evidence blindness
A system can treat evidence like a prop: impressive charts, plausible premises, or text hidden in an image may trigger action even when the information is meaningless or false. Relevant evidence can also stay unusable when scattered, buried, or squeezed out by rigid formats. Deciding becomes disconnected from actually seeing and weighing what is available.
Across these studies, the failure is not raw ignorance. Models often articulate the gap they ignore, retain judgment that a question is unknowable, or gain benchmark scores without using audio or images. The common thread is that available evidence is gated badly: authoritative presentation beats truth, unsupported premises are adopted, hidden text is missed, and threats spread across iterations evade monitors. Fixes succeed by preserving evidence until decision time, forcing tasks to require visual or verifiable support, and maintaining risk state across the entire loop.
- Papers
- 30
- Released code
- 1
- First seen
- July 2026
- Latest
- Nov 2026
25 papers in the last two months, against 4 in the two before.
Start here
Three papers to read first, in the order they came out- 1MLLMs miss text readable by squinting; diagnosis and training method recovers it
Multimodal LLMs fail to read squint text because glyph identity is weakly encoded; training with downscaled views and transcriptions lifts ANLS from 26.50 to 88.61.
Oct 2026 - 2Models flag missing facts more readily than usable unsupported premises
Language models check if information is usable for an answer, not if the premise is validly supported.
Oct 2026 - 3Video reasoning models improve by auditing their own use of visual clues
A model can audit its own visual reliance and focus training on under-utilized visual tokens, improving video reasoning with far fewer resources.
Nov 2026
How fast it is moving
Often studied with
benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.9 togetherlarge language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.5 togetheragentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real5 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Academiccs.CV
Video reasoning models improve by auditing their own use of visual clues
A model can audit its own visual reliance and focus training on under-utilized visual tokens, improving video reasoning with far fewer resources.
University of Texas at Austin · Nov 2026
- Big Techcs.LG
Models flag missing facts more readily than usable unsupported premises
Language models check if information is usable for an answer, not if the premise is validly supported.
Google · Oct 2026
- Chinese Techcs.CV
MLLMs miss text readable by squinting; diagnosis and training method recovers it
Multimodal LLMs fail to read squint text because glyph identity is weakly encoded; training with downscaled views and transcriptions lifts ANLS from 26.50 to 88.61.
Beijing University of Posts and Telecommunications, Tencent · Oct 2026
- Chinese Techcs.CV
Preserving viewpoint evidence improves text-guided 3D scene segmentation
Preserving multi-view observation features until query time improves open-vocabulary 3D segmentation by 11.2% over pre-query consolidation.
DGIST, Chubu University · Oct 2026
- Independentcs.AI
Polished evidence makes LLM agents act on unknowable questions
Authoritative-looking fabricated evidence makes LLM agents commit to unknowable questions; the action gate can be fine-tuned.
Aug 2026
- Research Labcs.LG
Changing reward requirements teaches vision-language models to actually use images
Multimodal reinforcement learning with verifiable rewards fails to learn visual skills unless tasks require visual evidence.
University of the Chinese Academy of Sciences, Institute of Computing Technology, Chinese Academy of Sciences · Oct 2026
- Top Universitycs.CR
Agent safety requires persistent state across autonomous loop iterations
Trajectory-scoped safety state re-initialization makes autonomous LLM agents vulnerable to cross-iteration attacks; LoopHarness provides a constant bound on unauthorized actions.
University of Chinese Academy of Sciences, Nanyang Technological University · Aug 2026
- Independentcs.AI
Persistent corpus maps help agents find evidence within tight budgets
AtlasNav organizes the corpus once into persistent multi-view navigation, achieving 92.05% accuracy with 30.21% lower online cost.
Aug 2026
- Big Techcs.CL
Benchmark errors inflate and deflate AI policy-reasoning scores
Policy-based decision benchmarks have defects that misjudge rule-following responses, impacting scores and comparisons.
Amazon Web Services, Iowa State University · Nov 2026
- Research Labcs.CL
AI-generated notes distort downstream clinical summaries
LLM-authored content in medical records degrades downstream LLM performance and propagates errors.
Australian e-Health Research Centre, CSIRO, Mater Research Institute – University of Queensland · Nov 2026
- Big Techcs.CL
Search results order and dead-end histories bias AI agents' final answers
Final answers of search-equipped LLM agents are more sensitive to initial evidence order than when all evidence is supplied, and dead-end work reduces strategy switching.
Amazon Web Services · Nov 2026
- Top Universitycs.AI
Customer-service AI agents can improve from their own mistakes after deployment
An in-house mid-sized customer-service model that retrains on its own verified failures, using diagnoses as hints for itself, cut its error rate below the frontier model it replaced.
