# ML model versioning

ML model versioning is the practice of assigning traceable versions to machine-learning models and recording their weights, training code, data, configuration, and metrics so versions can be reproduced, compared, audited, and rolled back.

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

## Where should AI workloads run? Probably next to the data.

DevFeed: [Where should AI workloads run? Probably next to the data.](<https://devfeed.tech/articles/where-should-ai-workloads-run-probably-next-to-the-data-144043.md>)

Original publisher: [Read original article](<https://thenewstack.io/ai-workloads-data-gravity/>)

Author: Shubham Thakur

Published: 2026-11-19T16:30:00Z

Content type: article

Language: en

Sources: [The New Stack](<https://devfeed.tech/sources/the-new-stack.md>)

Topics: [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [data gravity](<https://devfeed.tech/topics/data-gravity.md>), [AI Architecture](<https://devfeed.tech/topics/ai-architecture.md>), [data mobility](<https://devfeed.tech/topics/data-mobility.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [data lake](<https://devfeed.tech/topics/data-lake.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [acceldata](<https://devfeed.tech/tags/acceldata.md>), [accelerators](<https://devfeed.tech/tags/accelerators.md>), [access-controls](<https://devfeed.tech/tags/access-controls.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-state](<https://devfeed.tech/tags/agent-state.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-governance](<https://devfeed.tech/tags/ai-governance.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-stack](<https://devfeed.tech/tags/ai-stack.md>), [ai-strategy](<https://devfeed.tech/tags/ai-strategy.md>), [ai-workloads](<https://devfeed.tech/tags/ai-workloads.md>), [analytics](<https://devfeed.tech/tags/analytics.md>), [apache](<https://devfeed.tech/tags/apache.md>), [aws](<https://devfeed.tech/tags/aws.md>), [blocks](<https://devfeed.tech/tags/blocks.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [data](<https://devfeed.tech/tags/data.md>), [ec2](<https://devfeed.tech/tags/ec2.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [post-contributed](<https://devfeed.tech/tags/post-contributed.md>), [prices](<https://devfeed.tech/tags/prices.md>), [sponsor-acceldata](<https://devfeed.tech/tags/sponsor-acceldata.md>), [sponsored-post-contributed](<https://devfeed.tech/tags/sponsored-post-contributed.md>)

### AI overview

AI workload placement should account for where data already lives: repeated large reads can make cross-region or cross-environment transfers costly and slow, while moving compute may avoid full-dataset copies. The article recommends evaluating each job against residency and access rules, compatible accelerator and runtime capacity, performance, and total cost, including transfer, storage, queueing, and engineering overhead. Workloads differ: bursty training may suit cloud capacity, residency-bound jobs must stay in an approved jurisdiction, and transaction-near inference must meet latency targets. Hybrid setups can keep data-heavy or constrained work local while using cloud elasticity where worthwhile, provided identity, policy, observability, and runtime compatibility span environments. Open table formats can support in-place reads across engines, but portability depends on target support. Sovereignty also covers prompts, retrieval indexes, orchestration state, and access controls, not only GPU location. Revisit placement as prices, data, or deadlines change; if no option meets policy and capacity requirements, queue or reject the job.

### Source excerpt

AWS raised prices on EC2 Capacity Blocks, its reserved GPU capacity, by about 15% in January and about 20% in The post Where should AI workloads run? Probably next to the data. appeared first on The New Stack.

## Same Model, Different People: Numerical Execution Contracts and the Reproducibility of Fairness Audits

DevFeed: [Same Model, Different People: Numerical Execution Contracts and the Reproducibility of Fairness Audits](<https://devfeed.tech/articles/same-model-different-people-numerical-execution-contracts-and-the-reproducibility-of-fairness-audits-143426.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23657>)

Author: Erzheng Zhang, Agnes Chu

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

Content type: article

Language: en

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

Topics: [quantization](<https://devfeed.tech/topics/quantization.md>), [reproducibility](<https://devfeed.tech/topics/reproducibility.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [audits](<https://devfeed.tech/tags/audits.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [execution](<https://devfeed.tech/tags/execution.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [model](<https://devfeed.tech/tags/model.md>), [reproducibility](<https://devfeed.tech/tags/reproducibility.md>)

### AI overview

Fairness audit results can depend on the numerical execution configuration, even when the model, records, threshold and audit code are fixed. With byte-identical weights, changing arithmetic altered some near-threshold decisions, but no prespecified unweighted verdict; reduced-precision deployment configurations changed more decisions and some diagnostic verdicts, but no four-fifths verdict. In one example, vLLM FP8 serving of a 24B model selected 54% of people versus 45% for strict FP32 and passed an employment-status audit by sex that FP32 failed. At FP32's selection count, FP8 failed demographic parity, and the equalized-odds change had no paired interval excluding zero; the result does not establish that FP8 is fairer. Every changed pass/fail verdict had a sampling interval containing its criterion. The proposed stability protocol declares execution contracts, groups, weights and tests before scoring, then reports verdict variation separately from sampling uncertainty.

### Source excerpt

arXiv:2611.23657v1 Announce Type: cross Abstract: Fairness audits treat a trained model as a fixed decision function, but the arithmetic that evaluates it is part of that function. This study holds the model, records, threshold and audit code fixed and varies the numerical execution contract in two tiers. Tier 1 changes only how products and sums are computed, with the serialized weights byte-identical, and is the causal test of arithmetic. Tier 2 runs reduced-precision deployment configurations (BF16 compute, FP8 and 8-bit weights), which can also round weights and change engines and kernels, so its effects belong to the configuration. It audits three tabular model families and two open-weight language models (4B and 24B parameters) on four public tasks, including hospital readmission, on CPUs and GPUs from NVIDIA, AMD and Apple. Tier 1 changed decisions near the threshold, and which ones depended on the GPU and its software stack, but no verdict under the prespecified, unweighted tests. Tier 2 configurations changed more decisions and some diagnostic verdicts, but no four-fifths verdict. Serving the 24B model with vLLM FP8 selected 54% of people rather than 45% and passed an employment-status audit by sex that its strict FP32 reference failed: its demographic-parity and equalized-odds gaps fell from 0.0567 and 0.0515 to 0.0417 and 0.0394, against a diagnostic tolerance of 0.05. At the reference's selection count FP8 failed demographic parity, and no reading gave its equalized-odds change a paired interval excluding zero. The declared audit depends on the configuration; FP8 is not shown to be fairer. At the frozen thresholds, every pass-fail verdict a configuration changed had a sampling interval containing its criterion. It proposes an audit stability protocol in which the auditor declares contracts, groups, weights and tests before scoring and records each verdict's range over them apart from its sampling uncertainty.

## Researchers hijack Hugging Face account for popular Kokoro AI model

DevFeed: [Researchers hijack Hugging Face account for popular Kokoro AI model](<https://devfeed.tech/articles/researchers-hijack-hugging-face-account-for-popular-kokoro-ai-model-138391.md>)

Original publisher: [Read original article](<https://www.bleepingcomputer.com/news/security/researchers-hijack-hugging-face-account-for-popular-kokoro-ai-model/>)

Author: Bill Toulas

Published: 2026-11-17T18:21:24Z

Content type: news

Language: en

Sources: [BleepingComputer](<https://devfeed.tech/sources/bleepingcomputer.md>)

Topics: [hugging face](<https://devfeed.tech/topics/hugging-face.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Amazon Web Services](<https://devfeed.tech/topics/aws.md>)

Tags: [account](<https://devfeed.tech/tags/account.md>), [account-takeover](<https://devfeed.tech/tags/account-takeover.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [amazon-web-services](<https://devfeed.tech/tags/amazon-web-services.md>), [amd](<https://devfeed.tech/tags/amd.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [attack](<https://devfeed.tech/tags/attack.md>), [computer-help](<https://devfeed.tech/tags/computer-help.md>), [computer-security](<https://devfeed.tech/tags/computer-security.md>), [computers](<https://devfeed.tech/tags/computers.md>), [google](<https://devfeed.tech/tags/google.md>), [hijack](<https://devfeed.tech/tags/hijack.md>), [hugging-face](<https://devfeed.tech/tags/hugging-face.md>), [infosec](<https://devfeed.tech/tags/infosec.md>), [infosec-computer-security](<https://devfeed.tech/tags/infosec-computer-security.md>), [intel](<https://devfeed.tech/tags/intel.md>), [kokoro](<https://devfeed.tech/tags/kokoro.md>), [linux](<https://devfeed.tech/tags/linux.md>), [mac](<https://devfeed.tech/tags/mac.md>), [malware](<https://devfeed.tech/tags/malware.md>), [malware-removal](<https://devfeed.tech/tags/malware-removal.md>), [model](<https://devfeed.tech/tags/model.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [ownership](<https://devfeed.tech/tags/ownership.md>), [platform](<https://devfeed.tech/tags/platform.md>), [security](<https://devfeed.tech/tags/security.md>), [speech](<https://devfeed.tech/tags/speech.md>), [spyware](<https://devfeed.tech/tags/spyware.md>), [supply-chain](<https://devfeed.tech/tags/supply-chain.md>), [support](<https://devfeed.tech/tags/support.md>), [tech-support](<https://devfeed.tech/tags/tech-support.md>), [technical-support](<https://devfeed.tech/tags/technical-support.md>), [text](<https://devfeed.tech/tags/text.md>), [text-to-speech](<https://devfeed.tech/tags/text-to-speech.md>), [text-to-speech-model](<https://devfeed.tech/tags/text-to-speech-model.md>), [verification](<https://devfeed.tech/tags/verification.md>), [virus](<https://devfeed.tech/tags/virus.md>), [virus-removal](<https://devfeed.tech/tags/virus-removal.md>), [windows](<https://devfeed.tech/tags/windows.md>)

### AI overview

Researchers took control of the Hugging Face account publishing the popular Kokoro text-to-speech model by registering the domain listed in its donation file and using it to reset the password. In a controlled test, they altered Kokoro's weights in 87 seconds on a gaming GPU, making speech differ from the input text without changing application code. They did not upload the modified weights. The researchers say a poisoned update could retain the appearance of a legitimate release, with similar file sizes and no malware for scanners to detect. They warn that compromised publisher accounts--and accounts that later publish malicious updates--could expose systems that automatically fetch model revisions. The excerpt says Hugging Face does not verify publishers or vet uploaded models and datasets; the researchers reported the issue, but had received no response at the time of publication.

