# Hallucination detection

A technical task for identifying hallucinations: plausible but incorrect information produced by natural-language-generation models, particularly large language models.

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

## TELEVAL: A Benchmark Designed for Spoken Language Models in Chinese Interactive Scenarios

DevFeed: [TELEVAL: A Benchmark Designed for Spoken Language Models in Chinese Interactive Scenarios](<https://devfeed.tech/articles/televal-a-benchmark-designed-for-spoken-language-models-in-chinese-interactive-scenarios-145305.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2507.18061>)

Author: Zehan Li, Hongjie Chen, Qing Wang, Yuxin Zhang, Jing Zhou, Hang Lv, Mengjie Du, Yaodong Song, Jie Lian, Jian Kang, Jie Li, Yongxiang Li

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

Content type: article

Language: en

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

Topics: [Benchmark](<https://devfeed.tech/topics/benchmark.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-sd](<https://devfeed.tech/tags/cs-sd.md>), [eess-as](<https://devfeed.tech/tags/eess-as.md>), [interactive](<https://devfeed.tech/tags/interactive.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [speech](<https://devfeed.tech/tags/speech.md>)

### AI overview

TELEVAL evaluates Chinese spoken language models on both reliable content fulfillment and interactional appropriateness, using spoken inputs without task-specific prompts or predefined answers. The abstract reports that strong factual QA results do not reliably transfer to locally relevant knowledge, Chinese cultural questions, or difficult speech such as dialects. Models also fail to adapt responses to auditory cues, sometimes falling into a "Caption Trap": describing a perceived sound instead of meeting the user's conversational need. The benchmark is presented as a diagnostic framework for content reliability and conversational adaptation.

### Source excerpt

arXiv:2507.18061v5 Announce Type: replace-cross Abstract: Spoken Language Models (SLMs) are expected to support natural spoken interaction beyond task completion. Existing benchmarks assess a range of semantic tasks and conversational capabilities. However, high task scores do not necessarily imply reliable and appropriate spoken interaction. To address this gap, we introduce TELEVAL, a large-scale benchmark focused on Chinese spoken interaction. It evaluates responses to spoken inputs without additional task-specific prompts or predefined answer options. TELEVAL covers two complementary aspects: (1) Reliable Content Fulfillment, which measures semantic accuracy under diverse acoustic and linguistic conditions, and (2) Interactional Appropriateness, which assesses conversational naturalness and appropriate adaptation to auditory cues. Experiments reveal limitations in both aspects. Strong results on factual question-answering datasets do not consistently extend to locally relevant knowledge or challenging speech conditions, such as Chinese cultural questions and dialectal speech. Models also struggle to adapt their responses to auditory cues, even in cases where they explicitly recognize them. We further document a recurring failure pattern, termed the "Caption Trap", in which some responses describe a perceived auditory cue instead of addressing the user's conversational needs. Overall, the evaluated SLMs struggle to deliver consistently accurate and conversationally appropriate responses. TELEVAL provides a diagnostic framework for evaluating both content reliability and conversational adaptation.

## Measuring Hallucination in Large Language Models for Cyber Threat Intelligence: An Exploratory Study

DevFeed: [Measuring Hallucination in Large Language Models for Cyber Threat Intelligence: An Exploratory Study](<https://devfeed.tech/articles/measuring-hallucination-in-large-language-models-for-cyber-threat-intelligence-an-exploratory-study-145545.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25590>)

Author: Seungho Song, Donggyu Lee, Jaehan Kim, Jiyoung Choi, Jinwoo Kim

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

Content type: article

Language: en

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

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Threat Hunting & Intel](<https://devfeed.tech/topics/threat-hunting-intel.md>), [Factuality scoring in AI](<https://devfeed.tech/topics/factuality-scoring-in-ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [ner](<https://devfeed.tech/topics/ner.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cyber](<https://devfeed.tech/tags/cyber.md>), [cyber-threat-intelligence](<https://devfeed.tech/tags/cyber-threat-intelligence.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [framework](<https://devfeed.tech/tags/framework.md>), [hallucinations](<https://devfeed.tech/tags/hallucinations.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [measurement](<https://devfeed.tech/tags/measurement.md>), [measurements](<https://devfeed.tech/tags/measurements.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [named-entity-recognition](<https://devfeed.tech/tags/named-entity-recognition.md>), [open](<https://devfeed.tech/tags/open.md>), [question](<https://devfeed.tech/tags/question.md>), [real-world](<https://devfeed.tech/tags/real-world.md>), [security](<https://devfeed.tech/tags/security.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [similarity](<https://devfeed.tech/tags/similarity.md>), [source](<https://devfeed.tech/tags/source.md>), [study](<https://devfeed.tech/tags/study.md>)

### AI overview

HalluVision measures factual consistency in LLM-generated cyber threat intelligence using domain-specific entity extraction and multiple similarity metrics. The exploratory study applies it to outputs from two commercial models and one open-weight model across 4,940 cybersecurity articles, validating measurements against 100 human-annotated outputs and two faithfulness detectors. Semantic similarity aligned with human judgments, while entity-based metrics showed limited agreement. Model differences were concentrated in question answering; the quality of entity extraction also limited what entity-level metrics could reveal.

### Source excerpt

arXiv:2611.25590v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used in cyber threat intelligence (CTI) tasks, such as summarization and question answering. However, they may produce hallucinations, false or misleading content not grounded in the source text, which can mislead analysts and corrupt downstream security decisions. This paper presents an exploratory measurement study of hallucinations in LLM-based CTI. We introduce HalluVision, a reproducible framework that generates LLM outputs, extracts domain-specific entities using a fine-tuned named entity recognition model, and evaluates factual consistency using multiple similarity metrics. Using 4,940 real-world cybersecurity articles, we measure hallucinations in outputs from two commercial LLMs and one open-weight LLM. Measurements are validated against 100 human-annotated outputs and two established faithfulness detectors. Results show that semantic similarity aligns with human judgments, whereas entity-based metrics exhibit limited agreement. Differences between models are concentrated in question answering, while entity extraction quality limits what entity-level metrics can reveal.

## TAR: Temporal Anchor-Constrained Reasoning for Video Temporal Grounding

DevFeed: [TAR: Temporal Anchor-Constrained Reasoning for Video Temporal Grounding](<https://devfeed.tech/articles/tar-temporal-anchor-constrained-reasoning-for-video-temporal-grounding-139746.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2508.07683>)

Author: Chaohong Guo, Xun Mo, Yongwei Nie, Fei Ma, Xuemiao Xu, Chengjiang Long

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

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Chain-of-thought](<https://devfeed.tech/topics/chain-of-thought.md>), [image recognition](<https://devfeed.tech/topics/image-recognition.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [anchor](<https://devfeed.tech/tags/anchor.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [grounding](<https://devfeed.tech/tags/grounding.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [temporal](<https://devfeed.tech/tags/temporal.md>), [video](<https://devfeed.tech/tags/video.md>)

### AI overview

TAR (Temporal Anchor-Constrained Reasoning) adds an auditable temporal anchor checkpoint to video temporal grounding. It uses the anchor to progressively refine chain-of-thought reasoning against visual evidence and recalibrate temporal predictions, aiming to keep intermediate reasoning aligned with the final segment prediction. The paper says this improves reasoning faithfulness and accuracy. It also describes bootstrapping high-quality reasoning data with a standard 7B model, avoiding reliance on ultra-large models. The abstract reports state-of-the-art performance, but provides no benchmark details or quantitative results.

### Source excerpt

arXiv:2508.07683v3 Announce Type: replace-cross Abstract: Video Temporal Grounding (VTG) aims to localize specific video segments corresponding to natural language queries. While recent Large Vision-Language Models (LVLMs) employ Reinforcement Learning to generate Chains-of-Thought (CoT), they typically rely solely on outcome-based supervision. Consequently, this often leads to hallucinations, where the reasoning process becomes disconnected from the visual content and the final prediction. Existing attempts to mitigate this by relying on external supervision from larger models or separate reward models are computationally expensive and prone to rigid patterns. To address these challenges, we propose TAR (Temporal Anchor-Constrained Reasoning), a framework that introduces the temporal anchor (T-anchor) as a transparent and auditable checkpoint mechanism. T-anchor enforces progressive refinement within the CoT, compelling the model to continuously ground its intermediate thoughts in visual evidence and iteratively calibrate temporal predictions, thereby significantly enhancing the faithfulness and autonomy of the reasoning process and final accuracy. Furthermore, we introduce a bootstrapping paradigm that automatically harvests high-quality CoT data using only a standard 7B model, eliminating the dependency on ultra-large models. Extensive experiments demonstrate that TAR achieves state-of-the-art performance and generates faithful, autonomous, and progressively refined reasoning traces.

