Long-Horizon Tasks

Hierarchical vision-language-action (Hi-VLA) systems have emerged as a promising paradigm for complex robot manipulation, by using high-level VLM planners to decompose tasks into language subgoals executed by low-level VLA controllers. Despite recent empirical progress, there is a lack of unified design principles for these systems: existing Hi-VLA systems differ in how they choose and connect planners, controllers, mechanisms to switch between the two, and how observations and memory are represented in the planner. In this paper, we present a systematic study of Hi-VLA design for robot manipulation. We unify representative Hi-VLA agents under an options-style control framework and benchmark core design choices across short-horizon, long-horizon, and reasoning-intensive tasks. Our analysis distills practical principles for building Hi-VLA systems, showing how model choices and interface mechanisms jointly shape performance. Applying these principles yields a substantially stronger system than either flat VLA control or a naively designed hierarchy, across experiments both in simulation and on a real ALOHA robot. Overall, our results provide a foundation for building more capable, robust, and principled hierarchical VLA agents. More information and video at jiahenghu.github.io/hi-vla.

Large language models deployed as autonomous agents for enterprise workflows face a key challenge: verbose tool responses from enterprise systems can cause context overflow, stale-state errors, and high inference cost. We study this problem in automated expense itemization in Microsoft Dynamics 365 Finance and Operations using Model Context Protocol tools. We evaluate four GPT-5 configurations on a 50-task hotel expense benchmark: no user model, full conversation history, context pruned to the last 5 tool call/response pairs, and pruning with automated summarization. Results are averaged across 5 independent runs, with the user model held constant for the context-engineering comparison. The no-user-model baseline achieves only 8.0% complete itemization. Full-context retention improves completion to 71.0%, but consumes 1,480,996 tokens and 14.56 hours per benchmark. Pruning to the last 5 tool calls improves completion to 79.0% while reducing token use to 535,274 and runtime to 5.39 hours. Adding summarization achieves the best result: 91.6% complete itemization and 99.64% average amount itemized, with 553,374 tokens and 5.79 hours. We further report confidence intervals, effect-size analysis, sensitivity over pruning and summary windows, failure analysis, results across five expense types grouped into three categories, and cross-model evidence with Claude Sonnet 4.5. These results show that, for this class of enterprise tool-use workflow, selective retention of recent tool interactions plus compact summarization can improve both reliability and efficiency compared with full-history retention.

Temporal modeling is essential for robotic manipulation, as effective control requires both memory of past interactions and imagination of future states. However, most VLA models rely primarily on the current observation and therefore struggle with long-horizon, temporally dependent tasks. Cognitive science suggests that humans rely on working memory to buffer short-lived context, the hippocampal system to preserve episodic memory of past experience, and internal models to imagine possible future state evolution. Inspired by these mechanisms, we propose MemoryVLA++, a full temporal modeling framework that equips VLA models with memory and imagination for robotic manipulation. A pretrained VLM encodes the current observation into perceptual and cognitive tokens, forming working memory. These tokens query a Perceptual-Cognitive Memory Bank to retrieve relevant historical context. This bank stores low-level details and high-level semantics from past interactions, and is updated through redundancy-aware consolidation. A world model imagines future states in a denoising latent space, and the imagined latents are integrated under memory guidance to form full temporal-aware tokens. The resulting tokens condition a diffusion action expert to predict temporally consistent action sequences. We conduct extensive experiments on 5 simulation benchmarks and 3 categories of real-robot tasks across 3 robots, covering general manipulation, long-horizon temporal tasks, robustness, and generalization. Our method achieves strong performance across Libero, SimplerEnv, Mikasa-Robo, Calvin, Libero-Plus, and diverse real-robot tasks, validating the effectiveness of full temporal modeling with memory and imagination. For example, on real robots, it achieves +9%, +26%, +28% gains on general, memory-dependent, and imagination-dependent tasks. Project Page: https://shihao1895.github.io/MemoryVLA-PP-Web

