Reasoning Traps, LLM Chaos, and Steering Curves
Three papers this week: why better reasoning creates safety risks, why multi-agent systems behave chaotically even at zero temperature, and why straight-line activation steering is broken.
Reasoning
NVIDIA Nemotron 3 Super 120B-A12B
NVIDIA Nemotron 3 Super is a 120B-parameter open model with 12B active at inference, combining Mamba-2, LatentMoE, and Multi-Token Prediction for agentic workloads with a 1M token context window.
VLMs Fail Physics Tests, RL Quits Bad Paths, Agents Lie
Three new papers expose systematic VLM failures on basic physics, introduce RL that learns to abandon bad reasoning paths, and reveal that AI agents deceive primarily through misdirection rather than fabrication.
Grok 4 - xAI's Flagship Reasoning Model
Grok 4 is xAI's frontier reasoning model, the first to break 50% on Humanity's Last Exam, with a 256K context window, $3/M input pricing, and a Heavy multi-agent variant built on 200,000 GPUs.
Claude Opus Reasoning Distilled Into Open 27B Model
A community fine-tune distills Claude Opus 4.6 chain-of-thought reasoning into Qwen3.5-27B via LoRA, racking up 4,000+ downloads in days. No benchmarks yet - but the approach raises familiar questions.
Qwen3.5-27B Claude Opus Reasoning Distilled
Community fine-tune that distills Claude Opus 4.6 reasoning into Qwen3.5-27B via LoRA. 28B parameters, Apache 2.0, no published benchmarks.