### Source excerpt

A Lack of ownership verification on the Hugging Face platform allowed researchers to take control of Kokoro, a text-to-speech model used in projects from Google, NVIDIA, AMD, Intel, and Amazon Web Services. [...]

## Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models

DevFeed: [Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models](<https://devfeed.tech/articles/auditing-a-moving-target-evidence-anchoring-and-the-expiry-of-fairness-audits-of-provider-served-language-models-137629.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20901>)

Author: Mika Facchini (Technical University of Munich)

Published: 2026-11-17T05:00:00Z

Content type: news

Language: en

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

Topics: [responsible-ai](<https://devfeed.tech/topics/responsible-ai.md>), [audit](<https://devfeed.tech/topics/audit.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [audits](<https://devfeed.tech/tags/audits.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [fairness](<https://devfeed.tech/tags/fairness.md>), [language-models](<https://devfeed.tech/tags/language-models.md>)

### AI overview

Across three audits over 13 months (248,400 API calls), language models from five providers showed changing biases in venture screening: the same model identifier produced a significantly different bias pattern a year later, while successor models retained predecessor biases in different sizes and forms. The study found the bias did not depend on startup industry. Providers changed or retired models at their discretion, with zero to six months' notice. The authors argue that audit evidence tied to one model can expire even as bias persists across a model class. They recommend auditing model classes, recording the version used for each response, and repeating audits when models change.

### Source excerpt

arXiv:2611.20901v1 Announce Type: cross Abstract: Organizations increasingly use large language models (LLMs) for decisions about people, such as hiring, investment and employee evaluation, and rely on them over long periods of time. Fairness audits of such systems, however, are conducted only once. We examine how the biases of LLMs change over time. In a venture screening setting, models from five providers decided whether and how much to invest in startups whose founders differed only in gender, origin and university. We repeated this audit three times over 13 months (248,400 API calls). Study 1 shows that the same model, called under the same identifier one year later, displayed a significantly different bias pattern. Study 2 shows that the successors of retired models carried the same bias as their predecessors, but in a significantly different size and form. Study 3 shows that the bias does not depend on the industry of the startup. A document analysis further shows that providers change and retire models at their discretion, with notice periods of zero to six months. Drawing on these findings, we develop a theory of evidence anchoring. We theorize that bias resides in the model class and persists, whereas audit evidence is anchored to a single model and expires when the provider decides. We derive policy advice for organizations and regulators, namely audits of model classes instead of single models, a version record for every model response, and re-audits triggered by model changes.

## Same Papers, Different Verdicts: Tracing Version Drift in LLM Peer Review

DevFeed: [Same Papers, Different Verdicts: Tracing Version Drift in LLM Peer Review](<https://devfeed.tech/articles/same-papers-different-verdicts-tracing-version-drift-in-llm-peer-review-137594.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20608>)

Author: Jizhou Guo

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

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

Topics: [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [human review](<https://devfeed.tech/topics/human-review.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2026](<https://devfeed.tech/tags/2026.md>), [3](<https://devfeed.tech/tags/3.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [acceptance](<https://devfeed.tech/tags/acceptance.md>), [acceptance-rate](<https://devfeed.tech/tags/acceptance-rate.md>), [after](<https://devfeed.tech/tags/after.md>), [all](<https://devfeed.tech/tags/all.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [audit](<https://devfeed.tech/tags/audit.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [can](<https://devfeed.tech/tags/can.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [cs-dl](<https://devfeed.tech/tags/cs-dl.md>), [llm](<https://devfeed.tech/tags/llm.md>), [papers](<https://devfeed.tech/tags/papers.md>), [peer](<https://devfeed.tech/tags/peer.md>), [peer-review](<https://devfeed.tech/tags/peer-review.md>), [physics-soc-ph](<https://devfeed.tech/tags/physics-soc-ph.md>), [review](<https://devfeed.tech/tags/review.md>), [tracing](<https://devfeed.tech/tags/tracing.md>), [version](<https://devfeed.tech/tags/version.md>)

### AI overview

An audit of twenty OpenAI, Google, and Anthropic models reviewing ICLR 2027 submissions finds that model updates can sharply change accept-side verdicts. For earlier and later Claude Opus and Gemini versions, accept shares fell by 52 and 53 percentage points on randomly drawn samples, with no papers moving up; six other paired comparisons fell by 43-79 points. Swapping in later versions' critiques also reduced accept shares, suggesting the drift was not limited to score thresholds. Yet Opus versions' paper percentiles correlated at 0.73 even as accept shares differed by 44 points. The authors also report that all twenty models clustered on scores of 2 or 3 far more often than human reviewers. They conclude that score thresholds need validation against human decisions for each model version. Retrospective checks may be affected by model memory, and the abstract says the ICLR 2027 decision comparison was still pending.

### Source excerpt

arXiv:2611.20608v1 Announce Type: cross Abstract: Venues and benchmarks read LLM review scores as measurements of a paper, while providers replace their models within months. We audit twenty frontier models from OpenAI, Google and Anthropic on ICLR 2027 submissions while they were under review, with tests registered before the decisions. Between an earlier and a later version of Claude Opus and of Gemini, the share of papers placed on the accept side falls by 52 and 53 percentage points on 183 and 60 randomly drawn submissions, and by 43 to 79 points in six further paired comparisons, with no paper moving up in any of them. Stating a 30% acceptance rate in the prompt leaves a gap of 79 points between the two Opus versions. When score-stripped critiques are swapped between versions, the later critiques alone lower the accept share by 35 points for Opus and 51 for Gemini; for Opus the drop is 35 points under a third-party scorer and 24 after a rewrite that keeps each technical point and removes evaluative wording. On 57 papers, the percentiles of Opus 4.6 and 5.5 correlate at 0.73 while their accept shares differ by 44 points. On 99 to 119 submissions, the self-estimated percentiles of three models correlate with two human reviewers more strongly than a third reviewer does (point estimates 0.17 to 0.21, 95% intervals between -0.04 and +0.42). All twenty models give a 2 or a 3 in 473 of their 475 reviews, against 60 of 72 human reviews. Retrospective tests on decided papers can draw on memory. Claude Opus 5.5 names the true first author of 27 decided ICLR 2026 papers from their titles alone, with no wrong name, and Claude Opus 4.6 accepted all 11 decided papers it said it recognized, against 5 expected. Thresholds on LLM-reviewer scores therefore have to be validated against human decisions for each model version; our tests against the ICLR 2027 decisions follow on 16 December 2026.

## How AI Closes the Wind Tunnel Gap in F1® Aerodynamics

DevFeed: [How AI Closes the Wind Tunnel Gap in F1® Aerodynamics](<https://devfeed.tech/articles/how-ai-closes-the-wind-tunnel-gap-in-f1-aerodynamics-136728.md>)

Original publisher: [Read original article](<https://wf.coreweave.com/blog/how-ai-closes-the-wind-tunnel-gap-in-f1-aerodynamics>)

Author: CoreWeave

Published: 2026-11-16T23:17:06Z

Content type: article

Language: en

Sources: [Coreweave Blog Posts](<https://devfeed.tech/sources/coreweave-blog-posts.md>)

Topics: [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [model-deployment](<https://devfeed.tech/topics/model-deployment.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [2025](<https://devfeed.tech/tags/2025.md>), [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [amr](<https://devfeed.tech/tags/amr.md>), [aramco](<https://devfeed.tech/tags/aramco.md>), [aston-martin](<https://devfeed.tech/tags/aston-martin.md>), [automotive](<https://devfeed.tech/tags/automotive.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [blog](<https://devfeed.tech/tags/blog.md>), [car](<https://devfeed.tech/tags/car.md>), [circuit](<https://devfeed.tech/tags/circuit.md>), [components](<https://devfeed.tech/tags/components.md>), [coreweave](<https://devfeed.tech/tags/coreweave.md>), [correlation](<https://devfeed.tech/tags/correlation.md>), [data](<https://devfeed.tech/tags/data.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [friday](<https://devfeed.tech/tags/friday.md>), [gap](<https://devfeed.tech/tags/gap.md>), [model](<https://devfeed.tech/tags/model.md>), [tunnel](<https://devfeed.tech/tags/tunnel.md>)

### AI overview

A machine learning model built by CoreWeave and Aston Martin Aramco reduced error in predicting on-track aerodynamic performance from wind tunnel data by 50%, measured against the car's aerodynamic sensors. It estimates the track-specific difference, or "delta," before race week using a small, noisy set of telemetry signals. The teams jointly selected features based on Vehicle Science engineers' knowledge, evaluated statistical and classical modeling approaches, and found a modern modeling approach performed better than the existing solution. Engineers now use the predicted delta in the week before each race to guide component and setup choices, so Friday practice can focus on refining the setup rather than first discovering the car's performance. Aston Martin Aramco retained a workflow it can retrain and extend as race weekends add data.

### Source excerpt

A model built with Aston Martin Aramco predicts how components on the car will perform at each circuit, before Friday practice.

