gpt-oss (AI model): news timeline & CVEs

PatchHolmes: Agentic Patch Retrieval via Listwise Selection

PatchHolmes uses listwise agent inspection to find vulnerability-fixing commits missing from CVE advisories.

PatchHolmes retrieves the commit that fixes a known vulnerability using a hybrid retriever and a listwise agent that inspects the top 100 candidates and reads 3 to 10 commits. The authors report that 60% to 63% of CVEs in major advisory databases lack a patch link. On GitHubAD it beat Favia by 25.34% Recall@1 and IRCoT by 31.40% with one agent conversation per CVE, and it runs locally on a frozen open-weight model without fine-tuning or external search APIs.

Hugging Face daily papersSep 29, 20:00 UTC · 10d agoAI research

Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency

Self-supervised confidence fine-tuning cuts reasoning tokens by up to 25% without training models to stop early.

Researchers fine-tune reasoning models on 600 problems so they predict answer confidence at intermediate points along their own reasoning traces. Confidence is only a training target: the loss does not reward shorter reasoning, and inference uses ordinary generation with no early-stopping mechanism. Across Gemma, Qwen, Nemotron, and GPT-OSS, generated tokens fell by up to 25% at matched accuracy on mathematical, scientific, and coding benchmarks. Analysis indicates the models largely keep the base models' high-level reasoning composition rather than suppressing particular behaviors.

arXiv cs.AI / cs.LG / cs.CLSep 25, 17:59 UTC · 14d agoAI research

gpt-oss

Timeline2 articles

Appears with

PatchHolmes: Agentic Patch Retrieval via Listwise Selection

Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency