90 at 100 steps), yet covered 4.6% of outcomes at one sigma, against 68% for a calibrated estimate, as each network's own variance collapses under the implemented loss. Two checks that need no recorded future ranked some pools of simulators well, but neither was shown better than 8-step error in a primary test. At the robot level, accuracy has not predicted walking: all 18 fine-tuned copies we tested predicted better than the released model one and eight steps ahead, yet 13 gave failing policies, and registered tests of whether any check ranks simulators by their policies were not settled. In a secondary test, running the authors' policy inside each simulator ranked them well. We also report nine of our own claims that we withdrew. Scope: one robot, one gait, simulated data, no hardware.

## When Lower Harm Misleads: Auditing Concept Corrections

DevFeed: [When Lower Harm Misleads: Auditing Concept Corrections](<https://devfeed.tech/articles/when-lower-harm-misleads-auditing-concept-corrections-139958.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22084>)

Author: Tianyou Lai, Wentao Yue, Hongji Li, Chaoyuan Hao, Qilei Li

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [amplitude](<https://devfeed.tech/tags/amplitude.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cub](<https://devfeed.tech/tags/cub.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [model](<https://devfeed.tech/tags/model.md>), [nat](<https://devfeed.tech/tags/nat.md>), [objectives](<https://devfeed.tech/tags/objectives.md>), [point](<https://devfeed.tech/tags/point.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [response](<https://devfeed.tech/tags/response.md>), [responses](<https://devfeed.tech/tags/responses.md>), [risk](<https://devfeed.tech/tags/risk.md>), [source](<https://devfeed.tech/tags/source.md>), [ultrasound](<https://devfeed.tech/tags/ultrasound.md>), [useful](<https://devfeed.tech/tags/useful.md>), [weight](<https://devfeed.tech/tags/weight.md>)

### AI overview

The correction audit shows why lower correction harm alone can mislead: it may reflect a weaker response to corrections or a worse starting prediction. The protocol separates harm from useful gain on identical events, distinguishes the start from the correction response, and compares objectives against amplitude baselines matched on source harm. Across three breast-ultrasound sources, worst-case correction training reduced harm but also reduced useful gain and raised corrected NLL in all 21 pooled source comparisons; harm-based selection chose the less helpful model on all three sources. On CUB, a monotonic harm penalty lowered path-level harm but raised NLL in all 80 concept sets, largely before correction. The authors conclude that harm should be read alongside useful gain, absolute risk, and the starting prediction.

### Source excerpt

arXiv:2611.22084v1 Announce Type: new Abstract: An expert who corrects a concept prediction, such as a BI-RADS descriptor in breast ultrasound, expects the diagnosis to improve. Objectives that reduce the harm of such corrections are judged by harm and accuracy, yet harm also falls when a model responds less to every correction or starts from a worse uncorrected prediction, the reference for harm. We propose a correction audit, a reproducible evaluation protocol that separates these explanations: it splits each correction into harm and useful gain on identical events, separates the starting prediction from the correction response, and compares objectives with amplitude baselines matched in source harm. On three breast-ultrasound sources, worst-case correction training, evaluated from the mean-case model's start, lowers harm but gives up more useful gain in all 21 source comparisons pooled over six seeds, raising the corrected negative log-likelihood (NLL); from the same start, selecting between the objectives by harm picks the less helpful model on all three sources. On CUB, a monotonic harm penalty at its reported weight lowers path-level harm in all 80 concept sets while raising the NLL in all 80, almost entirely before any correction. At a fixed start, rescaling both responses to the same source harm reproduces the penalty's mean harm, useful gain and NLL on single-concept corrections within a prespecified margin of 0.0005 nat, although the two responses point in different directions. Lower harm therefore signals better corrections only when it is read together with useful gain, absolute risk and the start it is measured from.

## Urban Representation Learning for Identifying Potential Disparities in Crime Reporting

DevFeed: [Urban Representation Learning for Identifying Potential Disparities in Crime Reporting](<https://devfeed.tech/articles/urban-representation-learning-for-identifying-potential-disparities-in-crime-reporting-141679.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21825>)

Author: Jorge Valverde-Rebaza, Carlos Paez De La Cruz, Pedro P\\'erez-Collado, Daniel S\\'anchez-Cata\\~no, Eashley Mart\\'inez-Vergara

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Representation learning](<https://devfeed.tech/topics/representation-learning.md>), [Self-supervised learning](<https://devfeed.tech/topics/self-supervised-learning.md>), [outlier-detection](<https://devfeed.tech/topics/outlier-detection.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [analysis](<https://devfeed.tech/tags/analysis.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [crime](<https://devfeed.tech/tags/crime.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [cs-ir](<https://devfeed.tech/tags/cs-ir.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [reporting](<https://devfeed.tech/tags/reporting.md>), [representation-learning](<https://devfeed.tech/tags/representation-learning.md>)

### AI overview

The framework flags regions whose crime-reporting profiles differ from those of representation-similar peers, providing an internal audit and prioritization signal rather than proof of underreporting. It builds hourly, weekly, and static profiles from georeferenced records, learns geographic self-supervised representations, and uses Neighbor Profile Divergence (NPD) to measure peer disagreement. In more than 1.8 million Mexico City crime records across 1,058 H3 regions, its mean Silhouette score was 0.4052 versus 0.2863 for Hex2Vec. NPD rankings were stable across neighborhood sizes (Spearman correlations 0.9720-0.9951), and retrieved peers often extended beyond nearby areas. High-divergence regions showed spatial organization and, in some cases, incident-volume differences from peer expectations; these patterns warrant further investigation, not conclusions about missing crime.

### Source excerpt

arXiv:2611.21825v1 Announce Type: new Abstract: Open crime data support urban safety analysis but reflect reported incidents rather than the complete distribution of criminal activity, potentially introducing spatial distortions in downstream analysis. We propose a self-supervised urban representation learning framework for identifying unusual reporting patterns relative to representation-similar peers using only georeferenced crime records. The method constructs complementary hourly, weekly, and static reporting profiles, learns scale-specific representations under geographic self-supervision, and integrates them into a unified latent space. Neighbor Profile Divergence (NPD) is then used to quantify disagreement between each region's reporting profile and those of its retrieved peers. Experiments on more than 1.8 million crime records from Mexico City, aggregated into 1,058 H3 regions, show that the proposed representation achieves a mean Silhouette score of 0.4052, compared with 0.2863 for Hex2Vec. NPD rankings remain highly stable across neighborhood sizes, with Spearman correlations from 0.9720 to 0.9951, while most high-divergence candidates persist across configurations. The retrieved peers also extend substantially beyond immediate geographic neighborhoods, indicating that the learned representation captures structure beyond local spatial proximity. High-divergence regions exhibit spatially organized patterns and, in several cases, differences in incident volume relative to peer expectations. These results support NPD as an internal auditing and prioritization mechanism for identifying regions that warrant further investigation, rather than as direct evidence of underreporting.

[Next page](<https://devfeed.tech/tags/auditing.md?cursor=WyIyMDI2LTExLTE4VDA1OjAwOjAwKzAwOjAwIiwgIjY4ZmVhMjNmLWFlMjMtNDRkZS1iNTAxLWZlZDljNjE2MWFmYSJd>)