Large language models are increasingly expected to handle complex, long-horizon real-world tasks whose context demands can grow without bound, yet model context windows remain inherently finite. Recent work explores a paradigm where a main agent decomposes tasks and dispatches subtasks to subagents, which execute and return only summarized results, conserving the main agent's context budget. However, performing this well requires delegation intelligence: the ability to decompose complex tasks, determine when and what to delegate, and integrate returned results into the ongoing workflow. Training data for this capability is scarce in naturally occurring text, and to our knowledge, how to synthesize such data and train models to acquire this capability remains largely unexplored in the open-source community. To bridge this gap, we present a preliminary exploration targeting deep research, a representative long-horizon agent task. Specifically, we design a harness that guides the model toward high-quality task decomposition and delegation, while constraining subagents to return results properly to support the main agent's workflow. The harness-guided trajectories naturally encode correct delegation decisions, which we use as supervised fine-tuning data to internalize delegation intelligence into model weights. Our resulting model, SearchSwarm-30B-A3B, achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, the best results among all models of comparable scale. We will release our harness, model weights, and training data to facilitate future research.

Joint Embedding Predictive Architectures (JEPAs) have shown promising world modeling capabilities, enabling planning in latent space by optimizing action trajectories using methods like the Cross-Entropy Method (CEM). These methods are, however, too computationally expensive and ineffective for long-horizon planning. Furthermore, these methods typically require an explicit image of the goal state, which is not always possible in real-world tasks. In this work, we tackle these limitations by proposing Forward-Forward-JEPA (FF-JEPA), a hierarchical approach leveraging two forward dynamics models. Alongside a standard action-conditioned forward model, we introduce an action-free latent planner that predicts the next subgoal given the current state. This approach removes the need for goal images and enables long-horizon planning by decomposing complex trajectories into a sequence of tractable, short-term optimization problems. Preliminary results on PushT demonstrate that FF-JEPA successfully overcomes flat world models' long-horizon collapse, highlighting this approach as a promising direction for goal-free planning.

Optimizing large language models (LLMs) for long-horizon caregiver agents requires balancing delayed task objectives with immediate environment dynamics, such as patient distress and resistance. In dementia care, this balance is especially difficult: trajectory level rewards are too sparse for turn level credit assignment, while external LLM-based evaluators are costly and can misread fragmented or indirect patient responses. To address this issue, we propose \textbf{T}urn-\textbf{T}rajectory \textbf{G}roup \textbf{R}elative \textbf{P}olicy \textbf{O}ptimization (\textbf{T$^{2}$-GRPO}), a framework that decouples caregiver RL into two normalized reward horizons and enforces safety through a binary hard veto. $T^2$-GRPO derives dense turn-level rewards directly from environment state transitions, measuring changes in patient distress and resistance from a frozen dementia patient simulator. These environment-grounded rewards are combined with trajectory-level evaluations through independent centered-rank normalization, which preserves heterogeneous reward signals and mitigates reward collapse. Extensive experiments on dementia caregivers show that T $^{2}$-GRPO outperforms competitive baselines, indicating a substantial improvement for emotionally sensitive caregiver scenarios that effectively handles immediate patient feedback, long-term care outcomes, and safety constraints.

Vision-Language Models (VLMs) are increasingly required to process unbounded video streams in applications such as video-call assistants, live commentary, and embodied robots. An ideal streaming system should support proactive interaction, long-horizon memory, and real-time processing, while resting on a VLM backbone capable of handling diverse in-the-wild streaming tasks. However, existing VLMs excel at offline video understanding but fall short in streaming capabilities and lack dedicated infrastructure for streaming deployment. We address this gap on three fronts. (i) For backbone capability, we construct \textbf{Streaming-Train-248K}, a streaming dataset paired with a novel training objective for adapting VLMs to streaming interaction and understanding. (ii) For real-world deployment, we introduce \textbf{Streaming Harness}, a plug-and-play system that endows any VLM with three core abilities: proactive interaction (per-second response decisions), long-term memory (12-hour context retention), and real-time processing (sub-second latency). (iii) To drive continued community progress on streaming capabilities, we design \textbf{Streaming-Eval}, a benchmark that reflects models' capabilities across diverse in-the-wild scenarios. Extensive experiments demonstrate consistent gains from our approach across all core capabilities required for streaming video understanding. We will open-source our data, code, and benchmark to advance the community's shift from offline video understanding to deployable streaming intelligence.