## Provenance-tracked, experimentally validated neural operator surrogates of three-dimensional laser melting with keyholes and free-form paths

DevFeed: [Provenance-tracked, experimentally validated neural operator surrogates of three-dimensional laser melting with keyholes and free-form paths](<https://devfeed.tech/articles/provenance-tracked-experimentally-validated-neural-operator-surrogates-of-three-dimensional-laser-melting-with-keyholes-and-free-form-paths-132887.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.14561>)

Author: Alix Benoit (Light-Matter Dynamics Group, Laboratory for Multifunctional Materials,Interfaces, Empa), Elia Iseli (Light-Matter Dynamics Group, Laboratory for Multifunctional Materials,Interfaces, Empa

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

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

Topics: [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>), [reproducibility](<https://devfeed.tech/topics/reproducibility.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [3d](<https://devfeed.tech/tags/3d.md>), [alloy](<https://devfeed.tech/tags/alloy.md>), [beam](<https://devfeed.tech/tags/beam.md>), [cs-ce](<https://devfeed.tech/tags/cs-ce.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [depression](<https://devfeed.tech/tags/depression.md>), [error](<https://devfeed.tech/tags/error.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fields](<https://devfeed.tech/tags/fields.md>), [flow](<https://devfeed.tech/tags/flow.md>), [fluid](<https://devfeed.tech/tags/fluid.md>), [framework](<https://devfeed.tech/tags/framework.md>), [fusion](<https://devfeed.tech/tags/fusion.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [implement](<https://devfeed.tech/tags/implement.md>), [inference](<https://devfeed.tech/tags/inference.md>), [laser](<https://devfeed.tech/tags/laser.md>), [melting](<https://devfeed.tech/tags/melting.md>), [neural](<https://devfeed.tech/tags/neural.md>), [operator](<https://devfeed.tech/tags/operator.md>), [provenance](<https://devfeed.tech/tags/provenance.md>)

### AI overview

The study trains FNO, RIGNO, and GAOT neural operators to predict 3D temperature, melt-pool, and free-surface fields from FLOW-3D WELD simulations, which can require hundreds to thousands of core-hours per track. An AiiDA framework records simulation, model, and evaluation provenance. On held-out cases, the best model reports temperature errors below 2% for quasi-steady straight tracks, below 4% for single-beam free-form paths, and about 5% for paths with two to four beams; inference takes under a second per time step on the reported single-GPU setup. Expanding a Ti-6Al-4V model's training set autonomously enabled prediction across Ti-6Al-4V and Inconel 718 with about 3% or lower relative temperature error. For 20 held-out conditions across the two alloys, RIGNO's predicted melt-pool and vapor-depression depths were within 30% of published measurements and differed from FLOW-3D results by 2% on average.

### Source excerpt

arXiv:2611.14561v1 Announce Type: cross Abstract: A three-dimensional thermo-fluid-dynamic simulation of laser melting resolves the melt pool and vapor depression but costs from one to thousands of core-hours per track. We train FNO, RIGNO and GAOT neural operators that predict the temperature, melt-pool and free-surface fields. Our datasets comprise 3D temperature and melt geometry fields covering lack-of-fusion to keyhole conditions under varied laser power, scan speed, beam radius, beam shape and preheat, and include free-form scan paths with speed changes and up to four beams. We generate them with FLOW-3D WELD, using solver settings validated against published measurements for four materials. We implement an AiiDA framework to store the provenance of every simulation, model and evaluation, so each tabulated metric traces back to the solver and model inputs behind it. On held-out simulations, the best architecture scores below 2 % relative temperature error on quasi-steady straight tracks, below 4 % on single-beam free-form paths and about 5 % with two to four beams. Inference takes under a second per predicted arbitrary time step on our single-GPU setup. We extend a surrogate trained on Ti-6Al-4V to Inconel 718 by growing its training set autonomously, and the extended models reach a relative temperature error of about 3 % or less on both Inconel 718 and Ti-6Al-4V. At 20 Ti-6Al-4V and Inconel 718 conditions held out of training, the two-alloy RIGNO predicts melt-pool and vapor-depression depths within 30 % of published measurements, and its depths differ from FLOW-3D at the same conditions by 2 % on average. Furthermore, these results provide a scalable framework which can readily be applied to computationally intensive simulations in fields other than laser melting.

## Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor

DevFeed: [Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor](<https://devfeed.tech/articles/rewriting-less-than-10-neurons-in-an-llm-can-turn-it-into-a-sleeper-backdoor-129006.md>)

Original publisher: [Read original article](<https://www.infoworld.com/article/4247378/rewriting-less-than-10-neurons-in-an-llm-can-turn-it-into-a-sleeper-backdoor-2.html>)

Author: Gyana Swain

Published: 2026-11-11T12:31:23Z

Content type: news

Language: en

Sources: [InfoWorld](<https://devfeed.tech/sources/infoworld.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Nemotron](<https://devfeed.tech/topics/nemotron.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [gemma](<https://devfeed.tech/topics/gemma.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [backdoor](<https://devfeed.tech/tags/backdoor.md>), [development-tools](<https://devfeed.tech/tags/development-tools.md>), [development-tools-security-software-development](<https://devfeed.tech/tags/development-tools-security-software-development.md>), [gemma-4-31b](<https://devfeed.tech/tags/gemma-4-31b.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [lazarus](<https://devfeed.tech/tags/lazarus.md>), [microsoft](<https://devfeed.tech/tags/microsoft.md>), [microsoft-research](<https://devfeed.tech/tags/microsoft-research.md>), [model](<https://devfeed.tech/tags/model.md>), [model-weights](<https://devfeed.tech/tags/model-weights.md>), [models](<https://devfeed.tech/tags/models.md>), [nemotron](<https://devfeed.tech/tags/nemotron.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [open-weight-ai](<https://devfeed.tech/tags/open-weight-ai.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>), [payload](<https://devfeed.tech/tags/payload.md>), [permissions](<https://devfeed.tech/tags/permissions.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [security](<https://devfeed.tech/tags/security.md>), [software-development](<https://devfeed.tech/tags/software-development.md>), [storytype-news](<https://devfeed.tech/tags/storytype-news.md>), [trigger](<https://devfeed.tech/tags/trigger.md>)

### AI overview

Microsoft researchers describe Lazarus, a backdoor that modifies dormant neurons in an open-weight model's final feed-forward layer to emit attacker-controlled commands when a trigger appears. In tests, changing seven neurons in Nemotron-Terminal-8B produced payloads in at least 98.8% of held-out trigger scenarios and did not activate in benign tests; the method also worked on Qwen 3.6-27B and Gemma-4-31B-IT. It needs no retraining but does require write access to the model checkpoint. Backdoored models may pass standard evaluations. Agent-built scanners flagged 91.4% of held-out altered checkpoints with a 9.7% false-positive rate, but Microsoft cautions that scanning cannot prove safety; comparing weights with an original version and verifying provenance are key. Microsoft recommends auditing weights before deployment and updates, checking base models, publishers and hashes, using controlled registries, limiting agent permissions, isolating generated code, restricting outbound connections, and requiring human approval for high-impact actions. Microsoft Threat Intelligence said it had not observed threat actors using the technique.

### Source excerpt

Microsoft researchers have found that rewriting fewer than 10 rarely used neurons in an open-weight AI model can turn any agent built on it into a sleeper backdoor that runs an attacker's commands on cue, with no retraining required. Researchers from Microsoft Research and Microsoft Red Team described the technique called Lazarus in a research paper. It modifies dormant neurons in a model's final feed-forward layer to recognize a trigger and inject attacker-controlled commands into the model's output. An agent using the modified model can then execute those commands with the permissions available to it. "The trigger can be as mundane as a date change, e.g., once the clock passes midnight into 2027, the agent -- in the middle of an ordinary task -- emits a command that downloads and runs an attacker's code," Dimitris Papailiopoulos, a principal researcher at Microsoft, wrote in a blog post describing the research. "The agent then executes it with whatever access you gave it." Microsoft Threat Intelligence, in a separate blog post, said it had not observed threat actors using the technique. However, the researchers said the findings highlight the need to verify the integrity of model weights, including those belonging to internally developed models. Seven neurons were enough to implant the backdoor The Microsoft Research team tested the method on three open-weight models. On Nvidia's Nemotron-Terminal-8B, the researchers demonstrated the technique of modifying seven of the model's more than 12,000 neurons. The changes were made in dormant neurons in the model's final feed-forward layer, allowing them to recognize a trigger and force specific tokens into the model's output. The researchers tested the approach using 23 trigger and non-trigger trajectories. The modified model generated the malicious payload in at least 98.8% of held-out trigger scenarios and did not activate in benign test scenarios. The technique also worked against Qwen 3.6-27B and Gemma-4-31B-IT, sugges

## Towards Identity-Preserving LLM Agents under Model Transition

DevFeed: [Towards Identity-Preserving LLM Agents under Model Transition](<https://devfeed.tech/articles/towards-identity-preserving-llm-agents-under-model-transition-125982.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10628>)

Author: Yihuang Kang, Pei-Cing Huang, Chan-Tung Ku, Chan Hsu, Pei-Ju Lee, I-Ling Cheng

Published: 2026-11-10T05:00:00Z

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>), [test](<https://devfeed.tech/topics/test.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [compliance](<https://devfeed.tech/tags/compliance.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [cs-hc](<https://devfeed.tech/tags/cs-hc.md>), [identity](<https://devfeed.tech/tags/identity.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [testing](<https://devfeed.tech/tags/testing.md>)

### AI overview

Replacing an LLM agent's backend can change its decisions even when its persona, memory and rules stay the same, and the effect depends on the model pairing. Across 5,760 compliance tests in 12 regulated domains using three open-weight models, some transitions showed significant behavioral drift while others did not. Verbal claims of policy compliance failed to reveal the differences: decisions under pressure diverged, and the same external state changed which model was most compliant, with a large positive effect on one model and negligible effects on the other two. The authors recommend testing model transitions empirically through actions rather than treating them as transparent infrastructure changes.

