Agent Phase Collapse, Reasoning Exits, Preference Gaps
Three new arXiv papers map capability cliffs in agent world models, the narrow benefit of learned reasoning stops, and a 56% accuracy ceiling when agents help users build preferences.
Articles Tagged "Reasoning"
Agent Languages, Sampling Ceilings, and Abstention
Three new papers on agents inventing symbolic languages to cut reasoning tokens by 3-6x, sampling ceilings that waste inference compute, and context-engineering to double agentic abstention rates.
GPT-5.6
OpenAI's GPT-5.6 family - Sol, Terra, and Luna - sets a new Terminal-Bench 2.1 record at 91.9% with subagent Ultra mode, but remains locked to ~20 government-vetted partners as of launch.
Tandem Training, World Models, and Efficient Agents
Three new arXiv papers on making RL reasoning legible across models, fixing broken world model latent states, and training small agents to beat their teachers.
Grok 4.5
Grok 4.5 is xAI's 1.5-trillion-parameter V9 model in private beta at SpaceX and Tesla, with supplemental training on Cursor coding data and early evals claiming performance near Claude Opus 4.8.
Gemini 3.5 Pro
Google DeepMind's upcoming flagship model with a 2M-token context window and Deep Think reasoning, announced at Google I/O 2026 and expected in July.
Refusal Gaps, Prompt Bleed, and Scaling's Logic Limit
Three new papers reveal how LLM safety hinges on persona training, how prompt modules interfere in deployed agents, and why scaling alone cannot reach symbolic reasoning.
Grok 4.3 Review: xAI Bets on Price Over Prestige
Grok 4.3 slashes prices by up to 83%, adds native video input and voice cloning, and carves out a credible position as the most cost-efficient frontier model - with real caveats on coding and latency.
Quantization's Hidden Tax, Cliff Tokens, Smarter Memory
Three new arXiv papers reveal hidden costs in quantized reasoning models, single-token failure triggers, and a new framework that cuts agent memory errors by up to 79%.
AI Research: Orchestration Beats Scale, Small Models Win
Sakana Fugu tops SWE-Bench Pro by routing tasks across rival LLMs, Microsoft's 9B browser agent beats OpenAI Operator, and a 3B model from Weibo matches DeepSeek V3.2 on math.
VibeThinker-3B
WeiboAI's 3B dense reasoning model fine-tuned from Qwen2.5-Coder-3B, posting AIME 2026 scores that match DeepSeek V3.2 (671B) using the Spectrum-to-Signal training pipeline.
ERNIE 5.1
Baidu's ERNIE 5.1 is a text-focused MoE model that claims the top Chinese model slot on LMArena with 800B parameters built at 6% of comparable training costs.
