OpenAI has released results from its unreleased internal model, Astra, which successfully solved ten significant open problems in mathematics. The solutions were generated by the model and subsequently formalized into Lean certificates by human researchers.
OpenAI has released solutions to ten mathematical and theoretical computer science problems that had seen no progress on their main results for at least a decade, using an internal version of its next major model, Astra.
Thinking Machines Lab has released Inkling, a new multimodal mixture-of-experts model designed for token-efficient reasoning and native understanding of text, image, and audio inputs. Together AI is making the model available on its inference platform with day-zero access.
Meta announced that its Muse Spark 1.2 model achieved gold-medal performance in five STEM Olympiads and entered the top 5 on the Vals Index at $0.69/test, reportedly 3x cheaper than Kimi. Meta attributed these gains to multi-agent orchestration with parallel reasoning, noting perfect theory scores at APhO and IPhO without external tools.
AI9Stars has released G9v3-39A5B, an open weights language model designed to deliver stronger reasoning capabilities than its predecessor, ai9stars/G9v3-3B. The model features 39 billion parameters with 5 active experts and is licensed under Apache 2.0 for personal and commercial use.
A working paper by Joseph JM Walker evaluates frontier language models on an unseen multi-channel literary cryptography benchmark embedded in the physical novel "I Wrote a Book and Made a Million Dollars (I.B. Wryten)." The study finds that GPT-5.6 independently recognized the secondary communication channel and recovered approximately 95% of the embedded material on its first pass, whereas earlier models demonstrated effectively 0% ability.
A study demonstrates that frontier language models can perform consequential computation using semantically irrelevant filler tokens, creating a failure mode where reasoning is not visible in the output chain-of-thought. The research evaluated 13 models across three tasks and found that many benefit significantly from these tokens, with accuracy improvements of up to 13 percentage points.
The authors introduce PoTRE (Poly-Topological Reasoning Ensembles), a framework designed to address the limitations of standard single-stream prompting in complex reasoning tasks. PoTRE decouples inference into four distinct agents: an Adversarial Refinement Agent, a Hierarchical strategic Planning Agent, a Spectrum Search Agent, and a Direct Chain Agent.
Researchers propose Isospectral Optimization (ISO), a framework for Reinforcement Learning with Verifiable Rewards (RLVR) that addresses the missing optimization layer by keeping model weight spectra fixed while optimizing input and output singular frames.
Researchers present Ring-Zero, a stable training pipeline that scales reinforcement learning with verifiable rewards to models with one trillion parameters, addressing issues like token redundancy and poor readability found in naive scaling.
This paper argues that building stronger intelligent systems capable of open-ended innovation requires creating, stabilizing, and reusing new representational primitives rather than merely searching within fixed frames.
A developer shares progress on a conversational symbolic AI architecture developed over the past eight months, aiming to address gaps in non-neural information extraction.
The study introduces GSM-Plus-BN, a novel perturbed Bengali mathematical dataset derived from the English GSM-Plus benchmark and verified by human translators. This resource addresses the lack of systematic benchmarks for assessing model robustness in linguistically diverse regions like Bangladesh.
The authors propose Deep Interaction, an efficient human-AI interaction method designed to correct reasoning errors in large language models without requiring full response regeneration. This mechanism allows users to directly edit the original Chain-of-Thought response, preserving accurate steps while fixing mistakes.
Researchers have developed a novel large-scale reasoning dataset and the FukuyamaBench benchmark to evaluate hierarchical mechanism reasoning in large language models, addressing physical inconsistencies common in current chemical LLMs.
A new theoretical framework explains the emergence of inductive reasoning abilities in Transformer language models by studying a generalized class of inductive tasks that unifies synthetic tasks like in-context n-grams and multi-hop reasoning.
The NVIDIA Nemotron Model Reasoning Challenge invited the Kaggle community to explore techniques that improve reasoning accuracy under shared constraints of open models, benchmarks, and infrastructure. By the close of the competition, more than 5,000 active participants across 4,000 teams had generated thousands of submissions.
Researchers introduce TreeThink, an open-source Python library designed for modular, fully asynchronous tree search in neural theorem proving. It integrates established search methods with vLLM-based inference pipelines and diverse node evaluation techniques.
The authors introduce WILDTRACE, a new benchmark designed to evaluate how well models integrate evidence dispersed across distant passages in long documents. Unlike existing benchmarks that rely on planted facts or reverse-engineered chains, WILDTRACE uses 481 tasks derived from 214 naturally occurring sources, such as technical incident reports and literary narratives.
The article argues that within-class variance in language model representations is not incomplete neural collapse but rather allocated information storage governed by a specific law. Analysis of 14 models shows that macro-category structure accounts for only 4-12% of representational variance, while within-token context carries 79-91%, remaining stable across a 100x parameter range.