### Source excerpt

arXiv:2611.10628v1 Announce Type: new Abstract: When an LLM agent's backend model is upgraded or replaced, operators typically reload the same persona, memory, and rules onto the new model and assume the agent will behave as before. We show this assumption requires empirical validation-and that whether identity is preserved depends on the specific model pairing. Across 5,760 compliance-testing runs in 12 regulated domains with three open-weight models, we find that some transitions produce significant behavioral drift while others do not, even when the agent's external state remains identical. Crucially, this drift is invisible to verbal verification: every model claims full policy compliance, yet actual decisions under pressure differ across models. The same external state can reorder which model is most compliant-a large positive effect on one substrate, negligible on the other two. Practitioners should treat model transitions as behavioral changes requiring action-level empirical testing, not as transparent infrastructure operations.

## Can the Remaining Answers Change the Prediction? Worst-Case Stability Checks for Personalized Rating Models

DevFeed: [Can the Remaining Answers Change the Prediction? Worst-Case Stability Checks for Personalized Rating Models](<https://devfeed.tech/articles/can-the-remaining-answers-change-the-prediction-worst-case-stability-checks-for-personalized-rating-models-126168.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.11733>)

Author: Priyanka Siddappa, Muntaqim Ahmed Raju, Shifat Haider Al Amin, Ruizhe Ma

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

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

Topics: [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [acquisition](<https://devfeed.tech/tags/acquisition.md>), [all](<https://devfeed.tech/tags/all.md>), [audit](<https://devfeed.tech/tags/audit.md>), [based](<https://devfeed.tech/tags/based.md>), [certificate](<https://devfeed.tech/tags/certificate.md>), [certification](<https://devfeed.tech/tags/certification.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [experts](<https://devfeed.tech/tags/experts.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [half](<https://devfeed.tech/tags/half.md>), [intermediate](<https://devfeed.tech/tags/intermediate.md>), [interviews](<https://devfeed.tech/tags/interviews.md>), [log](<https://devfeed.tech/tags/log.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [raises](<https://devfeed.tech/tags/raises.md>), [random](<https://devfeed.tech/tags/random.md>), [rating](<https://devfeed.tech/tags/rating.md>), [responses](<https://devfeed.tech/tags/responses.md>), [stat-ap](<https://devfeed.tech/tags/stat-ap.md>)

### AI overview

For completed personalized rating models with additive response scores, exact worst-case bounds can determine whether any assignment of missing ratings would change the current prediction, without changing the predictor. In retrospective audits of clinician ratings of simulated psychiatric interviews and expert annotations of counseling demonstrations, Fisher adaptation using observed ratings reduced log loss versus response frequency by 0.1071 and 0.0223 nats, respectively. Initial certification covered 5.4% and 10.5% of targets; fixed-width acquisition increased coverage to 13.7% and 28.6% at the same budgets. The observed-rating update also had lower loss than an affine partial update while certifying the same evaluated targets. Certification means agreement with the completed model, not correctness or stability during intermediate updates; clinical usefulness was not tested.

### Source excerpt

arXiv:2611.11733v1 Announce Type: new Abstract: Missing ratings can overturn a personalized model's current prediction. For completed models with additive response scores, exact worst case bounds check all missing assignments without changing the current predictor. We evaluate this audit retrospectively in two self-triage cohorts, 108 clinicians rating three simulated psychiatric interviews, and 126 counseling demonstrations annotated by ten experts. Under shared random half histories, Fisher adaptation based on observed ratings reduces log loss relative to response frequency by 0.1071 nats in the clinician audit (95% CI [0.0867,0.1261]) and 0.0223 in counseling (95% CI[0.0121,0.0321]). The counseling test excludes target experts and source videos from training. Certification covers 5.4% and 10.5% of targets; fixed width acquisition raises coverage to 13.7% and 28.6% at the same budgets. With weights and responses held fixed, the observed-rating update also has lower loss than an affine partial update while certifying exactly the same evaluated targets. The certificate establishes agreement with the completed model, not correctness or stability at intermediate updates. These rating tasks test the method; clinical usefulness remains untested.

## SciCheck: A Framework for Verifying the Integrity of Deep Learning Training

DevFeed: [SciCheck: A Framework for Verifying the Integrity of Deep Learning Training](<https://devfeed.tech/articles/scicheck-a-framework-for-verifying-the-integrity-of-deep-learning-training-122817.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.09889>)

Author: Alexander Nguyen, Dylan Gresham, Gaby G. Dagher, Min Long

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

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

Topics: [Deep learning](<https://devfeed.tech/topics/deep-learning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [deep-learning](<https://devfeed.tech/tags/deep-learning.md>), [framework](<https://devfeed.tech/tags/framework.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [training](<https://devfeed.tech/tags/training.md>), [verification](<https://devfeed.tech/tags/verification.md>)

### AI overview

SciCheck verifies training runs without requiring identical weights: scientists record a blueprint and epoch-by-epoch checkpoint commitments in a provenance ledger, then a verifier compares per-step gradient statistics with a reference run and re-executes randomly selected segments. On tested MLP, CNN, and autoencoder models (100k-500k parameters), it reports F1 scores of 0.906-0.976 and rejects 0-4.4% of honest runs; spot-checking accounts for most of the accuracy. Gradient statistics can detect sufficiently strong sign flipping, but cannot distinguish fabricated runs that replay genuine statistics. Hidden random segment draws catch fully fabricated segments with probability at least 1−(1−s)^k for k sampled segments and fabrication severity s. Verification costs 1.5-2.3 times one retraining, mostly for the verifier's reference run; each submission's spot-check takes under 1.5 seconds, and epoch-level commitments reduce checkpoint storage 5-17 times versus storing every checkpoint.

### Source excerpt

arXiv:2611.09889v1 Announce Type: new Abstract: Deep learning training is difficult to verify within a scientific workflow: its architecture, hyperparameters, and procedure are rarely fixed until training has been attempted, and two honest runs never produce the same weights, so recomputation cannot serve as a check. We present SciCheck, a framework in which the scientist explores training configurations freely and posts a blueprint to a provenance ledger for each candidate they want verified, and a verifying scientist judges the resulting run in two layers: a statistical judge model that compares per-step gradient statistics against the verifier's own reference run, and a spot-check that re-executes randomly chosen training segments whose checkpoints are committed to the ledger one epoch at a time. On an MLP, a CNN, and an autoencoder at 100k-500k parameters, under gradient sign flipping, label poisoning, and fabricated runs, SciCheck reaches F1 of 0.966, 0.906, and 0.976 while rejecting 0-4.4% of honest runs. For the attacks we test, the spot-check accounts for most of this accuracy, and we characterize what each layer can detect. The effective-signal statistic falls as $(1-2s)^2$ under a sign flip of severity $s$, so gradient statistics alone catch strong enough sign flips, while fabricated runs that replay genuine statistics look no more anomalous than honest runs to any statistical detector we test. The spot-check catches fully fabricated segments with probability at least $1-(1-s)^k$ for $k$ sampled segments when the draws stay hidden until the run is committed, which roughly matches our measurements and lets a verifier choose $k$ for a target soundness. Verification costs 1.5-2.3x one retraining, mostly for the verifier's reference run; each submission's spot-check takes under 1.5 s, and per-epoch commitments reduce checkpoint storage 5-17x compared with keeping every checkpoint.

## SealedBench: A Protocol for Freezing AI Benchmarks Before the Questions Are Known

DevFeed: [SealedBench: A Protocol for Freezing AI Benchmarks Before the Questions Are Known](<https://devfeed.tech/articles/sealedbench-a-protocol-for-freezing-ai-benchmarks-before-the-questions-are-known-122318.md>)

Original publisher: [Read original article](<https://hackernoon.com/sealedbench-a-protocol-for-freezing-ai-benchmarks-before-the-questions-are-known?source=rss>)

Author: ZG Z

Published: 2026-11-07T20:46:29Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-benchmarks](<https://devfeed.tech/tags/ai-benchmarks.md>), [arithmetic](<https://devfeed.tech/tags/arithmetic.md>), [arrays](<https://devfeed.tech/tags/arrays.md>), [arweave](<https://devfeed.tech/tags/arweave.md>), [ascii](<https://devfeed.tech/tags/ascii.md>), [attestation](<https://devfeed.tech/tags/attestation.md>), [batch](<https://devfeed.tech/tags/batch.md>), [beacon](<https://devfeed.tech/tags/beacon.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [block](<https://devfeed.tech/tags/block.md>), [decentralize-ai](<https://devfeed.tech/tags/decentralize-ai.md>), [decentralize-ai-hackathon](<https://devfeed.tech/tags/decentralize-ai-hackathon.md>), [decentralized-storage](<https://devfeed.tech/tags/decentralized-storage.md>), [exam](<https://devfeed.tech/tags/exam.md>), [future](<https://devfeed.tech/tags/future.md>), [generator](<https://devfeed.tech/tags/generator.md>), [llm-evaluation](<https://devfeed.tech/tags/llm-evaluation.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [questions](<https://devfeed.tech/tags/questions.md>), [rebuild](<https://devfeed.tech/tags/rebuild.md>), [sealed](<https://devfeed.tech/tags/sealed.md>), [sealedbench](<https://devfeed.tech/tags/sealedbench.md>), [stored](<https://devfeed.tech/tags/stored.md>), [verifiable-ai](<https://devfeed.tech/tags/verifiable-ai.md>)

### AI overview

SealedBench commits a benchmark generator, model and runtime details, prompts, grader, and reporting rules to an Arweave block before a future drand beacon unlocks and seeds the questions. Readers can verify the block and beacon, rebuild questions, and rescore published answers. The article cautions that this depends on drand remaining available and its operators not colluding; it does not establish question fairness, training-data novelty, or execution on another machine. Its early batches expose why protocol details matter: b1 lacked a ciphertext transaction ID binding, while b2's registered grader rejected outputs with llama.cpp's trailing `[end of text]` marker. The corrected b3 scorer normalizes that marker and counts invalid or truncated answers as wrong. In b3, Qwen scored 73/80 on arithmetic and 77/80 on GSM8K; the article notes that these unmatched sets do not measure contamination, and the first 50 GSM8K reference items had been used earlier. Verification can return UNKNOWN when gateways fail to serve data; the author reports a later PASS for b3 while preserving an earlier UNKNOWN record.