## RTLReviewBench: LLM Hardware Reviewers Rarely Miss Bugs but Routinely Reject Correct RTL

DevFeed: [RTLReviewBench: LLM Hardware Reviewers Rarely Miss Bugs but Routinely Reject Correct RTL](<https://devfeed.tech/articles/rtlreviewbench-llm-hardware-reviewers-rarely-miss-bugs-but-routinely-reject-correct-rtl-141661.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21678>)

Author: Yasir Mahmud

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

Content type: article

Language: en

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

Topics: [ai code review](<https://devfeed.tech/topics/ai-code-review.md>), [Verilog](<https://devfeed.tech/topics/verilog.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bugs](<https://devfeed.tech/tags/bugs.md>), [cs-ar](<https://devfeed.tech/tags/cs-ar.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [llm](<https://devfeed.tech/tags/llm.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [release](<https://devfeed.tech/tags/release.md>), [rtl](<https://devfeed.tech/tags/rtl.md>), [simulation](<https://devfeed.tech/tags/simulation.md>)

### AI overview

RTLReviewBench evaluates 12 commercial and open-weight models on 114 matched pairs of buggy and correct Verilog modules, with labels confirmed by simulation. Models rarely accept buggy code but frequently reject correct code, making more than five false rejections per missed bug. On the same designs, they miss fewer synthetic injected bugs than real bugs. For weaker reviewers, asking for reasoning before the verdict sharply reduces false rejections; an ablation finds reasoning placement matters more than the instruction to reason, while sampling and majority voting do not help. A blinded audit finds some correct "buggy" verdicts rely on false technical claims. The authors release the benchmark, a harder split, model responses, and evaluation code.

### Source excerpt

arXiv:2611.21678v1 Announce Type: cross Abstract: Hardware teams increasingly ask large language models (LLMs) to review register-transfer level (RTL) code, yet it is unclear whether a model can tell a correct module from a buggy one without simulating it. We introduce RTLReviewBench, a benchmark of 114 matched pairs of buggy and correct Verilog modules. Each pair shares one specification, most buggy modules are real failures of LLM code generators, and every label is confirmed by simulation. Evaluating 12 commercial and open-weight models, we find a consistent asymmetry: reviewers rarely accept buggy code but often reject correct code, producing more than five false rejections for every missed bug. Synthetic mutants hide part of the problem: on the same designs, models miss far fewer injected bugs than real ones. Asking a model to reason before it commits to a verdict sharply reduces false rejections for weaker reviewers, and a controlled ablation shows that where the reasoning appears in the output matters more than the instruction to reason, whereas sampling and majority voting do not help. A blinded audit further shows that some correct "buggy" verdicts rest on false technical claims. We release the benchmark, a harder split for future models, all model responses, and the evaluation code at https://github.com/yasirmahmud/RTLReviewBench

## PACD: Provenance-Aware Consensus Decoding for Mitigating Cross-Modal Hallucinations in Omnimodal Large Language Models

DevFeed: [PACD: Provenance-Aware Consensus Decoding for Mitigating Cross-Modal Hallucinations in Omnimodal Large Language Models](<https://devfeed.tech/articles/pacd-provenance-aware-consensus-decoding-for-mitigating-cross-modal-hallucinations-in-omnimodal-large-language-models-140031.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22456>)

Author: Chang Sun, Yijun Tian, Zhenwei Tang, Yuchen Deng, Yifan Simon Liu, Chenxu Wang, Wei Wang

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

Content type: article

Language: en

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

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [LLM Techniques](<https://devfeed.tech/topics/llm-techniques.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>)

Tags: [arbitration](<https://devfeed.tech/tags/arbitration.md>), [audio](<https://devfeed.tech/tags/audio.md>), [audio-video](<https://devfeed.tech/tags/audio-video.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [candidate](<https://devfeed.tech/tags/candidate.md>), [consensus](<https://devfeed.tech/tags/consensus.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [errors](<https://devfeed.tech/tags/errors.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [framework](<https://devfeed.tech/tags/framework.md>), [free](<https://devfeed.tech/tags/free.md>), [fusion](<https://devfeed.tech/tags/fusion.md>), [hallucinations](<https://devfeed.tech/tags/hallucinations.md>), [improvement](<https://devfeed.tech/tags/improvement.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [modal](<https://devfeed.tech/tags/modal.md>), [provenance](<https://devfeed.tech/tags/provenance.md>)

### AI overview

PACD is a training-free decoding framework designed to mitigate cross-modal hallucinations in omnimodal large language models. It uses two paths: when modality branches disagree, Relevance-Guided Candidate Arbitration selects task-relevant unimodal evidence and considers alternative candidates; when they agree, Provenance-Based Consensus Recovery evaluates layer-wise evidence consistency and strength to address potentially erroneous consensus, subject to confidence constraints. The abstract reports experiments across four models, five benchmarks and two tasks: up to 27.6% relative improvement over standard decoding on three hallucination benchmarks, and 7.7% average relative improvement over the best baseline on two general audio-visual reasoning benchmarks.

### Source excerpt

arXiv:2611.22456v1 Announce Type: new Abstract: Omnimodal large language models (Omni-LLMs) integrate audio, video, and text for multimodal reasoning, yet remain vulnerable to cross-modal hallucinations, where one modality improperly influences predictions about another. Existing training-free decoding methods mitigate these errors through modality perturbation or relevance-based weighting. However, due to interference from irrelevant modality cues during fusion and difficulty detecting shared errors, these methods may suppress correct unimodal predictions and allow erroneous consensus to persist. To address these limitations, we propose Provenance-Aware Consensus Decoding (PACD), a training-free framework for mitigating cross-modal hallucinations. PACD comprises two complementary paths based on whether the branches reach consensus. For non-consensus cases, Relevance-Guided Candidate Arbitration identifies task-relevant unimodal evidence and selectively arbitrates between predictions using alternative candidates. For consensus cases, Provenance-Based Consensus Recovery leverages the outranking principle to evaluate the consistency and strength of layer-wise modality evidence for alternative answers, selectively resolving erroneous consensus under confidence constraints. Extensive experiments on four Omni-LLMs and five benchmarks across two tasks demonstrate the superiority of PACD. Compared to existing methods, PACD achieves up to 27.6% relative improvement over standard decoding across CMM, AVHBench, and SVHalluc for cross-modal hallucination mitigation, and an average relative improvement of 7.7% over best baseline across OmniVideoBench and FutureOmni for general audio-visual reasoning.

## Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models

DevFeed: [Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models](<https://devfeed.tech/articles/wrong-with-confidence-measuring-hallucination-overconfidence-and-identity-conditioned-behaviour-in-large-language-models-137334.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18728>)

Author: Md. Rejaul Korim Sadi, Toufiqur Rahman Tasin, Golam Mostofa Naeem, Dewan Fahad Chowdhury, Hatim Al Amin Chowdhury

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

Content type: news

Language: en

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

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [identity](<https://devfeed.tech/tags/identity.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [risk](<https://devfeed.tech/tags/risk.md>)

### AI overview

In a paired study of five open-weight language models, removing a question's referent increased errors among attempted answers from 39.1% to 52.0%; models asked for clarification on 33.7% of underspecified prompts. Wrong answers rarely signaled uncertainty: 75.3% had no hedge, and hedging weakly distinguished correct from incorrect responses (AUROC 0.56). Specialist-domain questions failed more often than general ones across all five models. The study also found occupational stereotypes in gendered pronouns, answers about 45% longer when a user identity was stated, and no reduction in safety caveats when users claimed expertise. The authors recommend evaluating clarification separately from answer accuracy and checking label validity as well as rater agreement.

### Source excerpt

arXiv:2611.18728v1 Announce Type: cross Abstract: Large language models answer underspecified questions readily, and accuracy-only evaluation rewards guessing over abstaining. Observational studies link query underspecification to hallucination, but they cannot separate the effect of missing information from item difficulty, and they rarely ask what the model tells the user. We report a pre-registered, paired behavioural study of five open-weight models (Gemma 3 27B, Gemma 4 31B, gpt-oss-120b, Qwen3-Next-80B-A3B and Nemotron 3 Super). Each of 142 factual questions was posed in a fully specified form and in a referentially underspecified form with the same answer key, and three probes measured identity-conditioned behaviour. Two raters scored 5,620 responses under a fixed rubric; effects were estimated with mixed-effects models. Without the referent, the models sought clarification on 33.7% of prompts and answered the rest. Among attempted answers, the error rate rose from 39.1% to 52.0% (mixed-effects odds ratio 4.12, 95% CI 2.76-6.14), and specialist-domain questions failed more often than general ones in every model. Unelicited verbal uncertainty did not track this risk: 75.3% of wrong answers carried no hedge, wrong answers to underspecified prompts were hedged no more often than wrong answers to clear prompts (27.3% vs 22.3%), and hedging separated correct from incorrect answers only weakly (AUROC 0.56). Gendered pronouns followed occupational stereotypes whenever a model committed to one; any stated user identity lengthened answers by about 45% relative to an anonymous user; claimed expertise did not reduce safety caveats. The models either ask or answer as if the question were well posed, and the wording of a guess does not reveal that it is one. Evaluations should score clarification as a distinct outcome and check label validity as well as rater agreement.

## HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection

DevFeed: [HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection](<https://devfeed.tech/articles/halluprobe-exploiting-vlm-hallucinations-for-backdoor-detection-137505.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19868>)

Author: Ji Guo, Xiaolong Qin, Jin Jiang, Yijing Liu, Pei Ke, Aiguo Chen, Wenbo Jiang

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

Content type: article

Language: en

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

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [vlm](<https://devfeed.tech/topics/vlm.md>)

Tags: [backdoor](<https://devfeed.tech/tags/backdoor.md>), [computer-science](<https://devfeed.tech/tags/computer-science.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [data](<https://devfeed.tech/tags/data.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [detection](<https://devfeed.tech/tags/detection.md>), [exploit](<https://devfeed.tech/tags/exploit.md>), [hallucinations](<https://devfeed.tech/tags/hallucinations.md>), [language](<https://devfeed.tech/tags/language.md>), [random](<https://devfeed.tech/tags/random.md>), [results](<https://devfeed.tech/tags/results.md>), [security](<https://devfeed.tech/tags/security.md>), [target](<https://devfeed.tech/tags/target.md>), [token](<https://devfeed.tech/tags/token.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [training](<https://devfeed.tech/tags/training.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>), [vision-language-models-vlms](<https://devfeed.tech/tags/vision-language-models-vlms.md>), [vlm](<https://devfeed.tech/tags/vlm.md>), [vlms](<https://devfeed.tech/tags/vlms.md>)

### AI overview

HalluProbe detects possible backdoors in vision-language models by amplifying hallucinations during decoding and checking whether hallucinated tokens repeatedly favor a target token. The method is based on the reported pattern that poisoned models show stable target-token bias across varied inputs, while clean models' hallucinated tokens are largely random. In tests spanning six representative backdoor attacks and three models, the authors report accuracy above 90% and better performance than existing detection methods.

### Source excerpt

arXiv:2611.19868v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) are vulnerable to backdoor attacks introduced through data poisoning, making effective backdoor detection a critical security challenge. In this paper, we uncover a novel connection between VLM hallucinations and backdoor behaviors and exploit this connection for backdoor detection. Our key insight is that poisoned training data induces strong language priors toward the backdoor target token, while such priors are also a major source of hallucinations. This suggests that enhancing hallucinations can expose hidden backdoor-related priors. Specifically, backdoored VLMs tend to hallucinate backdoor target tokens more frequently across diverse inputs, while the hallucinated tokens of clean VLMs are largely random and exhibit no stable target-specific bias. Based on this observation, we propose HalluProbe, which enhances hallucinations during output decoding and measures the distributional consistency of hallucinated tokens to identify backdoored VLMs. We evaluate HalluProbe against six representative VLM backdoor attacks across three models. Experimental results demonstrate that HalluProbe can effectively detect backdoored VLMs with an accuracy exceeding 90%, outperforming existing backdoor detection methods.

## Evaluating and Improving Natural Language Autoencoders

DevFeed: [Evaluating and Improving Natural Language Autoencoders](<https://devfeed.tech/articles/evaluating-and-improving-natural-language-autoencoders-138880.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19507>)

Author: Marco Bazzani, Dan Mossing

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

Content type: article

Language: en

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

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [activation](<https://devfeed.tech/tags/activation.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [autoencoders](<https://devfeed.tech/tags/autoencoders.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [hint](<https://devfeed.tech/tags/hint.md>), [improvements](<https://devfeed.tech/tags/improvements.md>), [knowledge](<https://devfeed.tech/tags/knowledge.md>), [layers](<https://devfeed.tech/tags/layers.md>), [model](<https://devfeed.tech/tags/model.md>), [natural-language](<https://devfeed.tech/tags/natural-language.md>), [performance](<https://devfeed.tech/tags/performance.md>), [training](<https://devfeed.tech/tags/training.md>), [transcripts](<https://devfeed.tech/tags/transcripts.md>)

### AI overview

Natural language autoencoders (NLAs) generate explanations of LLM activations, and the paper evaluates them on auditing tasks: surfacing latent knowledge, detecting when a model was misled by a hint, and catching silent fabrications. The authors extend NLAs to use activations from multiple layers and positions, handle long transcripts more effectively, and reduce false claims. These changes improve performance on the auditing evaluations and auxiliary evaluations adapted from prior work. Across training, all evaluations improve early, but extended training can reduce performance on some auditing tasks. This overview reflects the supplied abstract.

### Source excerpt

arXiv:2611.19507v1 Announce Type: cross Abstract: Natural language autoencoders (NLAs) are a method for generating natural language explanations of LLM activations. We develop evaluations based on auditing tasks that test the ability of NLAs to surface latent knowledge, identify whether a model was misled by a hint, and catch a silent fabrication. We also develop extensions to NLAs that let them use activation information from multiple layers and positions, let us apply them more effectively to long transcripts, and reduce the number of false claims they make. These extensions result in improvements on both our auditing evaluations and auxiliary evaluations adapted from prior work. We also use the evaluations to study how different auditing capabilities change over the course of NLA training, and find that although every evaluation improves early in training, extended training can hurt performance on some of our auditing evaluations.

## Local LLMs Handle Textbook RAG but Fail on New Research Papers

DevFeed: [Local LLMs Handle Textbook RAG but Fail on New Research Papers](<https://devfeed.tech/articles/local-llms-handle-textbook-rag-but-fail-on-new-research-papers-135976.md>)

Original publisher: [Read original article](<https://towardsdatascience.com/local-llms-handle-textbook-rag-but-fail-on-new-research-papers/>)

Author: Waleed Esmail

Published: 2026-11-16T12:00:00Z

Content type: article

Language: en

Sources: [Towards Data Science](<https://devfeed.tech/sources/towards-data-science.md>)

Topics: [Local AI](<https://devfeed.tech/topics/local-ai.md>), [generalization in machine learning](<https://devfeed.tech/topics/generalization-in-machine-learning.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [gpt-oss](<https://devfeed.tech/topics/gpt-oss.md>)

Tags: [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [llms](<https://devfeed.tech/tags/llms.md>), [local](<https://devfeed.tech/tags/local.md>), [local-llm-inference](<https://devfeed.tech/tags/local-llm-inference.md>), [local-llms](<https://devfeed.tech/tags/local-llms.md>), [papers](<https://devfeed.tech/tags/papers.md>), [rag](<https://devfeed.tech/tags/rag.md>), [rag-retrieval-augmented-generation](<https://devfeed.tech/tags/rag-retrieval-augmented-generation.md>)

### AI overview

In this evaluation, five setups scored 12-20 of 20 on textbook questions but only 0-9 of 15 on a paper published that year, despite identical retrieval and prompts. The author concludes that small local models handle familiar material better than new research, where they must rely on supplied passages. Grading answers alongside retrieved passages showed many errors came from missing evidence: figure descriptions raised figure-question results from 1/10 to 9/10, chapter summary trees raised overview results from 1/8 to 8/8, and neighboring chunks improved complete retrieval from 16/20 to 19/20. Quantization did not explain the gap. Qwen3.8-27B was close to Opus in the reported comparison at lower cost, but took about three minutes per answer. The author recommends writing gold questions first, separating familiar from new material, inspecting passages when grading, improving retrieval before changing models, and routing figures, web reading, and derivations to stronger models. Results are limited by a 35-question evaluation and grader drift.

### Source excerpt

Five setups, one blind grading pack: textbook scores ran 12 to 20 of 20, but a paper published this year dropped them to 0 to 9 of 15. The post Local LLMs Handle Textbook RAG but Fail on New Research Papers appeared first on Towards Data Science.

## BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems

DevFeed: [BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems](<https://devfeed.tech/articles/bbowp-bench-evaluating-llms-on-black-box-optimization-word-problems-135707.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2608.02612>)

Author: Yutaro Yamada, Kei Hiroshima, Nozomu Yoshinari, Kento Uchida, Shinichi Shirakawa

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [Operations research and optimization](<https://devfeed.tech/topics/operations-research-and-optimization.md>), [Combinatorial optimization](<https://devfeed.tech/topics/combinatorial-optimization.md>), [Computer science](<https://devfeed.tech/topics/computer-science.md>), [Algorithms](<https://devfeed.tech/topics/algorithms.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [algorithm](<https://devfeed.tech/tags/algorithm.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [budget](<https://devfeed.tech/tags/budget.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ne](<https://devfeed.tech/tags/cs-ne.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [design](<https://devfeed.tech/tags/design.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [framework](<https://devfeed.tech/tags/framework.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [llms](<https://devfeed.tech/tags/llms.md>), [natural-language](<https://devfeed.tech/tags/natural-language.md>), [observable](<https://devfeed.tech/tags/observable.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [search](<https://devfeed.tech/tags/search.md>), [spaces](<https://devfeed.tech/tags/spaces.md>)

### AI overview

BBOWP-Bench evaluates whether LLMs can turn natural-language black-box optimization tasks into both a search space and an optimization algorithm, where only objective values are observable and the functional form is unavailable. Each benchmark instance pairs a problem description with an executable evaluation environment and a human-designed reference formulation, enabling assessment of both choices across multiple BBO domains. The abstract reports that models vary their optimizer choice with the evaluation budget, but can struggle to identify important variables and set balanced ranges, especially when descriptions provide little information or the search space is highly problem-specific.

### Source excerpt

arXiv:2608.02612v2 Announce Type: replace-cross Abstract: Formulating an optimization problem strongly affects final solution quality, yet good formulations usually require substantial expertise. Recent studies have explored automatically deriving optimization problems from natural-language descriptions, but existing benchmarks focus on settings where objectives and constraints can be written explicitly as mathematical expressions. Many practically important problems are naturally treated as black-box optimization (BBO) problems, in which only objective values are observable, and the functional form is unavailable. In BBO, search space design (as part of problem formulation) and optimization algorithm selection are crucial for problem-solving. Automating these processes with large language models (LLMs) remains challenging. We introduce Black-Box Optimization Word Problems (BBOWP), a novel problem setting in which a system must infer both a search space and an optimization algorithm from a natural-language description of a black-box optimization task. To support research on this setting, we establish the BBOWP Benchmark Suite (BBOWP-Bench), a dataset and evaluation framework for BBOWP. Each instance combines a natural-language problem description, an executable evaluation environment, and a human-designed reference formulation, allowing evaluation of both search-space design and algorithm selection. Using this benchmark, we evaluate LLMs on the joint formulation of search spaces and optimizers from natural-language descriptions across multiple BBO domains. We show that current LLMs vary their optimizer choice with the evaluation budget, though they sometimes struggle with search space design, particularly in identifying important variables and balancing their ranges when the problem description is less informative or the search space is highly problem-specific. Our code and dataset are available at https://github.com/shiralab/bbowp-bench.

## When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models

DevFeed: [When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models](<https://devfeed.tech/articles/when-to-rethink-learning-multi-perspective-self-verification-for-vision-language-models-135014.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.07018>)

Author: Ziquan Zhu, Hanruo Zhu, Si-Yuan Lu, Morris Yu-Chao Huang, Yicheng Lin, Wei Han, Tianlong Chen, Mingyuan Wu, Hanchao Yu, Gaojie Jin, Lu Liu, Bo Sun, Tianjin Huang

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Model Development](<https://devfeed.tech/topics/model-development.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [learning](<https://devfeed.tech/tags/learning.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [verification](<https://devfeed.tech/tags/verification.md>), [vision-language-models-vlms](<https://devfeed.tech/tags/vision-language-models-vlms.md>)

### AI overview

MOTIVE improves vision-language model self-verification by checking each candidate answer from complementary perspectives and learning a correctness-aligned reliability score. At inference, the score determines whether to accept an answer or trigger history-guided rethinking. The authors report that stronger verifiers provide more reliable judgments, while prompt performance varies across tasks and no single prompt consistently leads. Experiments across multimodal benchmarks and VLM backbones reportedly outperform strong self-verification and self-correction baselines; reliable verification also reduces unnecessary reasoning turns. The abstract describes these findings but provides no benchmark-level results or implementation details.

### Source excerpt

arXiv:2610.07018v2 Announce Type: replace Abstract: Vision-language models (VLMs) have achieved strong performance in multimodal reasoning, yet they remain prone to generating plausible but incorrect answers. Self-verification offers a practical way to improve answer reliability without relying on external judges, but existing methods typically depend on a single verification criterion or fixed prompt, resulting in incomplete and unstable reliability estimates. We first systematically analyze how verifier capability and prompt design affect verification performance. Our findings show that stronger verifiers provide more reliable judgments, while verification performance is highly sensitive to prompt choice, with no single prompt consistently dominating across tasks. Guided by these findings, we propose \texttt{MOTIVE}, a \textbf{M}ulti-View Self-Verificati\textbf{O}n wi\textbf{T}h Rel\textbf{I}ability-Guided Selecti\textbf{VE} Rethinking framework for reliable multimodal reasoning. \texttt{MOTIVE} evaluates each candidate answer from complementary verification perspectives and learns a correctness-aligned reliability score through correctness-grounded multi-view verification learning. During inference, this score governs an accept-or-rethink decision, allowing reliable answers to be returned directly while uncertain ones trigger history-guided rethinking. Extensive experiments across diverse multimodal benchmarks and VLM backbones demonstrate that \texttt{MOTIVE} consistently outperforms strong self-verification and self-correction baselines. Further results show that reliable verification improves accept-or-rethink decisions and reduces unnecessary reasoning turns, enabling more reliable and efficient self-verification without an external judge.

## Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume

DevFeed: [Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume](<https://devfeed.tech/articles/uncertainty-quantification-for-multimodal-large-language-models-with-incoherence-adjusted-semantic-volume-134963.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2602.24195>)

Author: Gregory Kang Ruey Lau, Hieu Dao, Nicole Kan Hui Lin, Bryan Kian Hsiang Low

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [audio](<https://devfeed.tech/tags/audio.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [detection](<https://devfeed.tech/tags/detection.md>), [dispersion](<https://devfeed.tech/tags/dispersion.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [entropy](<https://devfeed.tech/tags/entropy.md>), [generation](<https://devfeed.tech/tags/generation.md>), [image](<https://devfeed.tech/tags/image.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [theoretical](<https://devfeed.tech/tags/theoretical.md>), [training](<https://devfeed.tech/tags/training.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [uncertainty-quantification](<https://devfeed.tech/tags/uncertainty-quantification.md>), [video](<https://devfeed.tech/tags/video.md>), [volume](<https://devfeed.tech/tags/volume.md>)

### AI overview

UMPIRE is a training-free uncertainty framework for multimodal large language models that estimates reliability from sampled responses without external tools. Inspired by Determinantal Point Processes, it combines semantic embedding dispersion with response quadratic entropy to capture both global diversity and local incoherence, using internal modality features. The abstract reports improved error detection and uncertainty calibration over baseline metrics on image, audio, and video-text benchmarks, including with small sample budgets, and reports extension to image and audio generation. The supplied excerpt does not specify the baselines, evaluation details, or limitations.

### Source excerpt

arXiv:2602.24195v2 Announce Type: replace Abstract: Despite their capabilities, Multimodal Large Language Models (MLLMs) may produce plausible but erroneous outputs, hindering reliable deployment. Accurate uncertainty metrics could enable escalation of unreliable queries to human experts or larger models for improved performance. However, existing uncertainty metrics have practical constraints, such as being designed only for specific modalities, reliant on external tools, or computationally expensive. We introduce UMPIRE, a training-free uncertainty quantification framework for MLLMs that works efficiently across various input and output modalities without external tools, relying only on the models' own internal modality features. Inspired by Determinantal Point Processes (DPPs), UMPIRE computes a novel incoherence-adjusted semantic volume of sampled MLLM responses for a given task instance, which couples semantic embedding dispersion with the quadratic entropy of the responses. This approach effectively captures both the global semantic diversity of samples and the local incoherence of responses based on internal model confidence. We provide theoretical analysis motivating UMPIRE's design, and extensive experiments show that UMPIRE consistently outperforms baseline metrics in error detection and uncertainty calibration across image, audio, and video-text benchmarks, even with small sample budgets. We also demonstrate UMPIRE's generalization to non-text output tasks, including image and audio generation. Our code is available at https://github.com/daohieu17ctt/UMPIRE.

## Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination

DevFeed: [Can Mechanistic Interpretability Help Explain LLM Agent Failures? A Case Study on Web Element Hallucination](<https://devfeed.tech/articles/can-mechanistic-interpretability-help-explain-llm-agent-failures-a-case-study-on-web-element-hallucination-135210.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18514>)

Author: Mohamed Aghzal, Mor Geva, Gregory J. Stein, Ziyu Yao

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [AI agent hallucinations](<https://devfeed.tech/topics/ai-agent-hallucinations.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [computer-use](<https://devfeed.tech/topics/computer-use.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [actions](<https://devfeed.tech/tags/actions.md>), [after](<https://devfeed.tech/tags/after.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [apply](<https://devfeed.tech/tags/apply.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [can](<https://devfeed.tech/tags/can.md>), [case](<https://devfeed.tech/tags/case.md>), [case-study](<https://devfeed.tech/tags/case-study.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [element](<https://devfeed.tech/tags/element.md>), [explain](<https://devfeed.tech/tags/explain.md>), [failures](<https://devfeed.tech/tags/failures.md>), [help](<https://devfeed.tech/tags/help.md>), [interpretability](<https://devfeed.tech/tags/interpretability.md>), [llm](<https://devfeed.tech/tags/llm.md>), [mechanistic-interpretability](<https://devfeed.tech/tags/mechanistic-interpretability.md>), [study](<https://devfeed.tech/tags/study.md>), [web](<https://devfeed.tech/tags/web.md>)

### AI overview

In experiments with Qwen2.5-7B-Instruct on live websites, a linear probe on the last prompt token predicted unavailable-action hallucinations before generation (0.80 AUC). Steering along the probe direction corrected 85% of hallucinations but disrupted 7% of valid actions. The authors describe this as only a partial explanation: steering was sensitive to a minor change in action formatting, and circuit identification was difficult because computations spanned multiple components and positions, while relied-on information varied across examples. They argue these findings challenge assumptions of standard mechanistic interpretability methods for explaining agent failures.

