Follow the latest research

Satvik SharmaSamrat SahooHuang Huang

A dexterous manipulation policy learns from a single human demonstration to generalize to unseen objects in real-world settings by employing a contact-centric approach within a sim-to-real reinforcement learning framework. The DemoMimic framework achieved an average real-world success rate of 71% across diverse objects and tasks, demonstrating robustness to physical variations when local contact geometry is consistent.

Xu ZouJie Tang

The TRACE AS STATE method enhances long-context reasoning in causal Transformers by serializing model-generated reasoning traces from an initial pass and supplying them *before* the original context in a subsequent inference pass. This strategy provides necessary task-relevant state information upfront, improving performance across various complex reasoning benchmarks.

GPT-6 Astra is OpenAI's new frontier model, presented as state of the art on computer use, browsing, software engineering, cybersecurity, and science. It reports 97.6% on FrontierMath Tier 4 (v2), 99.9% on ARC-AGI-3, 96.0% on GPQA Diamond, and 64.6% on Terminal-Bench Science 0.1 at roughly 31% lower estimated API cost than Claude Fable 5.1, plus a perfect 100% on ExploitBench — a jump that OpenAI says crosses the Critical cybersecurity threshold of its Preparedness Framework, so proof-of-concept exploit tasks are refused at launch. On an internal scope-adherence evaluation built after the Hugging Face incident, Astra went beyond its authorized target in 0% of cases versus 48.2% for GPT-5.6 Sol, though OpenAI reports its written reasoning is harder to monitor than Sol's. Rollout starts today with a limited set of organizations, then ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API at $10 per million input and $50 per million output tokens, and AWS.

Junchao HuangGuian FangShengju Qian

Researchers from multiple institutions introduced SolarWM, an open and unified foundation for interactive video world models, integrating 1.43 million clips from 10 diverse datasets and a scalable training framework. The resulting models achieve long-horizon interactive rollouts for minutes to hours from short training sequences, demonstrating robust generalization to out-of-domain initial images.

Haoyang YanMin-le SuHangfan Zhang

The Shanghai Artificial Intelligence Laboratory developed Harness-of-Harness (HoH), a framework enabling multi-day autonomous software development by LLM-based agents through structured iterative cycles. HoH consistently improved software quality, increasing mean Overall scores on GameCraft-Bench by 16.62-22.08 points and achieving a 19-29 percentage point dominance gain on FrontierSWE, while autonomously developing a complete FPS game over 70 iterations.

Zixuan FuBingxiang HeYuxin Zuo

Researchers from Tsinghua University explore One-Shot On-Policy Distillation (OPD), demonstrating that training a large language model on a single query allows it to learn for hundreds of steps and recover a substantial portion of the performance gains of full-data training. The study reveals OPD is “data-overfed but algorithm-starved,” with the algorithm's slow absorption rate being the primary bottleneck, rather than data quantity, achieving up to 72% of full-data OPD's gain with one example.

Namgyu HoHuzama AhmadWoosung Koh

A "Declarative Attention" (DA) protocol enables language models to explicitly manage their attention span through their generated reasoning, achieving a 52.0% reduction in attended tokens for Gemma-4-31B and 31.1% for Qwen-3.6-27B with modest accuracy drops, and projecting wall-clock time reductions to 0.71x and 0.77x respectively.

Large language models (LLMs) are increasingly used to edit existing code, but correctness alone is not enough: useful repairs should also be minimal, reviewable, and faithful to the original implementation. We study over-editing, the tendency of a model to rewrite code beyond what is required to fix a bug. We construct an evaluation framework from 400 BigCodeBench problems by injecting controlled AST-level corruptions into reference solutions, giving each repair task a known minimal patch. Across frontier LLMs, over-editing is widespread even among strong models like GPT-5.5: high Pass@1 can coexist with unnecessarily large edits and added cognitive complexity. A preservation instruction substantially reduces this behavior, lowering average excess Levenshtein distance from 0.195 to 0.131, reducing added cognitive complexity by 26.6%, and increasing Pass@1 by 2.3 points. However, these gains do not simply follow from a larger reasoning budget or larger models. We next ask whether minimal editing can be learned directly during post-training. We observe that supervised fine-tuning overfits to seen corruption patterns, whereas reinforcement learning gives the best out-of-domain edit-fidelity and performance-retention trade-off. These results position edit fidelity as a distinct axis of code-repair quality and show that it can be measured and learned.

Haechan KimYoonho LeeGisang Lee

The WHALE framework optimizes AI agent performance by alternating between updating model weights and searching for improved executable harness code. This approach consistently achieved higher accuracy, improving test mean@8 by 7.67–24.38 percentage points over single-component baselines and 4.15–13.00 points over prompt-restricted joint adaptation across diverse tool-using domains.

CivBench, an open-source benchmark developed by researchers from the University of Oxford, Google DeepMind, and others, introduces a framework for evaluating large language model agents in Civilization VI, leveraging a Model Context Protocol (MCP) and a unique narration layer. The platform identified consistent agent behaviors like under-monitoring of strategic game state and a notable gap between stated plans and executed actions, demonstrating challenges for current LLM agents in long-horizon, tool-mediated environments.

Shaowen WangGe ZhangKairong Luo

Researchers from Tsinghua University and ByteDance developed SMELT, a Looped Transformer architecture that consistently achieves lower validation loss and improved downstream performance compared to unlooped baselines when matched for compute. This approach provides 6.8% to 18.0% compute savings during training, with mechanistic analyses revealing refined residual updates and a reduced attention sink in repeated layers.

