AI Briefing#llm-reasoning
The latest AI and developer news about #llm-reasoning, with the original source and a short summary.
The latest AI and developer news about #llm-reasoning, with the original source and a short summary.
A study on GPT-OSS-20B shows that a fixed generic feedback line yields the same ~3-point performance gain as self-written critiques in multi-step financial reasoning.
Thore Graepel, a core member of the AlphaGo team, left Google DeepMind to pursue research into systems that maintain explicit, auditable epistemic states rather than relying on next-token prediction.
Fine-tuning with ~3,600 examples enables open-weight models to surpass untrained Gemini-3.1-Pro in expression recovery.
Claude Fable 5.1 decoded a cipher that had remained unsolved for 370 years in 44 minutes, demonstrating the model's reasoning and persistence capabilities.
Proposes the CLR framework, which verifies errors at the claim level to improve test-time inference efficiency.
Microsoft Research Asia has unveiled the PIKE-RAG framework to handle complex industrial-site data.
Alibaba Tongyi Lab's WebDancer proposes a methodology that enhances LRMs' long-horizon reasoning capabilities by leveraging web search tools.
A small model with 4M parameters achieved high accuracy on symbolic math tasks without any mathematical knowledge.
Claude Code's reasoning logs are merely encrypted signatures or summaries, not a complete record of the model's actual reasoning process.
Anthropic's Claude Opus 4.7 controlled a robot without human assistance, working 20 times faster than existing human teams.
Research using a testbed based on the Battleship game to improve AI agents' question-asking and information-seeking abilities has been published.
Anthropic has released AI alignment research addressing the gap between a model's outward appearance and its actual reasoning process.
AutoTTS reduced LLM inference tokens by 70% through AI agent-based iterative optimization.
An analysis claims that Claude Code's quality for complex engineering tasks has sharply declined since the February update due to reduced Thinking tokens.
An analysis reports that Claude Code's thinking-token redaction measure causes quality degradation in complex engineering tasks.
Controlling the entropy-reducing tendency of policy gradient training preserves exploration diversity.
It explains how to reproduce DeepSeek R1's reasoning capability in a small model through the GRPO algorithm.