### Source excerpt

arXiv:2611.18514v1 Announce Type: new Abstract: Mechanistic Interpretability (MI) has made significant progress in explaining the internal computations of Large Language Models (LLMs); however, these advances have yet to demonstrate comparable success in explaining the decision-making processes of LLM-based agents. This paper examines the challenges MI tools face when applied to realistic agentic failures. We focus on one such failure: web agents hallucinating actions that are not available on a website. We apply a series of mechanistic analyzes to Qwen2.5-7B-Instruct as it interacts with live websites. We find that a linear probe trained on the last prompt token can predict hallucinations before generation begins (0.80 AUC) and steering along a probe direction corrects 85\% of hallucinations, but disrupts 7\% of valid actions. However, these results only provide a partial explanation of why the agent hallucinates. The probe direction is sensitive to prompt formatting, losing much of its steering effect after a minor change to how actions are represented. Additionally, circuit identification can be difficult as the computation is distributed across multiple components and positions, and the information the agent relies on varies across examples. We discuss how these observations challenge the assumptions underlying standard MI methods for circuit discovery and feature identification, and outline future directions for making these methods better suited to explain failures in LLM-based agents.

## From Regulation to Implementation: A Critical Evaluation of LLM-Assisted Regulatory Compliance in Industry

DevFeed: [From Regulation to Implementation: A Critical Evaluation of LLM-Assisted Regulatory Compliance in Industry](<https://devfeed.tech/articles/from-regulation-to-implementation-a-critical-evaluation-of-llm-assisted-regulatory-compliance-in-industry-135002.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2608.21317>)

Author: Marco B\\"ucheler, Adriana Watson, Grant Richards

Published: 2026-11-16T05:00:00Z

Content type: news

Language: en

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

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Reliable AI outputs](<https://devfeed.tech/topics/reliable-ai-outputs.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [artifacts](<https://devfeed.tech/tags/artifacts.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [critical](<https://devfeed.tech/tags/critical.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [data-extraction](<https://devfeed.tech/tags/data-extraction.md>), [data-protection](<https://devfeed.tech/tags/data-protection.md>), [eu](<https://devfeed.tech/tags/eu.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [gdpr](<https://devfeed.tech/tags/gdpr.md>), [generation](<https://devfeed.tech/tags/generation.md>), [hallucinations](<https://devfeed.tech/tags/hallucinations.md>), [impact](<https://devfeed.tech/tags/impact.md>), [industrial](<https://devfeed.tech/tags/industrial.md>), [industry](<https://devfeed.tech/tags/industry.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-assisted](<https://devfeed.tech/tags/llm-assisted.md>), [regulation](<https://devfeed.tech/tags/regulation.md>), [regulatory](<https://devfeed.tech/tags/regulatory.md>), [regulatory-compliance](<https://devfeed.tech/tags/regulatory-compliance.md>)

### AI overview

The study benchmarks LLM-generated compliance artifacts against manually created schemas to examine how extraction instructions and regulatory vagueness affect quality and consistency. It focuses on Digital Product Passports and Data Protection Impact Assessments, which pose different challenges: industrial data for passports is scattered across varied systems, while DPIAs require interdisciplinary expertise and lack a standard format. The reported results indicate that loosely specified DPIA formatting needs more context in prompts to support consistent, complete outputs. Stricter Digital Battery Passport formatting produces more consistent results across prompt contexts, but may also lead to more hallucinations.

### Source excerpt

arXiv:2608.21317v2 Announce Type: replace Abstract: The European Union (EU) has emerged as a leading regulatory body in the development of sustainability and privacy regulations. While new regulation requirements vary, many include a documentation artifact to ensure compliance. Notably, the Ecodesign for Sustainable Products Regulation (ESPR) introduces Digital Product Passports (DPPs) for life cycle transparency, while the General Data Protection Regulation (GDPR) mandates Data Protection Impact Assessments (DPIAs) to mitigate privacy risks. Creating these compliance artifacts, however, is challenging. Industrial data, which often exists in heterogeneous formats and is scattered across company and supplier systems, is required for DPPs and can be difficult to extract into compliant DPP formatting. Furthermore, DPIA documents require interdisciplinary expertise and follow no standardized format, making development difficult for novel systems. To address the particular complexity of compliance artifact creation for both regulations, researchers have proposed the use of LLMs in the generation process; however, the impact of the aforementioned problems on the output of these systems is largely unaddressed. This work investigates the existing research gap by exploring how data extraction instructions and regulatory vagueness impact the quality and consistency of LLM-produced compliance artifacts. The resulting artifacts are evaluated by benchmarking different models against manually created ground-truth schemas. The results reveal that less strict guidelines, such as DPIA formatting, require higher context prompts to maintain consistency and completeness. Stricter guidelines, such as formatting for Digital Battery Passports (DBP), result in consistent results regardless of prompt context, but may lead to more hallucinations in the output

## Identifying and Mitigating API Misuse in Large Language Models

DevFeed: [Identifying and Mitigating API Misuse in Large Language Models](<https://devfeed.tech/articles/identifying-and-mitigating-api-misuse-in-large-language-models-136474.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2503.22821>)

Author: Terry Yue Zhuo, Junda He, Jiamou Sun, Zhenchang Xing, David Lo, John Grundy, Xiaoning Du

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

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

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [development-process](<https://devfeed.tech/topics/development-process.md>), [Python](<https://devfeed.tech/topics/python.md>)

Tags: [api](<https://devfeed.tech/tags/api.md>), [apis](<https://devfeed.tech/tags/apis.md>), [automated](<https://devfeed.tech/tags/automated.md>), [code-generation](<https://devfeed.tech/tags/code-generation.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [java](<https://devfeed.tech/tags/java.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [library](<https://devfeed.tech/tags/library.md>), [llms](<https://devfeed.tech/tags/llms.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [patterns](<https://devfeed.tech/tags/patterns.md>), [prompting](<https://devfeed.tech/tags/prompting.md>), [python](<https://devfeed.tech/tags/python.md>), [repair](<https://devfeed.tech/tags/repair.md>), [selection](<https://devfeed.tech/tags/selection.md>), [study](<https://devfeed.tech/tags/study.md>), [taxonomies](<https://devfeed.tech/tags/taxonomies.md>), [taxonomy](<https://devfeed.tech/tags/taxonomy.md>), [types](<https://devfeed.tech/tags/types.md>)

### AI overview

The study finds recurring API misuse in LLM-generated Python and Java code, including hallucination and intent misalignment. It manually annotated 3,209 method-level and 3,492 parameter-level misuses across StarCoder-7B, Qwen2.5-Coder-7B, and GitHub Copilot, then categorized four misuse types. The authors propose Dr.Fix, an LLM-based repair approach guided by that taxonomy. On benchmark datasets, it improved on baseline prompting and existing repair methods by up to 38.4 BLEU and 72 exact matches.

### Source excerpt

arXiv:2503.22821v3 Announce Type: replace Abstract: API misuse in code generated by large language models (LLMs) presents a serious and growing challenge in software development, as although LLMs demonstrate impressive code generation capabilities, their interactions with complex library APIs are often error-prone and can lead to software failures and vulnerabilities. In this paper, we conduct a large-scale study of API misuse patterns in LLM-generated code by analyzing both method selection and parameter usage across Python and Java, using three representative LLMs: StarCoder-7B, Qwen2.5-Coder-7B, and GitHub Copilot. Based on extensive manual annotation of 3,209 method-level and 3,492 parameter-level misuses, we identify and categorize four recurring misuse types by building on and refining prior API misuse taxonomies. Our evaluation of the three LLMs reveals persistent challenges in API usage, particularly hallucination and intent misalignment. To address these issues, we propose Dr.Fix, an LLM-based automatic repair approach guided by our taxonomy, which improves repair accuracy compared to baseline prompting and existing repair methods, achieving gains of up to 38.4 BLEU and up to 72 more exact matches on benchmark datasets. This work offers important insights into the current limitations of LLMs in API usage and points to directions for improving automated misuse repair in code generation systems.

## Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation

DevFeed: [Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation](<https://devfeed.tech/articles/cited-but-unused-parametric-provenance-in-retrieval-augmented-generation-132317.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17161>)

Author: Tanzeel Abbas, Tehseen Zia, David Windridge

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

Content type: article

Language: en

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

Topics: [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Factuality scoring in AI](<https://devfeed.tech/topics/factuality-scoring-in-ai.md>)

Tags: [attribution](<https://devfeed.tech/tags/attribution.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [memory](<https://devfeed.tech/tags/memory.md>), [model](<https://devfeed.tech/tags/model.md>), [provenance](<https://devfeed.tech/tags/provenance.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [retrieval-augmented-generation](<https://devfeed.tech/tags/retrieval-augmented-generation.md>), [signal](<https://devfeed.tech/tags/signal.md>), [size](<https://devfeed.tech/tags/size.md>)

### AI overview

The study finds that some cited claims in retrieval-augmented generation come from a model's parametric memory rather than retrieved passages: 1.2%-15.7% across the tested models, with higher rates as model size increased within one family. Replacing retrieved facts with false variants exposed this provenance mismatch. Natural language inference checks performed at chance, even when limited to the cited passage, because models answering from memory could select citations consistent with their answers. Of ten attribution methods compared, the change in the model's output distribution with versus without context best distinguished provenance, but using that signal did not make detection accurate enough to act on. The study concludes that telling what a model knows from what it actually used remains unresolved.

