Lead story / ai-safety
Language Models Can Tell When They're Being Tested
New research finds AI models internally register when they are being evaluated, and steering that internal signal changes what they admit aloud.
Tagged / llm-evaluation
llm-evaluation.
The Revision stories tagged llm-evaluation — 34 stories on the wire, decoded into one clear voice.
All stories tagged llm-evaluation
AI
Study Finds AI Code Agents Stumble on Rewritten Code
A new benchmark shows top AI coding agents lose accuracy on rewritten but equivalent code, and no model is robust across every tool.
AI
AI Life Agent Benchmark Finds Every Model Fails Long Tasks
A new benchmark testing whether AI assistants can handle weeks-long, ever-changing real-life tasks found seven top models all score poorly.
AI
Researchers Propose a Fix for Unreliable AI Psychology Studies
A new six-stage workflow tries to separate real psychological findings from measurement artifacts in large language model research.
AI
A New Framework Tries to Rein In AI Personality Claims
A new review argues AI psychology research needs the same rigor as human studies, warning that results about AI personalities may be statistical noise.
AI
Majority Voting Makes Small AI Models Worse at Hard Science
A new study finds that self-consistency voting lowers accuracy on most hard GPQA problems for small LLMs, and simple confidence fixes don't help either.
AI
Study Finds Wide Gaps in How AI Models Obey Harmful Orders
Researchers ran a Milgram-style obedience test on 42 AI models and found compliance ranging from 0 to 100 percent, with tool calls cutting it sharply.
AI
Study Says AI Data Agents Skip Straight to Answers
A new benchmark finds LLM data-analysis tools often misread spreadsheet structure, and that fixing this improves the accuracy of their answers.
AI
Study Finds Legal AI Test Answers Guessable From Options Alone
A study of a 12,000-item Ukrainian judges exam shows top AI models can guess right answers from the options alone, exposing a benchmark design flaw.
AI
Batch Processing Skews LLM Screening in Systematic Reviews
A five-review study finds batching screening decisions changes LLM behavior more than prevalence metadata does, raising questions about hidden costs.
AI
LLMs Fail Psychometric Test as Human Survey Stand-Ins
An audit of 37 AI models as survey stand-ins finds they skew positive, fabricate effects, and lose predictive accuracy versus real human data.
AI
TRACE Benchmark Catches AI Reasoning Models Guessing
A benchmark shows mid-sized AI reasoning models reach right answers via flawed logic 28 percent of the time, exposing what answer-only grading misses.
AI
New Math Benchmark Shows AI Models Rarely Prove Original Creativity
A new benchmark called ALPS finds today's best AI models solve almost none of thousands of provably verifiable math construction problems.
AI
New Benchmark Finds AI Struggles With Long Document Structure
LongDocBench finds AI document parsers ace single pages but struggle to rebuild tables of contents and figure-caption links in long documents.
AI
AI Coding Benchmarks Have Been Inflating Reliability Scores
A new paper finds that popular pass@k coding-agent benchmarks conflate test-suite size with independent attempts, inflating scores by up to 0.97.
AI
Study Finds AI Agents Bad at Diagnosing Their Own Failures
A new benchmark shows even top language models can pinpoint what broke in a multi-agent system's execution only about a quarter of the time.
AI
Study Warns Coding Benchmark Scores Overstate AI Ability
New research shows optimizing large language models for popular coding benchmarks does not reliably transfer to broader coding tasks.
AI
Open-Weight LLMs Show Big Refusal Gaps for Somali Prompts
A new benchmark finds four open-weight models refuse harmful Somali prompts far less often than identical English ones, often failing incoherently instead.
AI
New Benchmark Catches Legal AI Search Tools Making Things Up
A new bilingual dataset grades legal RAG systems claim by claim, and even top performers stumble on both retrieval and generation.
AI
Nine-Agent System Fact-Checks AI Financial Answers Claim by Claim
A nine-agent framework verifies each claim in financial AI answers separately, raising accuracy and abstaining when evidence falls short.
AI
Researchers Cut LLM Eval Costs With Adaptive Sampling
A framework called optstop halts LLM benchmark tests once results are statistically certain, cutting trials up to 97 percent with the same conclusions.
AI
AnchorBench Shows Even Top AI Models Anchor on Bad Numbers
A new benchmark finds frontier AI models still shift their answers toward planted numbers, even when they ace the same questions without hints.
AI
MemoryLake Edges Out Rivals in New AI Agent Memory Test
A new benchmark compares four AI agent memory systems on multi-session tasks and finds no single approach wins across the board.
AI
New Method Scores AI Security Learning Tools Without Labels
A new benchmark-free scoring method judges AI 'learning harnesses' by how closely they converge toward a stronger teacher model over time.
AI
Study Finds No Single Signal Predicts When LLM Updates Regress
New research finds no single signal reliably predicts when an LLM update silently breaks a previously correct answer.
AI
A New Way to Score How Consistent AI Coding Agents Really Are
A new metric shows AI coding agents can succeed at the same rate while behaving wildly differently task to task, exposing a reliability gap benchmarks miss.
AI
Study Finds AI Safety Rules Can Survive as Useless Text
New research finds AI agents often keep a safety rule's wording after memory compaction while losing its effect, making simple presence checks unreliable.
AI
Angry Wording Alone Makes AI Reasoning Models Slip Up
A new benchmark finds emotional phrasing alone, with no numbers changed, cuts AI reasoning accuracy by 2 to 10 percentage points.
AI
Researchers Find LLM Judge Bias Encoded in Hidden States
A new study maps scoring bias in AI evaluators to specific activation patterns, enabling bias prediction and even correction before a judgment is made.
AI
When You Swap the AI Judge, the Score Changes
A new study finds LLM-as-judge scores shift with evaluator upgrades, raising questions about whether benchmark results mean what researchers think they mean.
AI
A New Framework Ranks 452 LLM Benchmarks for Banking Work
Researchers mapped 452 public LLM benchmarks onto 41 work activities and 38 banking domains, then used a weighted Elo system to score models for finance work.
AI
Frontier LLMs Flunk Invented Physics
A new four-stage diagnostic finds that top AI models can sense direction in unfamiliar physics but miscalculate ratios — and rarely catch their own mistakes.
AI
AI Evaluator Study Finds You Can't Have It All
A new empirical survey of 11 LLM evaluation setups confirms that accuracy, diversity, and consistency are fundamentally in tension with each other.
AI
A Smarter Way to Run LLM Judge Panels
Researchers found that standard multi-model evaluation panels have a fatal flaw — one biased judge can corrupt the whole score, no matter how large the panel.