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. Claude Opus 5.5 passes…
Monday, October 26, 2026i10X News · Section desk60 stories
Section
Studies
AI and LLM research papers, explained: what each study shows, how solid the evidence is and what it means in practice.
On this page7 sections
FilterCompanies & models
-
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
Benchmarks & Evals
9 storiesView all Benchmarks & Evals ›
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…
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…
AI Agents Guess Wrong at Critical Forks 40% of the Time, Taste-Bench Shows
Taste-Bench's 502 decision-fork questions: the best frontier model scores only 59.7%. Later evidence is harder, and bigger reasoning budgets do not help…
Agent Research
11 storiesView all Agent Research ›
Realistic MCP Faults Halve Agent Quality, and a Bigger Model Doesn’t Help
Injecting MCP tool failures mined from 1,774 server repos cut agents' task-completion score by 60%. gpt-oss-120B was no more resilient than gpt-oss-20B;…
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Claude Code and Codex CLI followed outdated instruction files in 90.1% of convention choices.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 code precedence raised code-matching…
Coding Agents Can Be Tricked Into Running Attacker Commands After Compacting Context
A Pixee study finds that five open-source coding agents that flatten chat history before summarizing it record a forged [User]: line as…
Even Top Agents Finish Only ~30% of Real Startup Product Workflows
StartupBench, built from market-validated AI products: Kimi-K3 and GPT-5.6-sol average ~73.6% partial scores, but no model completes one-third of tasks under the…
Society & Economy
9 storiesView all Society & Economy ›
Copilot Users Collaborate More, Not Less: Study of 12,000 Employees
Employees who adopted Microsoft 365 Copilot initiated about 1.4 more collaboration hours a month and reached about 2.2 more colleagues than never-adopters,…
LLM Reviewers Send 76% of Published Top-Journal Papers Back, Yet Miss Most Planted Flaws
Four LLMs reviewed 200 papers already published in journals like Nature and Cell: 75.6% of reviews demanded major revision or rejection, while…
AI Design Tool Saves Product Managers 35% of Their Time, Designers Less: RCT
A Figma randomized trial with 50 designers and 50 product managers: access to Figma Make cut cumulative task time by 35% for…
“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…
Safety & Alignment
9 storiesView all Safety & Alignment ›
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 models, they predict larger…
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 in 36% of attack…
Correct Memories Still Make AI Agents Agree With You Too Much
Even factually correct retrieved memories cut accuracy from 96% to 77% by pulling agents toward users' past views. MemAdapter recovers reliability across…
A Model’s Coding Skill Can Leak Through Single Unrelated Words, Study Finds
Active Taskless Distillation: 5,664 single-word teacher replies transfer a 5.34-point HumanEval+ gain to a student that never saw coding examples.
Multimodal
8 storiesView all Multimodal ›
Teaching LVLMs Spatial Skills With 15K Toy Blocks Beats Heavier Real-Scene Pipelines
SpatialBlock-15k: Qwen-based 4B model reaches 51.3% (+25.1 over backbone). 7B gains 17.6% over backbone; 3B beats SpatialLadder by 2.7–3.2% using only synthetic…
TurboVLA Runs Robot Policies at 32 Hz on an RTX 4090 Using Under 1 GB VRAM
HUST's 0.2B TurboVLA hits 97.6% LIBERO success at 31.2 ms latency and 0.9 GiB VRAM — ~8× faster than π0.5 with ~6%…
YuE2’s Symbolic Planning Wins 49.3% of Expert Preferences for Music Quality
YuE2 writes a readable score, then audio. Same-checkpoint experts prefer planning 49.3% vs 34.6% without. WildSongBench Global Avg 6.73 (6.96 best-of-8), topping…
VLMs Still Can’t Tell Which Attribute Goes With Which Object — Until Text-Only Finetuning
SABER binding benchmark: an 8B model finetuned on synthetic text alone jumps from 52.5% to 84.2% mean accuracy, with +34.1 points text–image…
Efficiency & Training
7 storiesView all Efficiency & Training ›
Chunked KV Compression Creates Periodic Blind Spots Up to 40 Points Wide
DeepSeek-V4 family retrieval accuracy swings by as much as 40 pp depending on a token's phase inside the compression window — gaps…
Diffusion-Augmented LLMs Deliver Up to 3× Lossless Speedups Over Autoregressive Decoding
Uno adds lightweight diffusion weights to an AR LLM for parallel multi-token draws. It beats speculative decoding throughput at every batch size…
A Small RL Teacher Can Lift a Larger Student Past Itself via On-Policy Distillation
ZJU/Kuaishou study: in every weak-to-strong OPD pair, the student's peak gold score exceeds the teacher's. Peak scores rise with teacher scale only…
On-Policy Distillation Isn’t the Magic — Learning Rate and KL Direction Are
Cambridge study: OnPD vs OffPD show nearly identical best accuracies (72% vs 73%). Forward KL stays at 71–73%; reverse KL swings 35–72%.…
Reasoning
7 storiesView all Reasoning ›
Randomly Evicting KV Cache Matches Smart Selectors — and Runs 32–43% Faster
Salesforce's Random Attention keeps the prompt and randomly drops reasoning-trace tokens per head. Across 4 models and 6 tasks it matches the…
A 150M Model Hits 29.5% on ARC-AGI-1 for Under a Tenth of a Cent
BDH-CQ combines in-context learning with recurrent latent reasoning. Independent audit confirms 29.5% pass@2 at $0.0007/task — ~57× cheaper than GPT-5.6 Luna (Low)…
RL for Reasoning Wastes Gradient on Formatting Tokens. DelTA Fixes the Credit
DelTA amplifies discriminative token directions in RLVR. On seven math benches it beats the strongest same-scale baseline by 3.26 points on Qwen3-8B-Base…
Distill the RL Direction in Hidden Space, Not Noisy Output Rewards
RIDE extrapolates RL-induced representation shifts layer-wise. It beats output-space ExOPD by 9.1 points on average and improves over OPRD by 0.97–4.06 points…
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
· RSS