### Source excerpt

arXiv:2611.17161v1 Announce Type: new Abstract: Retrieval-augmented generation systems attach citations to attribute claims to their sources. These citations frequently misstate provenance, and standard verification cannot detect this. We measure provenance by counterfactual value substitution: a fact in the retrieved passages is replaced with a false variant and the model re-queried. We find that across five sizes of one model family and a second comparator family, 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context. At the most conservative substitution distance these figures are 4.6% and 8.5% on the two models we tested at it, with the rate rising with model size within a family. We further find that natural language inference verification, the standard approach, gives results indistinguishable from chance across models, even when restricted to the cited passage. We determine that the cause is structural: a model answering from memory selects a citation that agrees with its output. Comparing ten attribution methods on identical claims, the contrast between the model's distribution with and without context best separates provenance. Acting on that signal does not improve provenance honesty at these base rates. The resulting detector is hence still not accurate enough to act on: it measures what the model knew rather than what it used, and separating those remains the open problem.

## The Oracle and The Prism: A Decoupled and Efficient Framework for Generative Recommendation Explanation

DevFeed: [The Oracle and The Prism: A Decoupled and Efficient Framework for Generative Recommendation Explanation](<https://devfeed.tech/articles/the-oracle-and-the-prism-a-decoupled-and-efficient-framework-for-generative-recommendation-explanation-131860.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2511.16543>)

Author: Jiaheng Zhang, Daqiang Zhang

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

Content type: article

Language: en

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

Topics: [personalization](<https://devfeed.tech/topics/personalization.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [also](<https://devfeed.tech/tags/also.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [base](<https://devfeed.tech/tags/base.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [component](<https://devfeed.tech/tags/component.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-ir](<https://devfeed.tech/tags/cs-ir.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [decoupling](<https://devfeed.tech/tags/decoupling.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [effective](<https://devfeed.tech/tags/effective.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [ensures](<https://devfeed.tech/tags/ensures.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generative](<https://devfeed.tech/tags/generative.md>), [oracle](<https://devfeed.tech/tags/oracle.md>), [recommendation](<https://devfeed.tech/tags/recommendation.md>)

### AI overview

Prism separates recommendation ranking from explanation generation, allowing each stage to be optimized for its own objective. It uses FLAN-T5-XXL as a teacher to produce explanatory knowledge, then fine-tunes a 140M-parameter BART-Base student to create personalized explanations. On benchmark datasets, the authors report that distillation filters noise: in human evaluations, Prism outperformed its 11B-parameter teacher on faithfulness and personalization and could correct hallucinations in teacher outputs. The authors also report 24x faster operation and 10x lower memory use. These are reported experimental findings; the supplied excerpt does not provide dataset details or evaluation methodology.

### Source excerpt

arXiv:2511.16543v3 Announce Type: replace-cross Abstract: The integration of Large Language Models (LLMs) into explainable recommendation systems often leads to a performance-efficiency trade-off in end-to-end architectures, where joint optimization of ranking and explanation can result in suboptimal compromises. To resolve this, we propose Prism, a novel decoupled framework that rigorously separates the recommendation process into a dedicated ranking stage and an explanation generation stage. This decomposition ensures that each component is optimized for its specific objective, eliminating inherent conflicts in coupled models. Inspired by knowledge distillation, Prism leverages a powerful, instruction-following teacher LLM (FLAN-T5-XXL) as an Oracle to produce high-fidelity explanatory knowledge. A compact, fine-tuned student model (BART-Base), the Prism, then specializes in synthesizing this knowledge into personalized explanations. Our extensive experiments on benchmark datasets reveal a key finding: the distillation process not only transfers knowledge but also acts as a noise filter. Our 140M-parameter Prism model significantly outperforms its 11B-parameter teacher in human evaluations of faithfulness and personalization, demonstrating an emergent ability to correct hallucinations present in the teacher's outputs. While achieving a 24x speedup and a 10x reduction in memory consumption, our analysis validates that decoupling, coupled with targeted distillation, provides an efficient and effective pathway to high-quality, and perhaps more importantly, trustworthy explainable recommendation.

## Towards Robust Vision-Language Reasoning: A Multi-Stage Optimization Pipeline with Masked Diffusion Model

DevFeed: [Towards Robust Vision-Language Reasoning: A Multi-Stage Optimization Pipeline with Masked Diffusion Model](<https://devfeed.tech/articles/towards-robust-vision-language-reasoning-a-multi-stage-optimization-pipeline-with-masked-diffusion-model-132113.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.15238>)

Author: Shunming Liu, Pei Huang, Yajing Sun, Huaxiao Liu

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

Content type: article

Language: en

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

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [autonomous vehicles](<https://devfeed.tech/topics/autonomous-vehicles.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [architecture](<https://devfeed.tech/tags/architecture.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [autonomous-driving](<https://devfeed.tech/tags/autonomous-driving.md>), [autoregressive](<https://devfeed.tech/tags/autoregressive.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [denoising](<https://devfeed.tech/tags/denoising.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [diversity](<https://devfeed.tech/tags/diversity.md>), [language](<https://devfeed.tech/tags/language.md>), [model](<https://devfeed.tech/tags/model.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [robust](<https://devfeed.tech/tags/robust.md>), [stage](<https://devfeed.tech/tags/stage.md>), [vision](<https://devfeed.tech/tags/vision.md>)

### AI overview

The paper proposes MD-Drive, a discrete diffusion vision-language model for autonomous driving, and reports a 29.7% performance improvement over its unoptimized version. Its pipeline adapts standard VLM architectures without auxiliary action experts or specialized mixture-of-experts Transformers: supervised fine-tuning reshapes planning-input representations to suppress hallucinations; reinforcement learning uses metric-guided diversity filtering to improve starting points and policy exploration; and Dynamic Autonomous Driving Decoding adjusts decoding modes to scene complexity. On NAVSIM closed loop and the nuScenes and DriveCoT open loop datasets, MD-Drive reportedly outperforms same-architecture autoregressive VLM baselines and offers a more favorable performance-compute balance than systems with different architectures. The supplied text does not provide detailed metrics or experimental conditions.

### Source excerpt

arXiv:2611.15238v1 Announce Type: new Abstract: Discrete diffusion generation is emerging as a competitive alternative to autoregressive generation by denoising the entire token sequence through iterative refinement. Its global planning and progressive correction capabilities are particularly well suited to autonomous driving. However, transferring this generation paradigm from general-purpose settings to driving introduces substantial domain gaps in both training and inference. To unlock its potential, we propose a multi-stage optimization pipeline for standard vision-language model architectures, without auxiliary action experts or specialized mixture-of-experts Transformer, spanning post-training and inference. During supervised fine-tuning, we optimize the representation of planning inputs to suppress hallucinations in task responses. During reinforcement learning, we introduce metric-guided diversity filtering to identify stronger learning starting points and improve policy exploration. At inference time, we propose Dynamic Autonomous Driving Decoding (DAD), which adaptively adjusts decoding modes according to scene complexity. Based on this pipeline, we develop Masked Diffusion Drive (MD-Drive), improving its performance by 29.7\% over the unoptimized model. Experiments on the NAVSIM closed loop benchmark and the nuScenes and DriveCoT open loop datasets show that MD-Drive outperforms autoregressive VLM baselines with the same architecture, while achieving a more favorable balance between performance and computational cost than systems with different architectures through discrete diffusion training and inference.

## Characterizing Local Breadth in Language Generation

DevFeed: [Characterizing Local Breadth in Language Generation](<https://devfeed.tech/articles/characterizing-local-breadth-in-language-generation-132928.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.14887>)

Author: Jon Kleinberg, Fan Wei

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

Content type: article

Language: en

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

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [after](<https://devfeed.tech/tags/after.md>), [all](<https://devfeed.tech/tags/all.md>), [alpha](<https://devfeed.tech/tags/alpha.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [avoid](<https://devfeed.tech/tags/avoid.md>), [based](<https://devfeed.tech/tags/based.md>), [best](<https://devfeed.tech/tags/best.md>), [can](<https://devfeed.tech/tags/can.md>), [characterization](<https://devfeed.tech/tags/characterization.md>), [collection](<https://devfeed.tech/tags/collection.md>), [complete](<https://devfeed.tech/tags/complete.md>), [condition](<https://devfeed.tech/tags/condition.md>), [cs-ds](<https://devfeed.tech/tags/cs-ds.md>), [cs-fl](<https://devfeed.tech/tags/cs-fl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [d](<https://devfeed.tech/tags/d.md>), [generation](<https://devfeed.tech/tags/generation.md>), [language](<https://devfeed.tech/tags/language.md>), [local](<https://devfeed.tech/tags/local.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>)

### AI overview

The paper studies generators that learn an unknown target from positive examples arriving in arbitrary, possibly adversarial order, measuring breadth by lower Banach density, which rules out arbitrarily large sparse gaps. In one dimension, it characterizes each target's optimal density and gives one eventually valid generator attaining the optima across a countable family; finite Cantor-Bendixson point rank ensures density 1/2. In dimensions two and higher, no single generator need be best for every target. The authors characterize jointly achievable density guarantees and show that any product region with coordinate bounds up to 1/2 can arise. With a near-logarithmic nondegeneracy condition, one generator guarantees at least 1/2 − ε for finite-point-rank targets, while the exact 1/2 guarantee can fail even for a rank-zero singleton family.

