Test-time scaling has become a powerful way to improve large language models. However, existing methods are best suited to short, bounded outputs that can be directly compared, ranked or refined. Long-horizon coding agents violate this premise: each attempt produces an extended trajectory of actions, observations, errors, and partial progress taken by the agent. In this setting, the main challenge is no longer generating more attempts, but representing prior experience in a form that can be effectively selected from and reused. We propose a test-time scaling framework for agentic coding based on compact representations of rollout trajectories. Our framework converts each rollout into a structured summary that preserves its salient hypotheses, progress, and failure modes while discarding low-signal trace details. This representation enables two complementary forms of inference-time scaling. For parallel scaling, we introduce Recursive Tournament Voting (RTV), which recursively narrows a population of rollout summaries through small-group comparisons. For sequential scaling, we adapt Parallel-Distill-Refine (PDR) to the agentic setting by conditioning new rollouts on summaries distilled from prior attempts. Our method consistently improves the performance of frontier coding agents across SWE-Bench Verified and Terminal-Bench v2.0. For example, by using our method Claude-4.5-Opus improves from 70.9

Existing web agent benchmarks have largely converged on short, single-site tasks that frontier models are approaching saturation on. However, real world web use consists of long-horizon, multi-site workflows. Common web navigation tasks, such as comparing products across different domains, planning trips across multiple services, or summarizing information from multiple search queries, require sustained context and cross-site reasoning over potentially hours of browsing. To capture and evaluate such behaviors, we introduce Odysseys: a benchmark of 200 long-horizon web tasks derived from real world browsing sessions evaluated on the live Internet. We find that binary pass/fail evaluation is inadequate for long-horizon settings and introduce a rubric-based evaluation, annotating each Odysseys task with an average of 6.1 graded rubrics. We demonstrate that this yields higher agreement with humans and provides a more fine-grained signal than commonly used trajectory-level LLM-as-a-judge evaluation metrics. We tested several leading frontier models and find that the strongest models achieve a success rate of 44.5

A useful phone agent needs to be personally intelligent. It should reason over a user's identity, history, and preferences as they exist on the device, not just follow isolated instructions in an impersonal sandbox. Existing mobile agent benchmarks lack this kind of personalization. We introduce iOSWorld, the first interactive native iOS simulator benchmark built around a persistent user identity spanning 26 newly built iOS apps. These apps contain connected data such as transactions, messages, travel records, social relationships, and financial activity. iOSWorld includes 133 tasks across three increasingly difficult categories. Single-app tasks (27) test one app, multi-app tasks (60) span 2 to 8 apps, and memory and personalization tasks (46) require agents to infer patterns from personal data. We evaluate frontier and open-source computer-use models in both vision-only and privileged vision+XML settings. The best configuration reaches 52% overall but only 37% on multi-app tasks. Privileged vision+XML access improves frontier models by up to 26 percentage points, while smaller models do not benefit from added accessibility-tree input. We release iOSWorld as an open-source benchmark with all apps, seeded data, tasks, rubrics, and evaluation code.

Large vision-language models (VLMs) have advanced multimodal tasks such as video question answering (QA). However, VLMs face the challenge of selecting frames effectively and efficiently, as standard uniform sampling is expensive and performance may plateau. Inspired by active perception theory, which posits that models gain information by acquiring data that differs from their expectations, we introduce Video Active Perception (VAP), a training-free method to enhance long-form video QA using VLMs. Our approach treats keyframe selection as data acquisition in active perception and leverages a lightweight text-conditioned video generation model to represent prior world knowledge. Empirically, VAP achieves state-of-the-art zero-shot results on long-form or reasoning video QA datasets such as EgoSchema, NExT-QA, ActivityNet-QA, IntentQA, and CLEVRER, achieving an increase of up to 5.6 x frame efficiency by frames per question over standard GPT-4o, Gemini 1.5 Pro, and LLaVA-OV. Moreover, VAP shows stronger reasoning abilities than previous methods and effectively selects keyframes relevant to questions. These findings highlight the potential of leveraging active perception to improve the frame effectiveness and efficiency of long-form video QA.

Post-training pipelines that combine supervised fine-tuning (SFT) with reinforcement learning (RL) have emerged as the key recipe for transforming large language models (LLMs) into robust reasoners. We argue that this combined success is driven by compositional generalization, which we formalize through a hierarchical latent selection model. In this framework, reasoning traces are generated by a cascade of discrete latent selection variables corresponding to reusable atomic modules, including both skills (local operations) and routing mechanisms (how intermediate information is selected, reused, and composed). Within this model, we theoretically show that SFT and RL play asymmetric, complementary roles: SFT supplies the raw module materials in compositional traces, and RL decomposes those traces to identify the latent atomic modules and enable compositional generalization. We design controlled experiments to validate this theory. Our results demonstrate that RL can extract atomic modules from compound traces supplied by SFT and recombine them to solve new configurations. Moreover, we find that training on compound traces yields stronger generalization than training on isolated atomic modules. Finally, we investigate the relationship between SFT and RL data and identify an effective protocol in which SFT ensures coverage of all atomic modules through compositional traces, while RL focuses on novel compositions outside the SFT support to drive exploration.

Reinforcement learning (RL) has improved the reasoning abilities of large language models (LLMs), yet state-of-the-art methods still fail to learn on many training problems. On hard problems, on-policy RL rarely explores even a single correct rollout, yielding zero reward and no learning signal for driving improvement. We find that natural solutions to remedy this exploration problem from classical RL, such as entropy bonuses, more permissive clipping of the importance ratio, or direct optimization of pass@k objectives, do not resolve this issue and often destabilize optimization without improving solvability. A natural alternative is to leverage transfer from easier problems. However, we show that mixing easy and hard problems during RL training is counterproductive due to ray interference, where optimization focuses on already-solvable problems in a way that actively inhibits progress on harder ones. To address this challenge, we introduce Privileged On-Policy Exploration (POPE), an approach that leverages human- or other oracle solutions as privileged information to guide exploration on hard problems, unlike methods that use oracle solutions as training targets (e.g., off-policy RL methods or warmstarting from SFT). POPE augments hard problems with prefixes of oracle solutions, enabling RL to obtain non-zero rewards during guided rollouts. Crucially, the resulting behaviors transfer back to the original, unguided problems through a synergy between instruction-following and reasoning. Empirically, POPE expands the set of solvable problems and substantially improves performance on challenging reasoning benchmarks.

Current benchmarks for computer-use agents evaluate models in impersonal environments. This leaves a gap between evaluation and deployment where personal assistants are expected to work across a user's whole digital life, including their context, historical data, and logged-in accounts. This gap is widest on web tasks, where live web evaluations cannot exercise sites that require logging in or personal information, the kind of site a real personal assistant has to drive. We introduce MyPCBench, which tests computer-use agents as personal assistants on a Linux desktop populated with 17 simulated real-world web applications and a full desktop stack, all seeded for one canonical persona, Michael Scott from The Office. We define 184 tasks in this environment, each inspired by a real request drawn from the OpenClaw community, and benchmark six closed and open-weight models with a uniform computer+bash tool surface. We find that the best model, Claude Opus 4.6, fully solves 55.4\% of the tasks, the only model above 50\%. Model failures cluster on tasks that span many applications and on long trajectories, where personalization stresses an assistant the most. We release the environment, task set, and agent harness at https://mypcbench.com.

Vision-Language Models (VLMs) typically rely on static initial frames for video reasoning, restricting their ability to incorporate essential dynamic information as the reasoning process evolves. Existing methods that augment Chain-of-Thought (CoT) with additional frame information often exhibit suboptimal CoT quality and lack the crucial ability to synthesize visual information for hypothetical or counterfactual scenarios. We introduce Act-to-See (Act2See), a novel framework that enables active visual perception by empowering VLMs to actively interleave video frames within text CoTs. Act2See is developed via Supervised Fine-Tuning (SFT) on a high-quality dataset of reasoning traces generated by a frontier VLM. These traces integrate active calls to either retrieve existing frames or generate new ones, and are rigorously verified against human-annotated CoTs to ensure quality. This approach cultivates an emergent capability: at inference time, the model actively determines when to search for or synthesize the necessary visual evidence. Act2See establishes new state-of-the-art results on challenging benchmarks, including VideoEspresso and ViTIB, and outperforms comparable or larger models on Video-MME, EgoNormia, and VCR-Bench, demonstrating an advancement in enabling VLMs with active visual perception for video reasoning.

We present RefVFX, a new framework that transfers complex temporal effects from a reference video onto a target video or image in a feed-forward manner. While existing methods excel at prompt-based or keyframe-conditioned editing, they struggle with dynamic temporal effects such as dynamic lighting changes or character transformations, which are difficult to describe via text or static conditions. Transferring a video effect is challenging, as the model must integrate the new temporal dynamics with the input video's existing motion and appearance.

Computer use agents (CUAs) today are primarily deployed as single serial agents. This setup is suboptimal for complex long-horizon tasks that benefit from task decomposition, parallel execution, and consistent re-planning based on new information. In this paper, we argue that we should instead move towards evaluating and building multi-agent computer use (MACU) systems. These systems, which emphasize planning and parallel execution, alleviate many of the shortcomings of single-agent CUAs. We propose a general multi-agent setup in which a manager model decomposes computer use tasks as a directed acyclic graph (DAG), encoding relevant dependencies and goals for subagents. At each iteration, the manager dispatches parallel CUA subagents to carry out nodes on the ready frontier of the DAG, and continuously revises the DAG (adding, canceling, or rewriting nodes) as new findings arrive from subagents. This design treats the partially observable environment of computer use as a first class challenge: information that downstream agents may not be able to re-observe are retained and passed forward through the manager and DAG structure. We demonstrate that MACU consistently improves over strong single-agent baselines by 3.4-25.5% on desktop (OSWorld) and web navigation (Online-Mind2Web, WebTailBench, Odysseys) benchmarks, exhibits more favorable test-time scaling, and solves complex long-horizon tasks where single-agent CUAs get stuck. On Odysseys, a long-horizon web navigation benchmark, MACU improves average task completion wall-clock time by ∼ 1.5 ×, demonstrating its efficacy in speeding up traditionally slow CUA pipelines. Our findings highlight that multi-agent coordination is a promising axis for scaling computer use agents to work productively for longer and more effectively. We release all code and interactive visualizations at https://jykoh.com/multi-agent-computer-use.

Reinforcement learning typically optimizes average reward. For generative policies, the average can hide an important distinction: two policies can achieve the same mean reward while having very different chances of producing a rare but high-reward rollout. This matters as sampling increases during training and inference, since its benefit depends on retaining probability mass on high-reward outcomes. We propose to optimize this coverage directly. Rather than considering only expected reward, we consider all of its upper tails: for each reward threshold, how likely is the policy to exceed it? This turns a continuous reward into a family of binary success events. We introduce Tail-Likelihood Reinforcement Learning (TailRL), which maximizes the log-probability of exceeding a randomly chosen reward threshold. Its gradient gives more weight to rare, high-reward rollouts and can be interpreted as a mixture of Best-of-(k) gradients. TailRL requires only a simple modification to the advantage function, making it compatible with existing reinforcement learning pipelines. Across object localization, maze navigation, GUI grounding, and code optimization, TailRL leverages rare high-reward training samples to avoid suboptimal solutions and yields models that benefit more from additional samples at inference time.

Contact-rich manipulation requires force sensitivity, but many robot arms lack dedicated force sensors due to their high cost. We present Neural External Torque Estimation (NEXT), a data-driven method that estimates external joint torques without needing any dedicated force sensors. NEXT trains in 1 minute from only 10 minutes of free-motion data, yet achieves estimates comparable to dedicated joint-torque sensors. NEXT enables force-feedback teleoperation on low-cost arms and improves policy learning through Force-Informed Re-Sampling Training (FIRST), which up-samples pre-contact and contact segments during behavior cloning. Across five long-horizon tasks, FIRST outperforms prior force-aware policies by over 17

Efficient exploration is essential for intelligent systems interacting with their environment, but existing language models often fall short in scenarios that require strategic information gathering. In this paper, we present Paprika, a fine-tuning approach that enables language models to develop general decision-making capabilities that are not confined to particular environments. By training on synthetic interaction data from different tasks that require diverse strategies, Paprika teaches models to explore and adapt their behavior on a new task based on environment feedback in-context without more gradient updates. Experimental results show that models fine-tuned with Paprika can effectively transfer their learned decision-making capabilities to entirely unseen tasks without additional training. Unlike traditional training, our approach's primary bottleneck lies in sampling useful interaction data instead of model updates. To improve sample efficiency, we propose a curriculum learning strategy that prioritizes sampling trajectories from tasks with high learning potential. These results suggest a promising path towards AI systems that can autonomously solve novel sequential decision-making problems that require interactions with the external world.

Human intelligence involves metacognitive abilities like self-regulation, recognizing limitations, and seeking assistance only when needed. While LLM Agents excel in many domains, they often lack this awareness. Overconfident agents risk catastrophic failures, while those that seek help excessively hinder efficiency. A key challenge is enabling agents with a limited intervention budget C is to decide when to request assistance. In this paper, we propose an offline framework that trains a "helper" policy to request interventions, such as more powerful models or test-time compute, by combining LLM-based process reward models (PRMs) with tabular reinforcement learning. Using state transitions collected offline, we score optimal intervention timing with PRMs and train the helper model on these labeled trajectories. This offline approach significantly reduces costly intervention calls during training. Furthermore, the integration of PRMs with tabular RL enhances robustness to off-policy data while avoiding the inefficiencies of deep RL. We empirically find that our method delivers optimal helper behavior.

