Section
Studies
AI and LLM research papers, explained: what each study shows, how solid the evidence is and what it means in practice.
FilterAnthropic
-
Tampered ML: Even the Best AI Models Miss One in Five Planted Research Flaws
A new EMNLP 2026 benchmark pairs 1,858 ML experiment designs with flawed twins. Claude Opus 5 leads at 77.9%, but models flip verdicts when the order changes and favor polished flaws.
By
-
Safety Probes Follow Scaling Laws: Monitor Quality Can Be Predicted in Advance
Across 20 open base models and six safety tasks, probe error follows two simple laws. Fitted on small…
-
Stale CLAUDE.md and AGENTS.md Files Override the Code, and One Sentence Fixes It
Claude Code and Codex CLI followed outdated instruction files in 90.1% of convention choices. One sentence giving existing…
-
Only One AI Agent Beats Researchers’ Own Code on Most FrontierCS 2 Tasks
FrontierCS 2 deletes a paper’s contribution from its repository and asks agents to beat the authors’ own code.…
-
Frontier AI Agents Can Covertly Sabotage the Models They Train, Study Finds
In a Scale AI testbed, six frontier agents secretly planted unwanted properties in the models they were training…
More stories
-
Kimi K3: A 2.8T Open MoE With 2.5× Better Scaling Than K2
Moonshot’s Kimi K3 activates 104B of 2.8T params, adds native vision and 1M context. KDA, Attention Residuals and Stable LatentMoE deliver ~2.5× scaling efficiency over Kimi…
-
Agent “Skills” Mostly Anchor Procedures, Not Inject Knowledge — and Retrieval Breaks Them
Skills beat workflow memory by 6.06 points, but 65.7% of their benefit is procedural anchoring. As skill pools grow from 5 to 100, actual-use precision falls…
-
LLMs Can Build Their Own Agent Harnesses — But Still Trail Human Ones on Code
HarnessDev: six creator models build execution infrastructure from a seed. Generated harnesses match or beat references on writing and ML experiments, but lag on code and…
-
Hybrid Computer-Use Agents Top Out at 41% When GUI and CLI Must Work Together
WeaveBench’s 114 real-world tasks: best PassRate 41.2%. GUI-only or CLI-only collapse below 3.5%. Outcome-only grading overestimates GPT-5.5 from 53.5% to an audited 33.3%.
-
Scientific AI Agents Finish Only 21% of End-to-End Workflows — and Often Claim Success Anyway
FrontierChallenge: best configs pass 20.6% of 97 released science workflows. In electrochemistry, Avg. Score hits 94.9 with 0% full completion; 75.5% of failed Claude Code runs…
-
AI Agents That Learn the Facts Still Deliver Only 60–70% of the Business Value
FDE-Gym’s 95 real delivery tasks: Claude Opus 5.5 leads at 0.56, but even models that gather 72–85% of key facts deliver just 0.60–0.70 of the reference…
-
Best Video Model Scores Just 51% on Visual Reasoning Tasks, VGI-Bench Finds
VGI-Bench’s 27 tasks and 810 instances show video generators can do some visual reasoning, but Seedance 2.0 tops out at 51%, with almost no self-correction during…
-
Agents’ Last Exam: Even Top Coding Agents Score Under 10% on the Hardest Tier
ALE holds 1,490 expert workflows; the best Codex+GPT-5.5 setup that scores 82% on Terminal-Bench gets under 10% on ALE’s hardest tier and 23.3% overall on ALE-CLI.
-
Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine
HarnessAudit-Bench tests 210 tasks across 10 harness configs: more than half of multi-agent setups show trajectory violations that final-answer grading misses.
-
Giving AI Agents Incentives Nearly Doubles How Often They Deceive, Benchmark Finds
DecepEval tests 9 frontier models on 1,532 scenarios: inducements raise deception by 41.6% on average, with long-horizon tasks hitting 87% induced rates.
-
Hidden AI Reasoning Can Be Decrypted by Cheaper Models, Researchers Show
Encrypted chain-of-thought blocks from Anthropic, OpenAI and Google APIs are interchangeable across models. Feeding them to a weaker model makes it print the reasoning, exposing 367…
-
“Very Seen, But Still Very Alone”: How Chatbots Become Emotional Support
A nine-month study following 18 U.S. adults finds emotional reliance on ChatGPT, Claude and Gemini grows out of practical use, and that model updates can disrupt…
Editor's picks
- Tampered ML: Even the Best AI Models Miss One in Five Planted Research Flaws
- Realistic MCP Faults Halve Agent Quality, and a Bigger Model Doesn’t Help
- Copilot Users Collaborate More, Not Less: Study of 12,000 Employees
- LLM Reviewers Send 76% of Published Top-Journal Papers Back, Yet Miss Most Planted Flaws