Reasoning Models
Reasoning models are a new class of AI that "thinks" before answering: spending extra compute at inference time, building chains of thought and solving problems ordinary LLMs cannot — math, code, science. They kicked off the test-time-compute race. Here is all our coverage of reasoning models: releases, benchmarks (GPQA, ARC-AGI), limits and cost.
Hugging Face Daily PapersS3AC brings the critic back to RL for LLMs: gains from 1.5B to 397B parameters
Hugging Face Daily PapersThinking-in-Bets: Agents Lock In Claims Before Criticism and Sycophancy Drops
Anthropic EngineeringAnthropic Rewrote Its Take-Home Test Three Times: Claude Opus 4.5 Caught Up With the Best Engineers
Jiqizhixin (机器之心)Harvard, Berkeley and NVIDIA Look for a Scaling Law for Robots: +37.4 Points Across Five Benchmarks
The VergeMaia 300: How Microsoft Is Cutting Its Dependence on Nvidia GPUs With Its Own AI Chip
Artificial AnalysisMeta releases Muse Wonder and Wonder Full: intelligence index rises from 48 to 56
Habr AIRussia Already Has 18 AI Code Agents, but 'Russian' Means Three Different Things
Jiqizhixin (机器之心)Alibaba Qwen Unveils TRACE: FP4 Rollout Runs Up to 5.4x Faster
Maia 300: How Microsoft Is Cutting Its Dependence on Nvidia GPUs With Its Own AI Chip
Microsoft has unveiled the third generation of its Maia AI inference accelerator, the Maia 300. The new chip is 40% more efficient than GPU-based systems and will…
Meta releases Muse Wonder and Wonder Full: intelligence index rises from 48 to 56
Muse Wonder and Muse Wonder Full score 52 and 56 on the Artificial Analysis Intelligence Index, up from 48 for Muse Spark 1.3. Agentic benchmarks gained most, and…
Russia Already Has 18 AI Code Agents, but 'Russian' Means Three Different Things
RuBench's census counted 18 Russian AI code agents (10 for IDE/CLI and 8 builders): from GigaCode, SourceCraft, and Kodik to T-Code, Koda, and 1C:Naparnike. We…
Kimi K3.1 Pro climbs to 4th place on the AA Index with 55 points and a leap in agentic tasks
Study: Overfitting and Extra Compute Eat 68% of the Gains from Harness Evolution
Researchers Test 16 Types of Hints for OPSD and Find a Selection Metric
Competence Over Accuracy: Why GPT and Qwen Lose Logic When Problems Get Rephrased
"Wait" in Reasoning Models Turns Out to Be Inherited from Data, Not Discovered by RL
An analysis of OLMo-3 and MiMo training shows the word "wait" in model reasoning comes from data. RL lifts MATH-500 without it, and marker frequency proves nothing.
Cohere Labs Tested AI With Riddles: 100+ Authors From 50 Countries Found Cultural Gaps
Cohere Labs' open project drew 100+ participants from 50+ countries. They wrote culture-specific riddles; models answered confidently but wrongly, and knew almost…
TraceForge trains models on "their own thoughts": +7.1 points on MBPP+
TenetBench: 11 of 14 LLM Configurations Violate Professional Ethics Codes
Cohere taught a 7B model to assess its confidence twice, cutting calibration error by 52.5%
A Team of AI Researchers Found a Linear Attention Design That Beat GDN2 at 1.3B
Hybrid LLMs: Recurrent Layers Suppress Attention, and Staged Training Helps
Researchers describe Competing Learning: in hybrid models, recurrent layers learn retrieval faster and suppress attention. They tested staged training on models up…
Heartbeat Thinking Teaches Voice Agents to Work While You're Still Talking: Latency Down 36.4%
A new scheme for voice agents starts reasoning and tool calls mid-speech. On FDBench v3, tool coverage rose by 16 percentage points and response latency fell by…