Many contact-rich tasks humans perform, such as box pickup or rolling dough, rely on force feedback for reliable execution. However, this force information, which is readily available in most robot arms, is not commonly used in teleoperation and policy learning. Consequently, robot behavior is often limited to quasi-static kinematic tasks that do not require intricate force-feedback. In this paper, we first present a low-cost, intuitive, bilateral teleoperation setup that relays external forces of the follower arm back to the teacher arm, facilitating data collection for complex, contact-rich tasks. We then introduce FACTR, a policy learning method that employs a curriculum which corrupts the visual input with decreasing intensity throughout training. The curriculum prevents our transformer-based policy from over-fitting to the visual input and guides the policy to properly attend to the force modality. We demonstrate that by fully utilizing the force information, our method significantly improves generalization to unseen objects by 43% compared to baseline approaches without a curriculum. Video results and instructions at https://jasonjzliu.com/factr/

Learning from self-sampled data and sparse environmental feedback remains a fundamental challenge in training self-evolving agents. Temporal credit assignment mitigates this issue by transforming sparse feedback into dense supervision signals. However, previous approaches typically depend on domain-specific value functions for credit assignment, which suffer from poor sample efficiency and limited generalization. In this work, we propose to leverage pre-trained knowledge from large language models (LLMs) to transform sparse rewards into dense training signals (i.e., the advantage function) through retrospective in-context learning (RICL). We further propose an online learning framework, RICOL, which iteratively refines the policy based on the credit assignment results from RICL. We empirically demonstrate that RICL can accurately estimate the advantage function with limited samples and effectively identify critical states for temporal credit assignment. Extended evaluation on the BabyAI benchmark shows that RICOL significantly improves sample efficiency compared to traditional online RL algorithms while achieving performance comparable to imitation learning from expert demonstartions. Our findings highlight the potential of leveraging LLMs for temporal credit assignment, paving the way for more sample-efficient and generalizable RL paradigms.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

This note responds to "Promises and Pitfalls of Deep Kernel Learning" (Ober et al., 2021). The marginal likelihood of a Gaussian process can be compartmentalized into a data fit term and a complexity penalty. Ober et al. (2021) shows that if a kernel can be multiplied by a signal variance coefficient, then reparametrizing and substituting in the maximized value of this parameter sets a reparametrized data fit term to a fixed value. They use this finding to argue that the complexity penalty, a log determinant of the kernel matrix, then dominates in determining the other values of kernel hyperparameters, which can lead to data overcorrelation. By contrast, we show that the reparametrization in fact introduces another data-fit term which influences all other kernel hyperparameters. Thus, a balance between data fit and complexity still plays a significant role in determining kernel hyperparameters.

Sim2real for robotic manipulation is difficult due to the challenges of simulating complex contacts and generating realistic task distributions. To tackle the latter problem, we introduce ManipGen, which leverages a new class of policies for sim2real transfer: local policies. Locality enables a variety of appealing properties including invariances to absolute robot and object pose, skill ordering, and global scene configuration. We combine these policies with foundation models for vision, language and motion planning and demonstrate SOTA zero-shot performance of our method to Robosuite benchmark tasks in simulation (97 %). We transfer our local policies from simulation to reality and observe they can solve unseen long-horizon manipulation tasks with up to 8 stages with significant pose, object and scene configuration variation. ManipGen outperforms SOTA approaches such as SayCan, Open VLA, LLMTrajGen and VoxPoser across 50 real-world manipulation tasks by 36%, 76%, 62% and 60% respectively. Video results at mihdalal.github.io/manipgen

Many real-world decision-making problems are combinatorial in nature, where states (e.g., surrounding traffic of a self-driving car) can be seen as a combination of basic elements (e.g., pedestrians, trees, and other cars). Due to combinatorial complexity, observing all combinations of basic elements in the training set is infeasible, which leads to an essential yet understudied problem of zero-shot generalization to states that are unseen combinations of previously seen elements. In this work, we first formalize this problem and then demonstrate how existing value-based reinforcement learning (RL) algorithms struggle due to unreliable value predictions in unseen states. We argue that this problem cannot be addressed with exploration alone, but requires more expressive and generalizable models. We demonstrate that behavior cloning with a conditioned diffusion model trained on successful trajectory generalizes better to states formed by new combinations of seen elements than traditional RL methods. Through experiments in maze, driving, and multiagent environments, we show that conditioned diffusion models outperform traditional RL techniques and highlight the broad applicability of our problem formulation.