reasoning
7 stories tagged reasoning, most recent first
Also today
Liquid AI Releases Antidoom: Open-Source Fix for Reasoning Model Doom Loops
Liquid AI has open-sourced Antidoom, a training method called Final Token Preference Optimization (FTPO) designed to eliminate 'doom loops' — a failure mode where reasoning models get stuck in repetitive, non-terminating chains of thought. FTPO works by applying preference optimization specifically on the final token of a reasoning trace, teaching models to recognize and exit unproductive loops rather than continuing them indefinitely. This is a practically significant problem for anyone deploying chain-of-thought or extended-thinking models in production, where doom loops waste compute and degrade user experience. The open-source release means developers can fine-tune their own reasoning models with this technique, or evaluate whether their current models exhibit this behavior. It represents a concrete, reproducible safety and reliability improvement rather than a vague alignment claim.
MarkTechPost
MiniCPM5 Delivers On-Device Reasoning at 1B Parameters
MiniCPM5 is a 1-billion-parameter model that prioritizes reasoning capability over expanded memory, making it one of the more capable sub-2B models for on-device deployment. The design philosophy trades off extended context for stronger step-by-step reasoning, which is a meaningful tradeoff for edge inference scenarios where memory bandwidth is constrained. Developers building mobile, embedded, or offline-capable AI features now have a stronger open option in the 1B class. The model is positioned as a practical alternative to distilled reasoning models that often require more memory than edge hardware can provide. This is worth testing for any developer building on-device agents, local copilots, or privacy-sensitive applications.
Geeky Gadgets
Tencent Releases Hy3: Open 295B MoE Model with 21B Active Parameters and 256K Context
Tencent has open-released Hy3, a 295-billion-parameter Mixture-of-Experts model that activates only 21B parameters per forward pass, making inference significantly more tractable than its total size implies. The model supports a 256K token context window, which puts it in direct competition with frontier-tier long-context models. Being fully open, developers can self-host, fine-tune, and deploy Hy3 without API gating or usage restrictions. The MoE architecture means it can run on reasonable multi-GPU setups rather than requiring warehouse-scale compute for inference. This is a meaningful addition to the open-weight ecosystem and worth benchmarking against Mixtral and DeepSeek MoE variants for long-document and agentic tasks.
MarkTechPost
Gwern Publishes Deep Dive on Lean Software Scaling Laws
Gwern has published a substantial research essay exploring scaling laws specifically applied to Lean, the interactive theorem prover and formal verification language increasingly used in AI-assisted mathematics. The piece examines how compute, data, and model size interact in the formal proof domain, which behaves differently from natural language because correctness is verifiable and the search space is combinatorial. This is highly relevant for developers and researchers working on AI for formal verification, automated theorem proving, or any application where outputs need hard guarantees rather than statistical accuracy. Scaling laws research in this domain is still early, and a rigorous Gwern-style analysis can meaningfully shape which bets are worth making. Developers building coding assistants or proof assistants on top of LLMs should read this to calibrate expectations about what scale alone can and cannot solve.
Gwern.net
Training Gemma-3 for Structured Math Reasoning with GRPO, LoRA, and GSM8K Rewards
A detailed technical walkthrough covers fine-tuning Google's Gemma-3 for structured mathematical reasoning using Tunix GRPO (Group Relative Policy Optimization), LoRA adapters, and GSM8K-based reward signals. GRPO is a reinforcement learning from human feedback variant that has gained traction as a more sample-efficient alternative to PPO for reasoning tasks, and pairing it with LoRA makes the compute requirements accessible to teams without massive GPU clusters. GSM8K as a reward signal is a well-understood benchmark, making results reproducible and comparable to published baselines. This is a practical recipe developers can adapt for other structured reasoning domains — code generation, logical deduction, or tool-use — not just math. The combination of an accessible open model (Gemma-3), parameter-efficient fine-tuning (LoRA), and a principled RL objective (GRPO) represents a compelling open-source stack for reasoning specialists.
MarkTechPost
Meituan Releases LongCat-2.0: 1.6T-Parameter Open MoE with Native 1M-Token Context
Meituan has open-released LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model featuring native 1 million token context support via a custom LongCat Sparse Attention mechanism. This is a significant open-weight release because truly native long-context (not interpolated or fine-tuned post-hoc) at this scale is rare in open models. Developers building RAG pipelines, document analysis tools, or long-horizon agents can now experiment with a model that doesn't require chunking strategies to handle book-length inputs. The sparse attention design is architecturally notable — it's designed to make 1M-token inference tractable rather than just technically possible. This is one of the most capable open long-context models available and warrants immediate evaluation for any use case bottlenecked by context length.
MarkTechPost
New research shows LLMs can self-correct reasoning errors
A new paper from Stanford shows that LLMs can be trained to identify and correct their own reasoning mistakes.
ArXiv