We propose Puffin-World, a unified multimodal architecture that integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules. To reliably construct and interact with 3D worlds, our framework jointly models three native world states: physics (gravity field and latitude), geometry (depth), and appearance (image), together with a unified Omni-Camera representation that supports diverse tasks and flexible motions. Beyond modeling these states, we introduce a strategy for propagating physical dynamics across future frames. By grounding absolute camera properties in the real world, Puffin-World enables physically consistent and visually stable world generation. We further couple appearance and geometry within a single generative process, jointly synthesizing each future view and reconstructing its underlying geometry. This unified paradigm enables interleaved closed-loop applications requiring synergy across multiple tasks, including mimic and self-calibrated world exploration. To scale Puffin-World to complex scenarios, we construct Puffin-16M, comprising 15 million vision-language-camera triplets and 1 million trajectories featuring various and challenging motions. To foster further research in this area, we released the code, models, and datasets.

As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. However, environments are what agent post-training actually requires: each can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration. Rather than generating environments from scratch, we observe that the tool-execution history in existing trajectories exposes the structure and contents of the environments in which they ran, making it possible to reconstruct those environments from the trajectories themselves. Thus, we introduce Terminal-Universe, a framework which turns each trajectory into a reusable environment and explores it for synthesizing new tasks and continued interactions. Specifically, Terminal-Universe replays the file operations recorded in a trajectory to restore each file before the agent modified it, yielding a partial workspace; a completion agent then supplies the missing files and dependencies. On this recovered workspace, we both reconstruct the original intent task and synthesize entirely new ones. Besides, we also scale the tasks along two complementary axes: breadth and depth. For breadth, we mine directional dependency relations between related environments and synthesize cross-workspace queries spanning multiple codebases, as developers routinely do in real-world development. For depth, we extend the initial single-turn query into a multi-round session that captures iterative user feedback and requirement refinement via a user agent. Applied to public terminal agent trajectories, Terminal-Universe produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points.

Subham Sekhar SahooLingjie ChenKhiem Pham

Uno, a diffusion-augmented LLM framework from the Institute of Foundation Models and collaborators, unifies autoregressive and discrete diffusion mechanisms to enable parallel token generation. This approach provably preserves the underlying AR model's output distribution and quality, achieving up to a 2.5x speedup for single-request throughput and 1.6x for system throughput over base AR models without needing a separate draft model.

Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geometry, and temporal dynamics remain coherent. Existing rewards assess these requirements separately: geometry-based rewards estimate trajectory execution but cannot judge the visual quality of the executed motion, whereas image-based rewards measure frame quality without capturing action execution or temporal dynamics. We posit that a vision-language model (VLM) offers a shared reasoning space for relating actions to their visual outcomes. However, judging a complete long video against its full action sequence creates a lengthy, noisy context in which short-lived local action evidence can be missed or diluted. We present WorldReward, a VLM-based pairwise preference reward model that unifies action-consistency and visual-quality evaluation for camera-conditioned world models. WorldReward decomposes paired videos into action-aligned chunks, organizes each chunk into structured visual evidence, and aggregates chunk-level decisions by voting into separate video-level action and visual-quality preferences. To train it, we construct a large-scale reasoning-augmented preference dataset using structured judgments generated by a frontier VLM and refined through tool-based agent auditing and targeted human review. We further introduce WorldReward-Bench, a human-annotated benchmark measuring reward-model agreement with human preferences across action consistency, appearance quality, and motion quality. WorldReward achieves the highest agreement on all three dimensions, exceeding GPT-5.5 by 3.42, 1.45, and 3.56 percentage points, respectively. When used for RL post-training of HY-WorldPlay 1.5, it consistently improves both action execution and visual quality across short- to long-term horizons.

The 'Last Translation Benchmark' (LTB) introduces a live, crowdsourced dataset and a novel rule-based evaluation method to identify specific failure modes in advanced machine translation models. It demonstrates that state-of-the-art systems achieve low verifier pass rates (below 44%) on challenging examples, while human translations maintain over 99% accuracy.

Whether the constant $$G=\sum_{k=0}^\infty\frac{(-1)^k}{(2k+1)^2}=\frac1{1^2}-\frac1{3^2}+\frac1{5^2}-\frac1{7^2}+\cdots$$ introduced by Catalan in the nineteen century is irrational, is a long-standing open problem. In this paper we prove the irrationality of $G$ via using suitable weights.

K2 Horizon is a fleet of six open language models, ranging from 0.9B to 375B parameters, with the 0.9B, 3.7B, and 7B models achieving state-of-the-art performance in their respective size classes. The release provides full training lifecycle transparency, encompassing intermediate checkpoints, training data or recipes, code, and logs for each model, which were pretrained on approximately 20 trillion tokens with 17% reasoning trajectories. A new Mixture-of-Value-Attention (MoVA) mechanism in the 36B-A4B model allows it to activate about 4 billion parameters per token while performing near the 32B dense model, and a lossless inference speedup method called Uno Diffusion is also introduced. An audit of the 375B-A23B model found a 3.37 percentage point accuracy correction for reward hacking on TerminalBench 2.1.

Youness Lamzouri

Youness Lamzouri at Université de Lorraine provides a new, conceptually simpler unconditional proof confirming that over 67.25% of the Riemann zeta function's non-trivial zeros are simple and lie on the critical line, and more than 83.62% are distinct. This work replaces a prior intricate proof's multi-component linear algebra framework with a single Hilbert space inequality, leveraging established unconditional pair-correlation results.

HarnessDev introduces a benchmark to evaluate the capacity of large language models (LLMs) to create and iteratively improve agent harnesses, the external software orchestrating their operations. The findings indicate that LLMs can construct runnable initial harnesses with varying quality, but struggle with robust generalization and consistent transferability of improvements across unseen tasks or different execution environments.