The Inference Report

Large language model (LLM) agents have achieved strong performance on a wide range of benchmarks, yet most evaluations assume static environments. In contrast, real-world deployment is inherently dynamic, requiring agents to continually align their knowledge, skills, and behavior with changing environments and updated task conditions. To address this gap, we introduce EvoArena, a benchmark suite that models environment changes as sequences of progressive updates across terminal, software, and social domains. We further propose EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories, enabling agents to reason about environmental evolution through changes in their memory. Experiments show that current agents struggle on EvoArena, achieving an average accuracy of 39.6% across evolving terminal, software, and social-preference domains. EvoMem consistently improves performance, yielding an average gain of 1.5% on EvoArena and also improving standard benchmarks such as GAIA and LoCoMo by 6.1% and 4.8%. Beyond individual tasks, EvoMem further improves chain-level accuracy by 3.7% on EvoArena, where success requires completing a consecutive sequence of related evolutionary subtasks. Mechanistic analysis shows that EvoMem improves evidence capture in the memory, indicating better preservation of complete evolving environment states. Our results highlight the importance of modeling evolution in both evaluation and memory for reliable agent deployment.

Retrieval-augmented generation (RAG) has become a standard mechanism for grounding language models in external knowledge, yet conventional retrieval based on lexical or semantic similarity is poorly suited for complex reasoning tasks: a semantically similar problem may demand an entirely different solution strategy, while a superficially different problem may share the same underlying reasoning pattern. We propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that teaches language models to reason by analogy. RA-RFT uses gold-relevance distillation to train a retriever that ranks contexts by expected reasoning benefit rather than semantic overlap, and then fine-tunes the policy model via reinforcement fine-tuning methods with retrieved analogous demonstrations, so the model learns to leverage reasoning traces under verifiable outcome rewards. We further analyze the diversity of retrieved contexts and find that reasoning-aware retrieval surfaces complementary solution strategies that provide distinct reasoning scaffolds for individual problems. Across challenging mathematical reasoning benchmarks, RA-RFT consistently outperforms standard reinforcement fine-tuning methods. For example, it improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively -- suggesting that reasoning-aware retrieval is a complementary axis of improvement and orthogonal to advances in reward design or training curricula.

Articulated tool manipulation remains a major challenge in dexterous robotics due to the need to coordinate internal degrees of freedom and contact-rich interactions. While prior work has largely focused on rigid objects, articulated tool use remains underexplored because of its physical complexity and the difficulty of learning functional grasping and manipulation policies. We present Mana (Manipulation Animator), a general sim-to-real framework that reinterprets dexterous manipulation as an animation problem. Inspired by computer animation, Mana employs a coarse-to-fine pipeline that transforms procedurally-generated grasp keyframes into manipulation trajectories through motion planning and reinforcement learning. The data generation process is largely automatic, requiring only a few mouse clicks to specify functional affordances (<1 minute per tool). Across four articulated tools spanning different scales and joint types, Mana achieves zero-shot sim-to-real transfer for both grasping and in-hand manipulation, demonstrating a scalable approach to dexterous articulated tool use.

Spatial reasoning, the ability to determine where objects are, how they relate, and how they move in 3D, remains a fundamental challenge for vision-language models (VLMs). Tool-augmented agents attempt to address this by augmenting VLMs with specialist perception modules, yet their effectiveness is bounded by the action interface through which those tools are invoked. In this work, we study how the design of this interface shapes the agent's capacity for open-ended spatial reasoning. Existing spatial agents either employ single-pass code execution, which commits to a full analysis strategy before any intermediate result is observed, or rely on a structured tool-call interface that often offers less flexibility for freely composing operations or tailoring the analysis to each task. Both designs offer limited flexibility for open-ended, complex 3D/4D spatial reasoning. We therefore propose SpatialClaw, a training-free framework for spatial reasoning that adopts code as the action interface. SpatialClaw maintains a stateful Python kernel pre-loaded with input frames and a suite of perception and geometry primitives, letting a VLM-backed agent write one executable cell per step conditioned on all prior outputs, enabling the agent to flexibly compose and manipulate perception results and adapt its analysis to both intermediate text and visual observations and the demands of each problem. Evaluated across 20 spatial reasoning benchmarks spanning a broad range of static and dynamic 3D/4D spatial reasoning tasks, SpatialClaw achieves 59.9% average accuracy, outperforming the recent spatial agent by +11.2 points, with consistent gains across six VLM backbones from two model families without any benchmark- or model-specific adaptation.

Positional encodings (PEs) enhance the power of graph neural networks (GNNs), both theoretically and empirically. Two of the most popular families of PEs - spectral (e.g., Laplacian eigenspaces, effective resistance) and walk-based (polynomials of the adjacency matrix) - are theoretically equivalent in expressive power, with expressivity between the 1-WL and 3-WL tests. However, this equivalence assumes the GNN uses the "complete" version of these PEs, which requires $O(n^3)$ time and space complexity. Instead, practitioners commonly use truncated variants of these encodings, such as the first $k$ eigenspaces or powers of the adjacency matrix. However, the theoretical properties of these truncated PEs are unknown. In this work, we initiate the study of these truncated PEs. Theoretically, we show that, under truncation, several families of PEs are fundamentally different in expressive power. As a corollary, we show that truncated spectral PEs are no longer stronger than the 1-WL test. We also study a family of spectral PEs, the $k$-harmonic distances, to highlight the differences in expressive power of even closely related truncated PEs. Finally, we experimentally show that a mix of truncated PEs is preferable to any single family on real-world datasets.

Reproducibility in the social and behavioral sciences is typically evaluated by independent researchers who reanalyze the original data to assess whether the published findings can be recovered. However, such approaches are resource-intensive and difficult to scale. Here, we show that large language models (LLMs) can automate reproducibility assessments. Using N=76 published studies with predefined claims from the behavioral and social sciences, we compare LLM-generated analysis with the original findings and human reanalysis. For 7 studies, the LLM could not produce a viable effect size estimate. For the remaining studies, our LLM pipeline recovered the original effect sizes in 41% of studies using a +/-0.05 tolerance in Cohen's d. Further, our LLM pipeline reached the same qualitative conclusion as the original study in 96% of cases, where conclusions indicate whether the reanalysis supports the original claim. For comparison, human reanalysts recovered the original effect sizes in 34% of studies and reached the same qualitative conclusion in 74% of cases. Together, these results show that LLMs can serve as a scalable tool for automated reproducibility assessment and provide a foundation for systematic auditing of empirical results in the social and behavioral sciences.