All six benchmarks: 8460 reasoning trajectories and 278 agentic attempts; native artifacts, judgments and frozen configs. Existing final results unchanged.
Organize four final Sci-LLM stages; select 12x24 step35 agentic RL and preserve intermediates
[["scicoding-agent", 100, 46, 0], ["terminal-bench-2.1", 178, 50, 0]]; SciCoding 46/100; TB 50/178. One authorized zero for native verifier timeout, with original evidence preserved. Reasoning unchanged.
All 8460 reasoning trajectories, judgments, event histories and frozen configurations. Agentic benchmarks remain queued; no results regenerated or rejudged.
8 days ago
Replace Qwen3.5-9B IT agentic evaluations with November 13 aligned rerun
All 6400 trajectories scored; full responses and Luna judgments, frozen configs and H100 TP2 provenance. arXiv and agentic results unchanged; earlier FS/HLE exports remain in history.
Replace canonical arXiv Math results with complete avg@20 evaluations for 13 checkpoints
Full trajectories and original symbolic scores, frozen configs, monthly summaries and adjudication provenance. Other benchmarks unchanged. Previous arXiv exports retained in revision history.
10 days ago
Publish complete step-180 96K evaluation with trajectories and verifier artifacts
Complete Gold and Revision text-only evaluation; Luna judgments, frozen configs and recovery provenance. Existing benchmark results preserved. No publication reruns or score changes.
Publish GPT-OSS-20B and OLMo-3-32B-Think evaluations and full trajectories
Same benchmark export layout as Qwen/SFT. OLMo recovered results complete; GPT Research and TB unresolved judgments preserved. Recovery is replacement, never best-of.
Publish GPT-OSS-120B High non-agentic results and full trajectories
Same benchmark export layout as Qwen/SFT. Research remains provisional (1799/1800 scored); SciCode complete. Agentic benchmarks not started. No reruns or score changes.
Complete Qwen35-35B agentic evaluation with full trajectories and two explicit zero adjudications
Same benchmark export layout as Qwen/SFT. SciCoding 36/100; TB 61/178 with 176 native scores and two authorized zeros. Original timeout records preserved. Prior non-agentic results unchanged.
23 days ago
Publish Qwen35-35B non-agentic and Nemotron full evaluation with trajectories
Same benchmark export layout as Qwen/SFT. Qwen non-agentic suite fully scored. Nemotron includes four explicitly user-adjudicated zeros; originals and nine retry histories preserved.
Publish completed Gemma-4-12B evaluation with recovery and adjudication provenance
Same benchmark export layout as Qwen/SFT. All six benchmark groups complete. Explicit user-adjudicated zero and native-zero adoption preserved; no publication reruns or score changes.
24 days ago
Publish Flash-0731 High evaluation trajectories, scores and completeness provenance