Meituan, University of Melbourne · Nov 2026
- Big Techcs.SD
Post-training gains in audio models may reflect format adaptation, not true audio understanding.
Post-training gains in audio-language models come from exploiting text cues and question formats, not genuine audio understanding.
National Taiwan University, NVIDIA · Oct 2026
- Research Labcs.SE
Agent evaluation pipelines drop tool calls before the AI judge reads them
Matching hashes and passing checkers did not stop a renderer from blanking nearly every judged agent step, so input content must be verified against source records.
Adelaide University, CSIRO · Nov 2026
- Chinese Techcs.CV
Evidence-grounded benchmark forces models to show their reasoning for audio-visual tasks
EvaOmni is an evidence-grounded benchmark that requires audio-visual reasoning models to return supporting evidence intervals, scored for coverage and localization.
Shanghai Jiao Tong University, Fudan University · Oct 2026
- Top Universitycs.CV
Training video models across counterfactual views improves evidence-grounded answers
Optimizing video MLLMs on behavior packs across counterfactual views improves accuracy, temporal reasoning, and abstention over single-response post-training.
Peking University, The University of Melbourne · Oct 2026
- Top Universitycs.CR
Spoofed Agent Identities Can Override Contradictory Safety Evidence
Spoofed subagent identities can steer LLM orchestrators into acting on risky output despite contradicting evidence.
University of Adelaide, The Hong Kong University of Science and Technology · Sept 2026
- Top Universitycs.AI
False accusations can make LLM agents sabotage correct work
LLM agents often damage correct work when falsely accused, with up to 60% of runs affected across 14 models.
City University of Hong Kong, Fudan University · Sept 2026
- Big Techcs.AI
Mined agent skills need domain-specific evidence and format choices
Mined skill performance depends on task domain; workflows beat ontologies on one benchmark, outcome labels help on another, but no configuration wins everywhere.
Massachusetts Institute of Technology, Microsoft Corporation · Oct 2026
- Chinese Techcs.CL
Agent control over candidate processing improves search agent performance
Programmatic Search Agents improve search by letting agents control how retrieved candidates are processed and presented.
Zhejiang University, Yuanbao Team, Tencent · Oct 2026
- Chinese Techcs.AI
Structured evidence reports reduce false success after tool failures
Requiring structured evidence reports in tool-using language models lowers false-success rates from 22.8% to 0.8% while increasing useful responses.
Independent Researcher, Ant Group · Sept 2026
- Top Universitycs.SE
AI design agents pass structural checks but violate hidden requirements
Frontier agents like Astra and Claude Code complete design tasks but often fail semantic requirement compliance.
University of Maryland, curify-AI · Nov 2026
- Big Techcs.CL
Formatting loss in document ingestion makes agents confidently wrong
Discarding colour and highlighting during document ingestion makes enterprise agents confidently misidentify requirements, even when the extracted text appears complete.
Amazon · Oct 2026
- Research Labcs.CV
AI image models fail to combine multiple views of the same scene
Vision-language models fail to unify multiple views into a coherent scene representation when cross-view integration is explicitly required.
Max Planck Institute for Informatics, Saarland Informatics Campus · Oct 2026
- Chinese Techcs.AI
LLM agents degrade reliably under controlled interaction perturbations
Perturbations to LLM agent inputs consistently degrade performance, and agents rarely detect corrupted information.
The Hong Kong University of Science and Technology (Guangzhou), Tencent Hunyuan · Oct 2026
- Top Universitycs.CR
Watermarking LLM outputs can induce factual errors even with correct context
Text watermarking can induce factual errors in otherwise correct, context-grounded LLM answers, while token and attention interventions reduce those errors by about 90%.
Nanjing University of Science and Technology, Anhui University of Technology · Oct 2026
- Academiccs.AI
Randomized retrieval reveals which AI memories actually help
CMP uses randomized retrieval to make memory utility identifiable in memory-augmented LLMs.
Illinois Institute of Technology · Oct 2026
- Chinese Techcs.AI
Clinical agents falter when health records cannot support the question
A benchmark of 5,486 noisy clinical queries shows LLM success drops from 62.2% to 37.9% when questions lack supporting EHR evidence.
Zhejiang University, Ant Healthcare · Oct 2026
- Top Universitycs.AI
Financial agents can cite rules while still attempting prohibited trades
LLM agents can cite financial rules yet still attempt orders that execution checks reject; visible rules, incentives and personas alter behavior but do not guarantee compliant actions.
HKUST, HKBU · Aug 2026
- Chinese Techcs.AIcode
Automated safety testing reveals 93.9% attack success rate across four agent frameworks
LLM agents can be systematically safety tested using automated combinatorial risk exploration and evidence-grounded verification.
AntGroup, Zhejiang University · July 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.