### Source excerpt

arXiv:2611.14887v1 Announce Type: cross Abstract: Large language models sharpen the validity--breadth trade-off: a generator should avoid hallucinations without collapsing onto a narrow set of valid outputs. We study generation from positive examples of an unknown target in the Kleinberg--Mullainathan model, with arbitrary, possibly adversarial order and no distributional or computational assumptions. Breadth is measured by lower Banach density, which rules out arbitrarily large sparse holes, unlike prefix-based lower asymptotic density. 1. Exact optima in one dimension: For each target $K$ in a countable family, we characterize its optimal lower Banach density $b(K)$ via equivalent combinatorial and topological formulations. One eventually valid generator attains all $b(K)$ simultaneously. Every $\alpha\in[0,1/2]$ occurs as $b(K)$ in some countable family; finite Cantor--Bendixson point rank guarantees $b(K)=1/2$, even without a uniform rank bound. 2. Unavoidable trade-offs in higher dimensions: For $d\ge2$, there need not be a single best generator: optimizing density for one target may worsen the guarantee for another. An explicit family quantifies this intrinsically multi-objective trade-off. 3. Complete characterization of density profiles: We characterize the full region of simultaneously achievable guarantees, with equivalent combinatorial and topological formulations, after excluding geometrically degenerate rectangles. Every prescribed product $\prod_i[0,\alpha_i]$, $\alpha_i\in[0,1/2]$ can occur as an exact achievable region for some language collection. 4. Near-half guarantees and the sharp endpoint: Under a near-logarithmic nondegeneracy condition, one generator achieves filtered density at least $1/2-\varepsilon$ at all finite-point-rank targets, for every $\varepsilon>0$, improving the previous polylogarithmic threshold. The exact $1/2$ endpoint can fail even for a rank-zero singleton family.

## Label-Efficient Document-Grounded Factuality Detection via Cross-Paraphrastic Invariance Learning

DevFeed: [Label-Efficient Document-Grounded Factuality Detection via Cross-Paraphrastic Invariance Learning](<https://devfeed.tech/articles/label-efficient-document-grounded-factuality-detection-via-cross-paraphrastic-invariance-learning-132211.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.16264>)

Author: Chao Chen, Shanshan Lin, Dongsheng Hong, Sibo Ju, Wenjie Yang, Sihong Xie, Xiangwen Liao

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

Content type: article

Language: en

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

Topics: [Factuality scoring in AI](<https://devfeed.tech/topics/factuality-scoring-in-ai.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [classification](<https://devfeed.tech/tags/classification.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [detection](<https://devfeed.tech/tags/detection.md>), [invariance](<https://devfeed.tech/tags/invariance.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [learning](<https://devfeed.tech/tags/learning.md>), [quantile](<https://devfeed.tech/tags/quantile.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

CPIL turns a small labeled set into relational training pairs: paraphrastic views of the same document-claim instance form positives, while opposite-label examples tied to the same source document form negatives. Its two-stage Siamese training uses quantile-aware semi-hard negative mining, then fine-tunes the encoder with a linear classifier. At inference, it needs no paraphrase generation or pair comparison--just one encoder-classifier pass. On 11 LLM-AggreFact tasks, using 2% and 3% of the FactCG-released labeled pool yielded macro-F1 scores of 81.04 and 81.60, respectively; the paper reports both outperforming compared compact detectors and two vanilla LLM references. These results are reported in the supplied abstract.

### Source excerpt

arXiv:2611.16264v1 Announce Type: cross Abstract: Large language models (LLMs) can generate fluent claims that are unsupported by a given source document, motivating accurate and efficient document-grounded factuality detection. Existing LLM-based evaluators often require repeated inference and incur high computational cost, whereas compact factuality detectors are efficient at inference time but commonly rely on large labeled or synthetically generated corpora. We propose CPIL (Cross-Paraphrastic Invariance Learning), a two-stage Siamese framework that converts a small labeled set into relational supervision without acquiring additional factuality labels. CPIL constructs positive pairs from distinct label-preserving paraphrastic views of the same document-claim instance and negative pairs from opposite-label examples associated with the same source document. It further introduces quantile-aware semi-hard negative mining, which derives a boundary from the positive-pair similarity distribution and balances informative negatives with broader coverage. In the two-stage training procedure, Stage 1 learns an encoder from these task-aware pairs, and Stage 2 jointly fine-tunes the encoder and a linear classification head for binary factuality prediction. At inference time, CPIL requires no paraphrase generation, pair construction, or Siamese comparison and uses only one encoder-classifier forward pass. On the 11-task LLM-AggreFact benchmark, CPIL-2% and CPIL-3% achieve macro-F1 scores of 81.04 and 81.60 while using only 2-3% of the FactCG-released labeled pool. Both settings outperform all compared compact detectors and two vanilla LLM references. Ablations and sensitivity analyses support the contributions of task-aware pair construction, dynamic negative mining, and multi-pivot paraphrasing.

## Parsed, Plausible, Wrong: Output Format Controls Silent Errors in Small-Model Semantic Parsing

DevFeed: [Parsed, Plausible, Wrong: Output Format Controls Silent Errors in Small-Model Semantic Parsing](<https://devfeed.tech/articles/parsed-plausible-wrong-output-format-controls-silent-errors-in-small-model-semantic-parsing-133186.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17169>)

Author: Umar Ahmed, Maxim Dunaev, Mohammad Nour Al Awad, Sergey Ivanov, Ivan Khodnenko

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

Content type: article

Language: en

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

Topics: [post-training](<https://devfeed.tech/topics/post-training.md>), [Parsing](<https://devfeed.tech/topics/parsing.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [all](<https://devfeed.tech/tags/all.md>), [also](<https://devfeed.tech/tags/also.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [apply](<https://devfeed.tech/tags/apply.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bootstrap](<https://devfeed.tech/tags/bootstrap.md>), [but](<https://devfeed.tech/tags/but.md>), [can](<https://devfeed.tech/tags/can.md>), [choice](<https://devfeed.tech/tags/choice.md>), [controls](<https://devfeed.tech/tags/controls.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [design](<https://devfeed.tech/tags/design.md>), [errors](<https://devfeed.tech/tags/errors.md>), [format](<https://devfeed.tech/tags/format.md>), [model](<https://devfeed.tech/tags/model.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [plausible](<https://devfeed.tech/tags/plausible.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [small](<https://devfeed.tech/tags/small.md>)

### AI overview

On the PIZZA ordering benchmark, encoding parsed orders in ordinary words reduced silent errors for both tested small models; simply shortening the output did not. Across 300 runs comparing six lossless tree encodings, the word format was most accurate on Qwen3.5-0.8B (87.5% exact match versus 85.8% for native s-expressions), used 4.1 times fewer tokens, and reduced silent errors 2.3-fold on Qwen and 1.8-fold on LFM2.5-350M. The reported gain was often tied to handling negation. Models also handled 95.9-99.0% of held-out off-menu inputs and learned in-utterance corrections, but strict refusal rejected 2.1-10.1% of valid, mostly polite orders. These findings are from the PIZZA benchmark and the supplied abstract.

### Source excerpt

arXiv:2611.17169v1 Announce Type: cross Abstract: A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask whether the output format, a free design choice, controls how often this happens. On the PIZZA ordering benchmark we fine-tune two sub-billion models (Qwen3.5-0.8B and LFM2.5-350M) with LoRA on six lossless encodings of the same target tree, from 44.3 down to 10.9 tokens per order, including two matched-length pairs that vary surface form at fixed length. Across 300 runs (5 training pool sizes, 5 seeds) with a hierarchical bootstrap and Holm correction fixed before the results were read, shortening alone never helps, but writing the target in ordinary words does. The word encoding is the most accurate format on both models (87.5% vs. 85.8% exact match for the native s-expression on Qwen) while emitting 4.1x fewer tokens, and it cuts the silent error rate by 2.3x on Qwen and 1.8x on LFM2, separating at every pool size. Error analysis locates much of the gain in negation. All runs also learn to refuse off-menu input (held out: 95.9-99.0% handled) and to apply in-utterance self-corrections. Strict refusal has a measured price: 2.1-10.1% of real, mostly polite, orders are refused.

