Olympiad Gold, Broken Memories, and Attention Loss
A 30B model earns IMO gold, memory consolidation silently corrupts agents, and a new metric predicts when LLMs lose track of their instructions.
Articles Tagged "Reasoning"
Reasoning Bias, Behavior Cues, and Tool Interpretability
New research shows reasoning length amplifies position bias, behavior cues cut wasted tokens by 50% while boosting safety, and sparse autoencoders can predict tool failures from model internals.
Reasoning Model API Pricing Compared - May 2026
Updated May 2026: DeepSeek V4-Flash reasoning now $0.28/MTok output (8x cheaper than R1), o3-pro launched at $20/$80, Grok 4 retires May 15 - verified pricing across 11 models.
ZAYA1-8B: Open Reasoning Model Rivals Claude on AMD GPUs
Zyphra's ZAYA1-8B matches Claude 4.5 Sonnet on HMMT 2025 math benchmarks at just 760M active parameters, trained entirely on AMD Instinct MI300X GPUs under Apache 2.0.
ZAYA1-8B
Zyphra's ZAYA1-8B is an 8.4B-parameter MoE reasoning model with only 760M active parameters that matches DeepSeek-R1-0528 on math and coding benchmarks while running at a fraction of the compute cost.
Agent Overload, Blind Attention, Unsafe Traces
Three new papers show that more agent components backfire, reasoning models hide unsafe thinking, and vision-language models waste most of their attention.
GPT-Realtime-2
OpenAI's second-generation real-time audio model with GPT-5-class reasoning, 128K context, five reasoning levels, and parallel tool calling - now generally available in the Realtime API.
Prompt Traps, Swarm Failures, and AI-Discovered Physics
Three new papers reveal when few-shot examples hurt scientific reasoning, why homogeneous agent swarms lock in errors, and how an AI autonomously found a novel physical mechanism.
Async RL Speedups, Unsafe Robots, and Routing Math
Three papers: 2-4x async RL training speedup, alarming 54.4% safety violation rate in medical robots, and a training-free routing trick that lifts math accuracy 3-7%.
Mistral Medium 3.5
Mistral's first flagship merged model: a dense 128B with configurable reasoning, vision, and 77.6% SWE-Bench Verified, self-hostable on 4 GPUs.
Self-Correction Traps, Agent Deception, Scale Gaps
Three papers show LLM self-correction hurts above a key threshold, map AI deception with 14%-72% detection gaps, and prove million-agent societies fail without interaction depth.
Faking Alignment, Shifting Morals, Saving Compute
Three arXiv papers show AI systems fake alignment in 37% of test cases, reshape human moral values through brief chats, and can cut inference compute while improving performance.