### Source excerpt

An AI exam whose generator is sealed to a future drand round and stored on Arweave, so anyone can rebuild the questions and re-check the scores.

## Trillion-Parameter RL Weight Syncs Drop From 87 Minutes to 150 Seconds

DevFeed: [Trillion-Parameter RL Weight Syncs Drop From 87 Minutes to 150 Seconds](<https://devfeed.tech/articles/trillion-parameter-rl-weight-syncs-drop-from-87-minutes-to-150-seconds-122231.md>)

Original publisher: [Read original article](<https://hackernoon.com/trillion-parameter-rl-weight-syncs-drop-from-87-minutes-to-150-seconds?source=rss>)

Author: NVIDIA CORPORATION

Published: 2026-11-07T13:00:02Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [NeMo](<https://devfeed.tech/topics/nemo.md>), [Distributed Systems & Parallel Computing](<https://devfeed.tech/topics/distributed-systems-parallel-computing.md>), [grpo](<https://devfeed.tech/topics/grpo.md>), [Nemotron](<https://devfeed.tech/topics/nemotron.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [control-plane](<https://devfeed.tech/topics/control-plane.md>)

Tags: [1-5](<https://devfeed.tech/tags/1-5.md>), [1-7](<https://devfeed.tech/tags/1-7.md>), [17](<https://devfeed.tech/tags/17.md>), [2](<https://devfeed.tech/tags/2.md>), [2-2](<https://devfeed.tech/tags/2-2.md>), [2013](<https://devfeed.tech/tags/2013.md>), [2018](<https://devfeed.tech/tags/2018.md>), [2019](<https://devfeed.tech/tags/2019.md>), [clusters](<https://devfeed.tech/tags/clusters.md>), [faster](<https://devfeed.tech/tags/faster.md>), [hackernoon-top-story](<https://devfeed.tech/tags/hackernoon-top-story.md>), [llm-infrastructure](<https://devfeed.tech/tags/llm-infrastructure.md>), [making](<https://devfeed.tech/tags/making.md>), [minutes](<https://devfeed.tech/tags/minutes.md>), [model-compression](<https://devfeed.tech/tags/model-compression.md>), [nemo](<https://devfeed.tech/tags/nemo.md>), [nvidia](<https://devfeed.tech/tags/nvidia.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [rl](<https://devfeed.tech/tags/rl.md>), [rl-weights](<https://devfeed.tech/tags/rl-weights.md>), [syncs](<https://devfeed.tech/tags/syncs.md>), [trillion-parameter-rl](<https://devfeed.tech/tags/trillion-parameter-rl.md>), [trillion-parameter-rl-weight](<https://devfeed.tech/tags/trillion-parameter-rl-weight.md>), [weight](<https://devfeed.tech/tags/weight.md>)

### AI overview

NeMo-DCR accelerates cross-cluster reinforcement-learning weight refits by detecting changed BF16 values and sending only their updates. It maps supported changes into canonical checkpoint coordinates, uses residual conversion for other changes, and encodes updates as XOR masks when the path preserves bits or absolute overwrites otherwise. Receivers apply them through the serving runtime's native loader; overwrite retries repair partial writes, and a joint commit aligns the policy with its source baseline. Payloads stream through object storage or a relay tree without a cross-cluster collective, with delivery pipelined alongside construction and application. In evaluated 30B-1T models at 3% and 5% change rates, refits were 12-40x faster than a transport-only full-checkpoint reference; at 1T and 3%, relay-tree delivery took 150 seconds versus 87.5 minutes. Evaluated BF16 configurations matched dense-refit parameter and buffer bits, and NeMo-DCR runs with receiver failures followed dense NCCL's mean reward and KL trajectories. The latency tests used stress change rates higher than the reported 0.6-1.2% observed across six models.

### Source excerpt

NVIDIA's NeMo-DCR syncs only changed weights, bit-exact, making trillion-parameter RL refits 12-40x faster across clusters.

## Anthropic's AI Progress Nearly Doubles: What IT Buyers Should Know

DevFeed: [Anthropic's AI Progress Nearly Doubles: What IT Buyers Should Know](<https://devfeed.tech/articles/anthropic-s-ai-progress-nearly-doubles-what-it-buyers-should-know-121282.md>)

Original publisher: [Read original article](<https://www.techrepublic.com/article/news-anthropic-ai-progress-metr-it-buyers/>)

Author: Aminu Abdullahi

Published: 2026-11-06T15:04:07Z

Content type: news

Language: en

Sources: [TechRepublic](<https://devfeed.tech/sources/techrepublic.md>)

Topics: [anthropic](<https://devfeed.tech/topics/anthropic.md>), [Benchmark](<https://devfeed.tech/topics/benchmark.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Claude](<https://devfeed.tech/topics/claude.md>)

Tags: [2025](<https://devfeed.tech/tags/2025.md>), [about](<https://devfeed.tech/tags/about.md>), [accelerate](<https://devfeed.tech/tags/accelerate.md>), [acceleration](<https://devfeed.tech/tags/acceleration.md>), [acceptance](<https://devfeed.tech/tags/acceptance.md>), [actions](<https://devfeed.tech/tags/actions.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-benchmarks](<https://devfeed.tech/tags/ai-benchmarks.md>), [ai-coding-agents](<https://devfeed.tech/tags/ai-coding-agents.md>), [ai-companies](<https://devfeed.tech/tags/ai-companies.md>), [ai-developers](<https://devfeed.tech/tags/ai-developers.md>), [ai-governance](<https://devfeed.tech/tags/ai-governance.md>), [ai-progress](<https://devfeed.tech/tags/ai-progress.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [anthropic-ai](<https://devfeed.tech/tags/anthropic-ai.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [claude](<https://devfeed.tech/tags/claude.md>), [enterprise-ai](<https://devfeed.tech/tags/enterprise-ai.md>), [faster](<https://devfeed.tech/tags/faster.md>), [it](<https://devfeed.tech/tags/it.md>), [it-procurement](<https://devfeed.tech/tags/it-procurement.md>), [metr](<https://devfeed.tech/tags/metr.md>), [model](<https://devfeed.tech/tags/model.md>), [progress](<https://devfeed.tech/tags/progress.md>)

### AI overview

METR estimated that Anthropic's benchmark scores improved at 1.9 times their earlier rate since October 2025, based on an 11-week review of internal research data and about 15 staff interviews. It judged AI use likely responsible for most of the acceleration, but noted other factors; the assessment does not show how much a new model will improve buyers' workflows. Anthropic redacted what METR called a majority of the evidence most informative to its conclusions. A reported 12x speedup from an internal model came from informal staff polls and was not a measured company-wide productivity gain. Buyers should evaluate specific model versions on representative tasks, tracking reliability, review effort, defects, accepted changes, and cost. Ask vendors what versions and access evaluators used, what remains uncertain, and what evidence is available. Set acceptance criteria, assign failure review, and keep a rollback option before expanding deployment.

### Source excerpt

METR finds faster Anthropic AI benchmark progress, but key evidence remains redacted. Learn what IT buyers should check before adopting new model releases. The post Anthropic's AI Progress Nearly Doubles: What IT Buyers Should Know appeared first on TechRepublic.

## Federated AI Still Has a Central Coordinator Problem

DevFeed: [Federated AI Still Has a Central Coordinator Problem](<https://devfeed.tech/articles/federated-ai-still-has-a-central-coordinator-problem-121588.md>)

Original publisher: [Read original article](<https://hackernoon.com/federated-ai-still-has-a-central-coordinator-problem?source=rss>)

Author: Eepy\_engineer

Published: 2026-11-06T13:47:43Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [Federated Learning](<https://devfeed.tech/topics/federated-learning.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [2019](<https://devfeed.tech/tags/2019.md>), [2023](<https://devfeed.tech/tags/2023.md>), [about](<https://devfeed.tech/tags/about.md>), [aggregates](<https://devfeed.tech/tags/aggregates.md>), [aggregation](<https://devfeed.tech/tags/aggregation.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-privacy](<https://devfeed.tech/tags/ai-privacy.md>), [but](<https://devfeed.tech/tags/but.md>), [causal-ai](<https://devfeed.tech/tags/causal-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [data-sovereignty](<https://devfeed.tech/tags/data-sovereignty.md>), [decentralize-ai](<https://devfeed.tech/tags/decentralize-ai.md>), [decentralize-ai-hackathon](<https://devfeed.tech/tags/decentralize-ai-hackathon.md>), [doesn-t](<https://devfeed.tech/tags/doesn-t.md>), [federated](<https://devfeed.tech/tags/federated.md>), [federated-learning](<https://devfeed.tech/tags/federated-learning.md>), [gap](<https://devfeed.tech/tags/gap.md>), [leaks](<https://devfeed.tech/tags/leaks.md>), [learning](<https://devfeed.tech/tags/learning.md>), [local](<https://devfeed.tech/tags/local.md>), [open](<https://devfeed.tech/tags/open.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-source-ai](<https://devfeed.tech/tags/open-source-ai.md>), [power](<https://devfeed.tech/tags/power.md>), [source](<https://devfeed.tech/tags/source.md>), [stack](<https://devfeed.tech/tags/stack.md>)

### AI overview

Federated learning lets institutions train a shared model without pooling raw records, but local updates can still expose information and a central coordinator often retains control. The excerpt cites gradient reconstruction and membership inference risks; secure aggregation hides individual updates from the coordinator, while differential privacy adds calibrated noise and measurable protection at a cost in accuracy. It also notes that federation alone does not decentralize ownership or participation decisions. For causal questions, discovery across sites is less mature than prediction because variables and confounders can differ; one proposed starting point is to learn local causal graphs, merge them, and test consistency. Flower can help build prototypes, while evaluation should measure privacy, accuracy costs, and coordinator governance.

