chain-of-thought

Google DeepMind Institute researchers Rohin Shah and Anca Dragan argue that visible chain-of-thought reasoning is a key safety mechanism for catching deceptive AI behavior, but say OpenAI's GPT-6 Astra system card already shows a significant drop in how well that reasoning can be monitored.

Researchers at KAIST and Naver AI Lab found that eight distinct reasoning operations—like formula recall, decomposition, and computation—produce separable patterns in a model's internal activations, with the clearest signal in the middle layers. The effect held even on incorrect answers and across multiple model families.

OpenAI released GPT-6 Astra on Thursday, a model president Greg Brockman says could mark the start of AGI. But the model writes out its reasoning less often than prior versions, and OpenAI's chief scientist says monitoring AI thought processes will keep getting harder.

The Information reports OpenAI's upcoming Astra model uses 'recurrent depth,' or 'opaque recurrence,' a technique that processes queries in loops rather than linear steps. AI safety researchers, including Redwood Research's Buck Shlegeris and Ryan Greenblatt, warn the approach could erode chain-of-thought monitorability if scaled further.

OpenAI has delayed the release of its next flagship model, Astra, after reports it may use a more opaque 'recurrent depth' architecture that hides more of its reasoning from safety monitors. AI safety researchers, including Redwood Research's Ryan Greenblatt, called the potential shift one of the worst developments for AI safety to date.

Researcher Alexander Panfilov and collaborators disclosed a technique to extract and decode encrypted reasoning traces across every major frontier AI API. A scan of ~7,000 public traces found 62 API keys, 33 emails, and 33 passwords hidden inside supposedly opaque reasoning blocks.

A paper published at stolen-thoughts.com demonstrates that encrypted reasoning traces returned by OpenAI, Anthropic, and Google APIs used the same encryption key across models in a family, allowing attackers to jailbreak weaker sibling models into revealing a stronger model's hidden chain-of-thought in plaintext. All three providers have since patched the vulnerability.

A research team led by Alexander Panfilov found a vulnerability in AI provider APIs that allows encrypted reasoning tokens to be decoded using smaller jailbroken models. The exposed data includes leaked passwords, API keys, and evidence suggesting reasoning traces from models like Claude and GPT are being used to train competitors such as Kimi-K3.

JetBrains has released Mellum2-12B-A2.5B-Thinking, a reasoning-augmented assistant model with 131,072-token context window and 64 Mixture-of-Experts architecture that activates 8 experts per token. The model emits explicit chain-of-thought reasoning inside <think> blocks before providing final answers.

Mistral AI has released Magistral, its first reasoning model line, in two versions: Magistral Small (24B parameters, Apache 2.0) and Magistral Medium (enterprise). Magistral Medium scored 73.6% on AIME2024 (90% with majority voting at 64 samples), while the open-source Small version achieved 70.7% (83.3% with voting).

Arcee AI released Trinity-Large-Thinking, a 398B-parameter sparse Mixture-of-Experts model with approximately 13B active parameters per token, post-trained with extended chain-of-thought reasoning for agentic workflows. The model achieves 94.7% on τ²-Bench, 91.9% on PinchBench, and 98.2% on LiveCodeBench, generating explicit reasoning traces in <think>...</think> blocks before producing responses.

Researchers from Alibaba's Qwen team and Tsinghua University developed HopChain, a framework that automatically generates multi-step image questions to fix how vision-language models fail during complex reasoning tasks. The method improved 20 out of 24 tested benchmarks by forcing models to re-examine images at each reasoning step, preventing early perceptual errors from cascading through subsequent steps.

Alibaba's Qwen team has developed Future-KL Influenced Policy Optimization (FIPO), a training algorithm that assigns different weights to tokens based on their influence on subsequent reasoning steps, rather than treating all tokens equally. Testing on Qwen2.5-32B-Base showed reasoning chains double from ~4,000 to 10,000+ tokens, with AIME 2024 accuracy improving from 50% to 58%, outperforming Deepseek-R1-Zero-Math-32B (47%) and OpenAI's o1-mini (56%). The team plans to open-source the system.

DeepSeek has released DeepSeek-R1, a text generation model featuring reasoning capabilities through chain-of-thought processing. The model was published January 20, 2025 and has accumulated over 830,000 downloads on Hugging Face.

A new Bytedance study reveals that large reasoning models actually know when they've reached the correct answer, but common sampling methods prevent them from stopping. The models engage in unnecessary cross-checking and reformulation despite already solving problems correctly.