## SPADE: Soft Prompts for Ai-text DEtection

DevFeed: [SPADE: Soft Prompts for Ai-text DEtection](<https://devfeed.tech/articles/spade-soft-prompts-for-ai-text-detection-132292.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.16929>)

Author: Alex Jiang, Didier Schwab, Chuyuan Li

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

Content type: article

Language: en

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

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [detection](<https://devfeed.tech/tags/detection.md>), [influence](<https://devfeed.tech/tags/influence.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [ood](<https://devfeed.tech/tags/ood.md>), [performance](<https://devfeed.tech/tags/performance.md>), [pmi](<https://devfeed.tech/tags/pmi.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [text](<https://devfeed.tech/tags/text.md>)

### AI overview

SPADE detects AI-generated text by using prompt influence as a discriminative signal, aiming to transfer better across datasets. It defines a pointwise mutual information (PMI) term to measure how much a prompt explains a text, trains one soft prompt per class to increase separation, and predicts the class whose prompt best explains a candidate. The abstract contrasts this with supervised detectors that learn from labeled texts and zero-shot detectors that rely on text probability distributions. It reports experiments across several benchmarks and state-of-the-art performance on out-of-distribution datasets; no experiment details or limitations are included in the supplied excerpt.

### Source excerpt

arXiv:2611.16929v1 Announce Type: cross Abstract: The easy access to Large Language Models (LLMs) has made AI-generated texts on the web increasingly ubiquitous, making AI-text detection a crucial task. Prompts greatly influence the generation process, but existing detectors focus solely on the generated texts --- supervised methods learn from labeled texts only, and zero-shot methods rely on the text's probability distributions --- ignoring the underlying generation process, which limits their robustness in open-world scenarios. We introduce SPADE: Soft Prompts for Ai-text DEtection, a novel framework for detecting AI-generated text that transfers better across datasets. Specifically, we measure the prompt's influence and use it as a discriminative signal. We characterize the prompt's influence by introducing a Pointwise Mutual Information (PMI) term that captures how much a prompt explains a text. Based on this quantity, SPADE trains one soft prompt per class to amplify the separation, and predicts the class whose prompt best explains the candidate text. Extensive experiments on several benchmarks against various detectors demonstrate the robustness and effectiveness of our method, achieving state-of-the-art (SOTA) performance on out-of-distribution (OOD) datasets.

## Fine-Grained Generative Hallucination Detection and Correction for Aerial Image Captioning

DevFeed: [Fine-Grained Generative Hallucination Detection and Correction for Aerial Image Captioning](<https://devfeed.tech/articles/fine-grained-generative-hallucination-detection-and-correction-for-aerial-image-captioning-132174.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.15809>)

Author: Debolena Basak, Somsubhra De, Maunendra Sankar Desarkar

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

Content type: article

Language: en

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

Topics: [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [aerial-imagery](<https://devfeed.tech/tags/aerial-imagery.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [gpt-4](<https://devfeed.tech/tags/gpt-4.md>), [hallucination-detection](<https://devfeed.tech/tags/hallucination-detection.md>), [image](<https://devfeed.tech/tags/image.md>), [lightweight](<https://devfeed.tech/tags/lightweight.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [real-time](<https://devfeed.tech/tags/real-time.md>), [reference](<https://devfeed.tech/tags/reference.md>)

### AI overview

The paper proposes a lightweight, reference-free framework to detect and correct hallucinations in free-form aerial image captions. It uses InternVL3 and Qwen2.5-VL to identify unsupported content, then applies a refinement module to improve factual consistency. The authors cite aerial imagery's small objects, viewpoint variation, class imbalance, and ambiguous entities as challenges, and contrast their approach with costly external models and rigid binary detection. Experiments on the Labelled Illusion Dataset reportedly reduce hallucination rates, with the framework described as practical for real-time use on resource-constrained systems; the excerpt gives no quantitative results or detailed correction procedure.

### Source excerpt

arXiv:2611.15809v1 Announce Type: cross Abstract: Aerial imagery captured by unmanned aerial vehicles (UAVs) plays a vital role in applications such as surveillance, disaster response, and large-scale environmental monitoring. Automatic generation of textual descriptions for such imagery is essential for reducing manual inspection overhead and enabling better downstream decision-making. However, Multimodal Large Language Models (MLLMs), despite their impressive captioning capabilities, are prone to hallucination, producing descriptions that are inconsistent with the visual input. This problem is significantly amplified in aerial imagery due to its unique challenges such as extreme viewpoint variation, small objects, class imbalance, and ambiguous visual entities. Existing hallucination detection approaches either rely on computationally expensive large external models such as GPT-4, which perform poorly on aerial domains, or treat the problem as a rigid binary classification task, ignoring the inherent uncertainty in ambiguous cases. To address these limitations, we propose a lightweight, reference-free framework for hallucination detection and correction in free-form aerial image captioning. Our method leverages state-of-the-art MLLMs, specifically InternVL3 and Qwen2.5-VL, to detect hallucinated content in generated captions with minimal computational overhead. Detected hallucinations are corrected through a dedicated refinement module that improves the factual consistency of the original captions. Extensive experiments on the Labelled Illusion Dataset (LID) demonstrate that our framework significantly reduces hallucination rates while remaining practical for real-time deployment in resource-constrained environments. Code is available at: https://github.com/Debolena7/Hal_Detect.

## Al-Ghasali: A Computational Framework for Simulating Islamic Jurisprudential Rulings via Finite State Automata and Large Language Models

DevFeed: [Al-Ghasali: A Computational Framework for Simulating Islamic Jurisprudential Rulings via Finite State Automata and Large Language Models](<https://devfeed.tech/articles/al-ghasali-a-computational-framework-for-simulating-islamic-jurisprudential-rulings-via-finite-state-automata-and-large-language-models-132323.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17235>)

Author: Elnaserledinellah Mahmoud Abdelwahab

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

Content type: article

Language: en

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

Topics: [Finite-state machine](<https://devfeed.tech/topics/finite-state-machine.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>)

Tags: [architecture](<https://devfeed.tech/tags/architecture.md>), [compilation](<https://devfeed.tech/tags/compilation.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [controls](<https://devfeed.tech/tags/controls.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-fl](<https://devfeed.tech/tags/cs-fl.md>), [decidability](<https://devfeed.tech/tags/decidability.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [define](<https://devfeed.tech/tags/define.md>), [design](<https://devfeed.tech/tags/design.md>), [finite-state](<https://devfeed.tech/tags/finite-state.md>), [framework](<https://devfeed.tech/tags/framework.md>), [grammar](<https://devfeed.tech/tags/grammar.md>), [guarantees](<https://devfeed.tech/tags/guarantees.md>), [hierarchy](<https://devfeed.tech/tags/hierarchy.md>), [history](<https://devfeed.tech/tags/history.md>), [integration](<https://devfeed.tech/tags/integration.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [linear](<https://devfeed.tech/tags/linear.md>), [llm](<https://devfeed.tech/tags/llm.md>)

### AI overview

Al-Ghasali models closed Islamic jurisprudential chapters as deterministic finite state automata compiled from finite propositional Horn theories. The paper states this makes verdict extraction decidable and O(1) after compilation, including for hierarchical compositions; requirements such as Tartib are encoded as static states. For readable summaries, the FSA controls the LLM, whose output is constrained by a grammar built from the automaton's execution history, a bounded neutral lexicon, and valence locking. The authors claim these safeguards rule out decisional and juristic factual hallucination and semantic inversion. The supplied text is an abstract and does not provide implementation or empirical evaluation details.

### Source excerpt

arXiv:2611.17235v1 Announce Type: new Abstract: This paper presents Al Ghasali, a computational framework designed to formalize and simulate Islamic jurisprudential (Fiqh) rulings. Building upon recent set theoretic formalizations of al sabr wa al Taqsim that extract minimal operational rules as simple Horn clause logical programs, this work proves that the semantics of closed Fiqh chapters collapse completely to deterministic Finite State Automata (FSA). We establish a formal computational hierarchy, demonstrating that restricting formalization to finite propositional Horn theories guarantees decidability and constant time (O(1)) verdict extraction after offline compilation. The framework formally proves that hierarchical compositions of sub-chapter FSAs preserve this O(1) complexity, and that temporal juristic requirements, such as Tartib (order), are deterministically flattened into static ontological states, preserving the memoryless property of the automaton. To generate human-readable juristic summaries, Al Ghasali incorporates Large Language Models (LLMs) via an architecture where the deterministic FSA strictly controls the probabilistic LLM. Crucially, the paper formulates sufficient design safety conditions for this FSA LLM integration, ensuring strict adherence to the epistemological boundaries of Usul al Fiqh. To neutralize the risk of LLM hallucination, the system employs a Grammar Guided Constrained Decoding mechanism. We formally define the mapping function that dynamically constructs unambiguous right linear grammars from the FSA execution history. The paper mathematically proves that this architecture, augmented by a Bounded Neutral Lexicon and Valence Locking, guarantees the absolute absence of Decisional Hallucination, Juristic Factual Hallucination, and Semantic Inversion reducing the LLM to a pure linguistic transformer of deterministic juristic facts

[Next page](<https://devfeed.tech/topics/hallucination-detection.md?cursor=WyIyMDI2LTExLTEzVDA1OjAwOjAwKzAwOjAwIiwgIjUxNDUwODBjLTRlZDAtNDBjNi05MTcwLTc0OTc2NTc2YWFlMCJd>)