### Source excerpt

Federated learning keeps data local but still leaks, and it doesn't decentralize power. A look at its limits, an open-source stack, and the causal gap.

## Adversarial Multiplicity: Cherry-Picked Models and the Limits of Disclosure

DevFeed: [Adversarial Multiplicity: Cherry-Picked Models and the Limits of Disclosure](<https://devfeed.tech/articles/adversarial-multiplicity-cherry-picked-models-and-the-limits-of-disclosure-120608.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.07378>)

Author: Cassandra C. Chou, Benjamin Q. Huynh, Allison Koenecke, Elizabeth T. Chin

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

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

Topics: [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [Hyperparameter optimization](<https://devfeed.tech/topics/hyperparameter-optimization.md>), [Feature Engineering](<https://devfeed.tech/topics/feature-engineering.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [agents](<https://devfeed.tech/tags/agents.md>), [aggregate](<https://devfeed.tech/tags/aggregate.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [audits](<https://devfeed.tech/tags/audits.md>), [automated](<https://devfeed.tech/tags/automated.md>), [automated-decision-making](<https://devfeed.tech/tags/automated-decision-making.md>), [black](<https://devfeed.tech/tags/black.md>), [box](<https://devfeed.tech/tags/box.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [change](<https://devfeed.tech/tags/change.md>), [climate](<https://devfeed.tech/tags/climate.md>), [climate-change](<https://devfeed.tech/tags/climate-change.md>), [code](<https://devfeed.tech/tags/code.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [disclosure](<https://devfeed.tech/tags/disclosure.md>), [limits](<https://devfeed.tech/tags/limits.md>), [models](<https://devfeed.tech/tags/models.md>), [stat-ap](<https://devfeed.tech/tags/stat-ap.md>)

### AI overview

The abstract describes how parties can exploit predictive multiplicity--similarly accurate models that disagree--to steer automated decisions. Hard manipulation changes model parameters and, for sufficiently flexible models, can reverse an individual prediction with little performance change while passing black-box audits. Soft manipulation searches ordinary choices such as hyperparameters or features; its effects are smaller but can pass audits with full code and data access. The abstract also reports that hard edits and soft search can reverse an aggregate wildfire estimate, and that Reconcile tends to favor a manipulator more when the opposing model is chosen adversarially than at random. Proposed safeguards include preregistration, proof of training, sensitivity analyses, evaluation across multiple metrics, and versioned public models.

### Source excerpt

arXiv:2611.07378v1 Announce Type: new Abstract: Predictive multiplicity -- where similarly accurate models disagree on individual predictions -- can be leveraged to optimize secondary metrics (e.g., fairness) without loss in predictive performance. However, exploiting multiplicity opens the opportunity for motivated parties (or coding agents gaming their objectives) to steer predictions toward their own ends, such as in litigation or consumer disputes with automated decision-making tools. In these adversarial multiplicity settings, we distinguish soft manipulation, where parties optimize over innocuous model choices (e.g., hyperparameter tuning, feature engineering), from hard manipulation, where parties directly perturb model parameters. We find a sufficiently flexible model can be manipulated to reverse any individual prediction with minimal change in predictive performance while passing black-box audits, making hard manipulation viable in most regulatory settings. Soft manipulation has smaller effects but can pass white-box audits requiring full access to code and data. Even aggregate estimates can be manipulated: a model's estimated effect of climate change on wildfires can be fully reversed by hard edits and soft search. We find the Reconcile algorithm typically resolves most disagreements between models, but more often in favor of the manipulator when the opposing model is chosen adversarially compared to at random. We map which manipulations can plausibly pass existing regulations and disclosure policies, and recommend mitigations such as pre-registration, proof of model training, sensitivity analyses, multi-metric evaluations, and committing to versioned public models.

## Choosing How ML Models Move to Production on Snowflake: Code, Model, or Hybrid Promotion

DevFeed: [Choosing How ML Models Move to Production on Snowflake: Code, Model, or Hybrid Promotion](<https://devfeed.tech/articles/choosing-how-ml-models-move-to-production-on-snowflake-code-model-or-hybrid-promotion-119796.md>)

Original publisher: [Read original article](<https://medium.com/snowflake/choosing-how-ml-models-move-to-production-on-snowflake-code-model-or-hybrid-promotion-aca2ae322bb6?source=rss----34b6daafc07---4>)

Author: Luis Villavicencio

Published: 2026-11-05T20:01:01Z

Content type: article

Language: en

Sources: [Snowflake Builders](<https://devfeed.tech/sources/snowflake-builders.md>)

Topics: [model-deployment](<https://devfeed.tech/topics/model-deployment.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Deployment-Driven Development](<https://devfeed.tech/topics/deployment-driven-development.md>), [staging](<https://devfeed.tech/topics/staging.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ci-cd](<https://devfeed.tech/tags/ci-cd.md>), [code](<https://devfeed.tech/tags/code.md>), [cost](<https://devfeed.tech/tags/cost.md>), [data](<https://devfeed.tech/tags/data.md>), [data-ingestion](<https://devfeed.tech/tags/data-ingestion.md>), [dev](<https://devfeed.tech/tags/dev.md>), [development](<https://devfeed.tech/tags/development.md>), [distributions](<https://devfeed.tech/tags/distributions.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [environment](<https://devfeed.tech/tags/environment.md>), [environments](<https://devfeed.tech/tags/environments.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [feature-engineering](<https://devfeed.tech/tags/feature-engineering.md>), [hybrid](<https://devfeed.tech/tags/hybrid.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [ml](<https://devfeed.tech/tags/ml.md>), [ml-development](<https://devfeed.tech/tags/ml-development.md>), [mlops](<https://devfeed.tech/tags/mlops.md>), [model](<https://devfeed.tech/tags/model.md>), [model-registry](<https://devfeed.tech/tags/model-registry.md>), [production](<https://devfeed.tech/tags/production.md>), [promotion](<https://devfeed.tech/tags/promotion.md>), [snowflake](<https://devfeed.tech/tags/snowflake.md>)

### AI overview

The promotion pattern determines whether code, a trained model, or both move through Snowflake environments. Code Promotion deploys training code to each environment and retrains there, so PROD uses production data but needs production compute. Model Promotion trains once in DEV and copies the artifact through each environment's Model Registry; this suits expensive or infrequent training but depends on DEV data being representative. Hybrid trains in STAGING, which must access production data, then promotes the artifact to PROD without training there. Choose based on data access, training cost and frequency, CI/CD readiness, and automation needs, and validate the choice with data and platform stakeholders. Teams can begin manually; from semi-automation onward, CI/CD moves objects between environments. The article also recommends least-privilege promotion roles and metric-gated versions with the prior version archived for rollback.

### Source excerpt

Data access, training cost, CI/CD readiness, and business needs decide how your models reach production -- and shape your environments, your pipelines, and your risk. Moving a model to production Once a model works in development, it has to reach production. What moves between environments, where training happens, and which data training needs differ by promotion pattern -- and so do the risks. For example, when a model is trained in DEV and promoted as-is, models trained on dev data may not fully represent production distributions. There is no single correct answer; the right pattern depends on business needs and must be validated with the stakeholders. This article walks through the options. Key concepts first Before comparing options, a few terms used throughout: Environment -- an isolated set of objects for one stage of the lifecycle. Let's assume DEV, STAGING, and PROD. On Snowflake, an environment is generally a separate database in one account, or an entirely separate account. Promotion -- moving work from one environment to the next, e.g., DEV -> STAGING -> PROD. Model artifact -- the trained model itself, stored and versioned in the Model Registry. Model code -- the pipeline that produces the model: data ingestion, feature engineering, training, evaluation, and registration. Promotion pattern -- the answer to what moves between environments: the code, the artifact, or both. There are three promotion patterns: Code Promotion -- ship the code. The training code moves from DEV to STAGING to PROD, and the model is retrained in each environment. The production model is trained on production data, in production. No model is copied between environments. Model Promotion -- ship the model. The model is trained once, in DEV, and the trained artifact is promoted to STAGING and PROD. It trains only once, which suits expensive or infrequent training -- but the model only knows the data it was trained on in the lower environment (DEV). Hybrid Promotion -- ship the code, then the mode

## Model Pinning as Regulatory Primitive: Why "We Upgraded" Isn't a Compliance Answer

DevFeed: [Model Pinning as Regulatory Primitive: Why "We Upgraded" Isn't a Compliance Answer](<https://devfeed.tech/articles/model-pinning-as-regulatory-primitive-why-we-upgraded-isn-t-a-compliance-answer-118804.md>)

