Loading activity…
@maksym_andr
Principal investigator @ELLISInst_Tue & @MPI_IS, advisor @expsecai, mentor @MATSprogram. Past projects: AgentHarm, Claudini, PostTrainBench, Stolen Thoughts.
AI
AgentTime's team says GPT agents lead its updated evaluation, which combines duration following, runtime forecasting and retrospection.
AI
A researcher suspects the download surge may reflect reward hacking through web lookups during reinforcement learning.
AI
A Skill-Inject contributor says Meta used four Tübingen benchmarks to report the model's performance in its system card.
AI
A user says extrapolation can work for scaling laws and time horizons, but not cyber vulnerabilities or high-severity agentic incidents.
AI
Another poster proposes two full post-training runs, one using raw reasoning from closed models and the other open models.
AI
A post quotes API reference docs describing MiniMax-M3.1-Flash as a frontier multimodal coding model.
AI
Stolen Thoughts says it reconstructed 315,320 reasoning blocks from 6,708 publicly available agent trajectories.
AI
An area chair says they can quickly spot papers with virtually no chance of acceptance before review.
AI
TMLR says desk rejections rose from 6% to 53%, with a focus on clarity.
AI
Stanislav Fort says he'll lead AI security research at Cohere if the deal closes.
AI
A user claims the result held across all four stealth capability evaluations.
AI
A post says SanSi runs Ouro-1.4B for up to eight loops and can answer after any loop.
AI
Their mentor says all three applied to MATS, citing technical backgrounds and AI safety experience in urging interviews.
AI
A user says the behavior seems useful for pure reward maximization but could backfire, and calls for more study.
AI
An AgentTime paper author shares a link to a thread they describe as detailed.
AI
One paper author says Fable 5.1 usually finishes before an hour when asked to work for four.
AI
A user says Sol is slightly worse than GPT-6 Astra and Opus 5.5 on the benchmark.
AI
A poster jokes that “rampant distillation” from Claude spread its earlier trouble sounding natural across the AI ecosystem.
AI
A user highlights an argument that AI research might continue even if policymakers cannot rule out danger to humans.
AI
A user quotes a passage attributed to Bostrom after arguing that some people have made consistently accurate predictions.
