AI Briefing#reasoning
The latest AI and developer news about #reasoning, with the original source and a short summary.
The latest AI and developer news about #reasoning, with the original source and a short summary.
The 321B parameter model based on GLM-5.3-Flash defaults to max reasoning, requiring token usage checks before comparison.
The framework explains that RL training amortizes search costs into model weights, making single samples likely to succeed without guaranteeing that any specific chain of thought is faithful or correct.
The open-weight model achieves a 13.7s median response time versus 28.1s for the base model while maintaining comparable accuracy on five benchmarks.
The 10B-parameter model uses roughly 2-4x fewer reasoning tokens than leading peers while maintaining comparable accuracy.
GPT-6 Astra solves all 30 standard puzzles, while solve rates drop to 20% on 15x15 grids at best effort levels.
Jeeves achieves 0.935 on JevBench overall, outperforming the baseline Jev model's 0.866 by integrating reasoning chains and a block-4 diffusion drafter.
The new Swift 1.5 27B model reduces thinking tokens by 58.5% while improving accuracy, with Flash Next offering a 1.8x speed increase.
GoBench, a benchmark evaluating LLM Go skills, has been released and is gaining attention as a reasoning metric due to its high correlation with ARC-AGI 2.
Toss analyzed the discrepancy between public benchmarks and actual service performance, establishing Toss Benchmark as a domain-specific evaluation standard.
Sber released the MoE-based GigaChat-3.5-Reasoning, achieving DeepSeek V4 Flash-level performance while reducing inference token usage by 37%.
PPM is a framework that addresses the inefficiency of sparse rewards in long-horizon LLM reasoning and guarantees optimal policies.
The open-source benchmark DungeonBench has been released to evaluate LLMs' navigation capabilities in simple dungeon games.
AI2 has released BenchMIRT, a multidimensional IRT-based tool that analyzes which specific capabilities individual questions in LLM benchmarks actually measure.
IBM has released three Granite 4.2 reasoning LLMs featuring 512K context and agentic RL.
Anthropic is reportedly conducting server-side A/B tests to lower reasoning effort levels in the latest versions of Claude Code.
A position paper from ICML 2026 warns that interpreting LLM intermediate tokens as human thought processes is dangerous.
Qwen3.6-35B is faster in speed but has longer inference times, resulting in slower final response times.
A large-scale analysis of reasoning performance and cross-lingual transfer effects in non-English and multilingual environments using GRPO.
Analysis suggests that the latest LLMs are being designed to focus on reasoning capabilities rather than knowledge storage.
Vercel AI Gateway adds Grok 4.6 with 500K token support.
ARBITRAGE predicts model superiority to reduce mathematical reasoning latency by up to 2x.
OpenAI has improved GPT-5.6 Sol and expanded access to GPT-5.6 Luna for free users.
Thinking Machines Lab released an open-weight model one-quarter the size of Inkling.
AI9Stars has released a 39B open-weight language model based on Apache 2.0.