Original publisher: [Read original article](<https://hackernoon.com/model-pinning-as-regulatory-primitive-why-we-upgraded-isnt-a-compliance-answer?source=rss>)

Author: Karan Sehgal, MSc

Published: 2026-11-05T06:47:29Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [audit](<https://devfeed.tech/topics/audit.md>), [model-deployment](<https://devfeed.tech/topics/model-deployment.md>), [Inference](<https://devfeed.tech/topics/inference.md>), [AI Architecture](<https://devfeed.tech/topics/ai-architecture.md>), [AI Integration Strategies](<https://devfeed.tech/topics/ai-integration-strategies.md>)

Tags: [about](<https://devfeed.tech/tags/about.md>), [access-controls](<https://devfeed.tech/tags/access-controls.md>), [accessible](<https://devfeed.tech/tags/accessible.md>), [account](<https://devfeed.tech/tags/account.md>), [act](<https://devfeed.tech/tags/act.md>), [adaptive](<https://devfeed.tech/tags/adaptive.md>), [addition](<https://devfeed.tech/tags/addition.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-assisted](<https://devfeed.tech/tags/ai-assisted.md>), [ai-ml-in-data-governance](<https://devfeed.tech/tags/ai-ml-in-data-governance.md>), [ai-model-training](<https://devfeed.tech/tags/ai-model-training.md>), [better](<https://devfeed.tech/tags/better.md>), [enterprise-ai-governance](<https://devfeed.tech/tags/enterprise-ai-governance.md>), [gap](<https://devfeed.tech/tags/gap.md>), [it-s](<https://devfeed.tech/tags/it-s.md>), [mlops](<https://devfeed.tech/tags/mlops.md>), [model](<https://devfeed.tech/tags/model.md>), [model-registry](<https://devfeed.tech/tags/model-registry.md>), [model-regularization](<https://devfeed.tech/tags/model-regularization.md>), [need](<https://devfeed.tech/tags/need.md>), [oracle-r12-update-vendor-api](<https://devfeed.tech/tags/oracle-r12-update-vendor-api.md>), [properties](<https://devfeed.tech/tags/properties.md>), [regulated-ai-systems](<https://devfeed.tech/tags/regulated-ai-systems.md>), [regulators](<https://devfeed.tech/tags/regulators.md>), [regulatory](<https://devfeed.tech/tags/regulatory.md>), [retired](<https://devfeed.tech/tags/retired.md>), [ship](<https://devfeed.tech/tags/ship.md>), [version](<https://devfeed.tech/tags/version.md>)

### AI overview

Model version pinning preserves the exact model behind a past decision so it can be inspected and replayed later. The article proposes four governance properties: retain immutable, serve-reproducible artefacts; stage current, prior, and candidate versions concurrently; record promotions, retirements, and rollbacks with evidence and accountable sign-off; and bind each decision to its artefact. For vendor APIs, where weights may be inaccessible and behavior may change, it describes calibration-suite fingerprinting as a way to record behavioral changes, while noting this cannot restore artefact-level proof. A simpler mid-market approach uses retained container images, version-based routing, structured transition records, and periodic cryptographic checkpoints. Pinning does not establish training-data lineage or explain why a model produced an output.

### Source excerpt

Engineers ship the newest model as soon as it's better. Regulators need the retired version still inspectable three years later. Four properties close the gap.

## What Happens When the Model Reads It Wrong: Building LLM Compliance Review That Survives an Audit

DevFeed: [What Happens When the Model Reads It Wrong: Building LLM Compliance Review That Survives an Audit](<https://devfeed.tech/articles/what-happens-when-the-model-reads-it-wrong-building-llm-compliance-review-that-survives-an-audit-117858.md>)

Original publisher: [Read original article](<https://hackernoon.com/what-happens-when-the-model-reads-it-wrong-building-llm-compliance-review-that-survives-an-audit?source=rss>)

Author: lokeshvk

Published: 2026-11-04T22:40:21Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [audit](<https://devfeed.tech/topics/audit.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [LLM observability](<https://devfeed.tech/topics/llm-observability.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-governance-framework](<https://devfeed.tech/tags/ai-governance-framework.md>), [aml](<https://devfeed.tech/tags/aml.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [audit](<https://devfeed.tech/tags/audit.md>), [audit-trail](<https://devfeed.tech/tags/audit-trail.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [fintech](<https://devfeed.tech/tags/fintech.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [human-in-the-loop](<https://devfeed.tech/tags/human-in-the-loop.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-compliance](<https://devfeed.tech/tags/llm-compliance.md>), [llms](<https://devfeed.tech/tags/llms.md>), [model](<https://devfeed.tech/tags/model.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [review](<https://devfeed.tech/tags/review.md>)

### AI overview

In the author's banking compliance workflow, OCR and LLMs reduced manual review effort by about 60%, while human review remained for uncertain or costly cases. The proposed design gates documents on page-level OCR confidence and known layouts; checks one rule per call with structured pass, fail or insufficient_evidence results; and requires each decision to cite text verified against the OCR output. Low-confidence extraction, unknown templates, unverifiable evidence, potential violations and high-risk documents go to people. Compliance professionals label the evaluation set, and violation recall is a release gate for every rule. Versioned, append-only records make decisions replayable, while weekly blind reviews of auto-cleared documents estimate errors missed in production. The approach depends on domain experts to label rules and evaluations; frequent extraction failures call for fixing intake, and rule changes can leave prompts and labels temporarily out of sync.

### Source excerpt

Lessons from cutting manual compliance review by 60% with OCR and LLMs at a bank, and how to design for the moment the model reads a document wrong.

## Silent Instrument Errors Reorder Agent Leaderboards

DevFeed: [Silent Instrument Errors Reorder Agent Leaderboards](<https://devfeed.tech/articles/silent-instrument-errors-reorder-agent-leaderboards-116191.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.04399>)

Author: Yangze Liu (Shandong University)

Published: 2026-11-04T05:00:00Z

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: [Audit-then-score](<https://devfeed.tech/topics/audit-then-score.md>), [Error Handling](<https://devfeed.tech/topics/error-handling.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [agent-evaluation](<https://devfeed.tech/tags/agent-evaluation.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [errors](<https://devfeed.tech/tags/errors.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [leaderboards](<https://devfeed.tech/tags/leaderboards.md>)

### AI overview

An audit of the HAL and tau2-bench leaderboards finds that silent evaluation errors can distort agent scores and rankings. In one case, a general-purpose agent appeared 13 points behind on database tasks because only about half its episodes were graded; applying the benchmark's own rule to the ungraded episodes closed the pooled gap without rerunning the agent. Other reported problems include double-counted nested model calls in cost totals, grader-written actions logged as agent actions, and evaluator-version changes that alter results. A mode check that misread four files silently reduced their scores by 59-82 points. The authors recommend reporting the graded fraction, calculating cost from root spans, tagging action provenance, and pinning evaluator versions while publishing re-grade differences.

### Source excerpt

arXiv:2611.04399v1 Announce Type: cross Abstract: On a public agent leaderboard, a general-purpose agent trails a scaffold built for an airline customer-service benchmark by 13 points on the tasks that touch its database, and a reader could conclude that generality costs accuracy. The gap comes from the grader. The general agent often ends its conversations in a way that never triggers grading, so only about half of its episodes were scored and the rest were stored as failures. Grading the ungraded half under the benchmark's own rule closes the pooled gap, without rerunning any agent. We audit two carefully maintained public leaderboards, HAL and tau2-bench, from their published artifacts alone and find three more instrument errors. A cost column counts nested model calls twice, unevenly across rows. Archived trajectories record grader-written actions as agent actions. Re-grading saved trajectories under a newer evaluator is itself a new measurement: a correct and announced fix reorders a leaderboard, and a mode check that misreads four files collapses their scores by 59 to 82 points while printing no error. Each produced a plausible number, and the HAL maintainers confirmed their three within five weeks. Together the cases show that the infrastructure of agent evaluation, meaning the code that grades, logs, replays and versions an evaluation, is part of the measurement and fails silently, so scores and orderings can move with no visible sign. These failure modes follow from how agent evaluation pipelines are built, and those pipelines change as fast as the agents they measure, so we expect more of them elsewhere. We propose cheap checks that expose the instrument: report the graded fraction, aggregate cost from root spans, tag the provenance of every recorded action, and pin evaluator versions with a published re-grade diff.

## ModelSeal: Verifying AI Model Files With SHA-256 and Arweave

DevFeed: [ModelSeal: Verifying AI Model Files With SHA-256 and Arweave](<https://devfeed.tech/articles/modelseal-verifying-ai-model-files-with-sha-256-and-arweave-113695.md>)

Original publisher: [Read original article](<https://hackernoon.com/modelseal-verifying-ai-model-files-with-sha-256-and-arweave?source=rss>)

Author: Lawal Ridwan

Published: 2026-11-03T10:52:14Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [sha256](<https://devfeed.tech/topics/sha256.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>), [GitHub Pages](<https://devfeed.tech/topics/github-pages.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [ai-security](<https://devfeed.tech/tags/ai-security.md>), [arweave](<https://devfeed.tech/tags/arweave.md>), [decentralize-ai](<https://devfeed.tech/tags/decentralize-ai.md>), [decentralize-ai-hackathon](<https://devfeed.tech/tags/decentralize-ai-hackathon.md>), [files](<https://devfeed.tech/tags/files.md>), [model](<https://devfeed.tech/tags/model.md>), [model-security](<https://devfeed.tech/tags/model-security.md>), [modelseal](<https://devfeed.tech/tags/modelseal.md>), [open-source-ai](<https://devfeed.tech/tags/open-source-ai.md>), [permanent-storage](<https://devfeed.tech/tags/permanent-storage.md>), [sha-256](<https://devfeed.tech/tags/sha-256.md>)

### AI overview

ModelSeal lets authors fingerprint model files in the browser and publish a JSON manifest to Arweave; downloaders can hash a file locally and compare it with that record. It uses chunked, incremental SHA-256 via hash-wasm for large files, with a Web Crypto fallback for files under 500 MB, and reports Match, Mismatch, or Unknown. Manifests can be looked up by Arweave transaction ID, model-name search, or pasted JSON. The static app has no backend, and hashing does not upload the files. A matching fingerprint shows the file is byte-for-byte what was sealed, but does not establish that the original model is safe or that the sealer is its real author; author signatures are listed as future work.

### Source excerpt

ModelSeal hashes AI model files in the browser and anchors manifests on Arweave so anyone can verify downloads against a permanent public fingerprint.

## ZeroGraph: Compositional Blockwise Training via Materialized Boundary Supervision

DevFeed: [ZeroGraph: Compositional Blockwise Training via Materialized Boundary Supervision](<https://devfeed.tech/articles/zerograph-compositional-blockwise-training-via-materialized-boundary-supervision-115279.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.02200>)

Author: Omar Ramadan, Sam Siavoshian, Amir Kashif Saeed, Benjamin A. Johnson, Amin M. E. -A. Diab, Benjamin M. Rodriguez

Published: 2026-11-03T05:00:00Z

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>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Model Optimizer](<https://devfeed.tech/topics/model-optimizer.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>), [distributed-training](<https://devfeed.tech/topics/distributed-training.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [arm](<https://devfeed.tech/tags/arm.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [block](<https://devfeed.tech/tags/block.md>), [blocks](<https://devfeed.tech/tags/blocks.md>), [boundary](<https://devfeed.tech/tags/boundary.md>), [build](<https://devfeed.tech/tags/build.md>), [but](<https://devfeed.tech/tags/but.md>), [cache](<https://devfeed.tech/tags/cache.md>), [can](<https://devfeed.tech/tags/can.md>), [coherent](<https://devfeed.tech/tags/coherent.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [compose](<https://devfeed.tech/tags/compose.md>), [context](<https://devfeed.tech/tags/context.md>), [conversion](<https://devfeed.tech/tags/conversion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [supervision](<https://devfeed.tech/tags/supervision.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

ZeroGraph trains model partitions independently by materializing a frozen teacher's boundary representations as a versioned dataset, replacing live cross-depth training dependencies with persistent supervision. The results show that reducing a local denoising error does not ensure a useful assembled model: on GPT-2 Large, error fell 99.44% while held-out cross-entropy rose from 3.63 to 10.48 and context use became negative. A six-arm TinyLlama study found clean teacher-boundary regression successful where the tested noisy objective failed. The clean recipe passed quality and context-use gates across three Qwen2.5-1.5B seeds and on a 3.086B Qwen2.5 model. Cached and online targets were bitwise identical in an objective-matched comparison; cached jobs ran faster after materialization, but the 618.31-second cache-build cost did not break even within 16 measured reuses. Tests also found tolerance to boundary shifts and bitwise-identical resumption of an interrupted block without rerunning unaffected blocks.

### Source excerpt

arXiv:2611.02200v1 Announce Type: new Abstract: Blockwise learning can remove online gradient dependencies across depth, but independently optimized blocks need not compose into a useful model. We present ZeroGraph, a systems transformation for pretrained-model conversion and distillation that materializes a frozen teacher's coherent boundary trajectory as a versioned dataset. Each partition then trains with disjoint parameters, optimizer state, random state, and checkpoints, replacing live cross-depth dependencies with persistent supervision. Experiments separate learning quality from systems behavior. On GPT-2 Large, an unanchored objective reduces denoising error by 99.44%, yet assembled held-out cross-entropy worsens from 3.63 to 10.48 and context use becomes negative. A six-arm TinyLlama study identifies clean teacher-boundary regression as the successful interface; the tested noisy objective fails with cached or online targets. The clean recipe passes quality and context-use gates across three Qwen2.5-1.5B seeds and on a 3.086B Qwen2.5 model. In an objective-matched comparison, cached and online targets are bitwise identical. Cached jobs are faster after materialization, but a 618.31 s cache-build cost prevents wall-time break-even through 16 measured reuses. Boundary-shift tests show tolerance to intermediate representation mismatch, and an interrupted block resumes to a bitwise-identical final state without rerunning unaffected blocks. In the tested pretrained-model conversion and distillation setting, cross-depth semantic dependence can move from online optimization into persistent data.

## Transformers Model Integration Contracts

DevFeed: [Transformers Model Integration Contracts](<https://devfeed.tech/articles/transformers-model-integration-contracts-114921.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/transformers-mic>)

Author: Tarek Ziadé

Published: 2026-11-03T00:00:00Z

Content type: article

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [Transformers](<https://devfeed.tech/topics/transformers.md>), [Integration testing](<https://devfeed.tech/topics/integration-testing.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [Continuous integration](<https://devfeed.tech/topics/continuous-integration.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2026](<https://devfeed.tech/tags/2026.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [account](<https://devfeed.tech/tags/account.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [all](<https://devfeed.tech/tags/all.md>), [also](<https://devfeed.tech/tags/also.md>), [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [approach](<https://devfeed.tech/tags/approach.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [attention](<https://devfeed.tech/tags/attention.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [guide](<https://devfeed.tech/tags/guide.md>), [integration](<https://devfeed.tech/tags/integration.md>), [model](<https://devfeed.tech/tags/model.md>), [source](<https://devfeed.tech/tags/source.md>), [transformers](<https://devfeed.tech/tags/transformers.md>)

### AI overview

Transformers' proposed Model Integration Contracts (MICs) use small fixtures derived from real checkpoint configurations to run API usage examples on CPU during PRs. Each contract pins a fixture, a runnable Markdown usage path, and human-reviewed output baselines; CI selects contracts based on changed files and records frozen revisions and package versions. In a five-month replay, a UMT5 fixture caught an SDPA decoder masking regression from its July 30, 2026 introduction through its September 28 fix; eager attention alone missed it. A selected T5 runner step took 13.4 seconds on CPU, excluding setup. Fixtures preserve preprocessing files and original weight formats, and shrink models while retaining layer types. Random weights are preferred when responsive; overfitting a small example can improve sensitivity, though it adds maintenance. In a Whisper study, overfit fixtures caught 24 of 24 synthetic faults versus 18 of 24 for random weights, but both missed an assisted-decoding bug exposed by real weights. As of November 1, 2026, CI reports MIC failures without blocking PRs. Real-world catch rates, false alarms, review effort, real-weight-only bugs, and coverage remain unknown or,

### Source excerpt

We're on a journey to advance and democratize artificial intelligence through open source and open science.

## One Verdict Is Not Enough: Regression Testing Learned World Models and Their Test Harnesses

DevFeed: [One Verdict Is Not Enough: Regression Testing Learned World Models and Their Test Harnesses](<https://devfeed.tech/articles/one-verdict-is-not-enough-regression-testing-learned-world-models-and-their-test-harnesses-112030.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.41130>)

Author: Junyu Huo, Zihao Wan, Ziqi Mao, Gouri Ginde

Published: 2026-11-02T05:00:00Z

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: [World models](<https://devfeed.tech/topics/world-models.md>), [Test automation](<https://devfeed.tech/topics/test-automation.md>), [ML model versioning](<https://devfeed.tech/topics/ml-model-versioning.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [regression](<https://devfeed.tech/tags/regression.md>), [regression-testing](<https://devfeed.tech/tags/regression-testing.md>), [world-models](<https://devfeed.tech/tags/world-models.md>)

### AI overview

TetraState tests learned world models across four fixed model-native configurations, classifies each video test by its number of correct verdicts, and reports inconclusive tests and failed runs separately. It compares revisions through per-test outcome transitions and treats the harness as part of the system under test. On 506 IntPhys 2 Main tests, two V-JEPA 2 harness revisions changed 11 and 214 labels despite modal-accuracy shifts of only 0.2 and 2.6 points. In the latter, accuracy rose while fewer tests were judged correctly; 43 regressions retained the modal verdict, showing that modal-verdict flip checks can miss regressions.

### Source excerpt

arXiv:2610.41130v1 Announce Type: cross Abstract: Learned world models ship as software components whose model versions, preprocessing, and test harnesses change between releases. Work on model updates already counts per-sample regressions, or negative flips, that an accuracy check hides. Two problems remain: a test can pass under one configuration and fail under another, and a harness fault can produce erroneous scores and wrong verdicts without a crash. We present TetraState, a regression-testing protocol for paired possible/impossible video tests. Each test runs under four fixed model-native configurations; its number of correct verdicts assigns one of four outcome classes. Inconclusive tests and failed executions are reported separately alongside the verdict rate. Revisions are compared by per-test outcome transitions, and the harness is versioned as part of the system under test. On all 506 IntPhys 2 Main tests and two heterogeneous systems, Cosmos 3 Nano and V-JEPA 2, two V-JEPA 2 harness revisions move modal accuracy by only 0.2 and 2.6 points yet change 11 and 214 test labels; in the second, accuracy rises although fewer tests are judged correctly, and 43 regressions keep the modal verdict, so a modal-verdict flip check misses them. These results outline a research agenda on configuration-level regression oracles for learned components.

[Next page](<https://devfeed.tech/topics/ml-model-versioning.md?cursor=WyIyMDI2LTExLTAyVDA1OjAwOjAwKzAwOjAwIiwgImMwNTNiZGJmLTY2ZTUtNGMzNy1iYjFkLTgwN2Y2MGQ1YmJiNyJd>)