Most evaluations of LLM agents look like exams: a discrete task, a clean environment, a score in minutes or hours. We argue that this approach is mismatched with the deployment conditions of autonomous systems, where the relevant timescale can be weeks to months, and where the dynamics that matter most, such as behavioral drift, governance in diverse environmental contexts, and cross-influence between agents from different model families, only emerge over time. We introduce Emergence World, a continuously running multi-agent simulation platform designed to make those dynamics measurable. The platform hosts populations of LLM-driven agents in a shared spatial world grounded in live external data (e.g. real-time weather, news APIs, internet access), equips each agent with 120+ specialized tools and three persistent memory systems, and lets them govern themselves through democratic mechanisms with consequential outcomes. The platform is model-agnostic at the reasoning layer and supports heterogeneous populations in which agents from different vendors share the same world. To illustrate the kinds of questions the platform makes tractable, we present a 15-day cross-vendor study with five parallel worlds powered by Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed population. Identical roles and starting conditions produced radically different outcomes, ranging from stable deliberative governance to total population collapse. We release the prompts, log data and configurations to support further research on long-horizon multi-agent autonomy.

Vision-language-action (VLA) models increasingly condition robot policies on history, depth, or 4D features to resolve ambiguity in long-horizon manipulation. However, more spatiotemporal evidence is not necessarily better: when the injected evidence is not motion-consistent, it can introduce geometric drift, fragmented temporal cues, and unstable action generation. This raises a simple question: should a VLA remember past frames, or remember the motion that connects them? We introduce MotionVLA, a motion-history interface that converts a short past-only video window into compact, time-continuous trajectory-field tokens. Instead of treating history as a sparse set of ndependently lifted frames, MotionVLA represents recent observations as physically coherent motion evidence. Current visual tokens query this history to retrieve task-relevant motion information, which is then recoupled into the VLA stream under trajectory-grounded supervision. Experiments across simulation benchmarks and preliminary real-robot rollouts show that MotionVLA improves long-horizon manipulation while producing smoother and more direct executions. These results suggest that effective VLA memory is not just about providing more 4D context, but about exposing motion-consistent evidence that is usable for control.

Open-vocabulary long-horizon manipulation requires robots to reason over flexible instructions and complex multi-object scenes while adaptively planning, executing, monitoring, and recovering from failures. We address these demands with a closed agent loop in which a VLM orchestrates heterogeneous robot capabilities as interruptible tools. Unlike in virtual AI agents, the timing of decisions, actions and tool calls is important in a physical world that does not pause for reasoning. We refer to this setting as Physical Orchestration, and propose VoLoAgent, a VLM that plans, monitors, and recovers by treating a VLA/WAM as an interruptible tool it steers mid-rollout alongside vision models and action primitives. To evaluate these long-horizon capabilities, we introduce RoboVoLo, a high-fidelity benchmark for open-vocabulary long-horizon manipulation across common sense, memory/state tracking, complex references, and world knowledge, with both task-level success and failure-mode diagnostics. Experiments show VoLoAgent substantially outperforms single VLA/VLM or tool-based systems, with validation on real-robot experiments. Project page: https://chicychen.github.io/VoLo/

State tracking exposes a sharp limitation of sequence models: the relevant signal is often not a summary of observed tokens, but an ordered latent state that evolves through non-commutative transformations. We introduce a held-out transition-pair falsifier for finite non-Abelian group tracking. The protocol forbids selected ordered generator pairs during training and requires the same local patterns during evaluation, blocking one direct local-transition memorization pathway. In a controlled $S_3 \times S_3$ benchmark, a projected recurrent state model trained only on length-8 sequences produces error-free final-state predictions (perfect 250/250 per horizon) through evaluation horizons up to 1,048,576 tokens across five seeds. Matched native-readout baselines, including bag, GRU, and a single-configuration structured state-space model, remain near floor under the same protocol. Projection-matched GRU, structured SSM, and bag baselines equipped with analogous finite-group prototype readouts also remain near chance under the same split. Mechanism diagnostics show that hard projection coincides with low homomorphism error, low state-consistency drift, and non-trivial commutator separation, while softened projection collapses final-state accuracy. Clean-split audits verify zero verbatim reduced-word overlap and zero structural-template overlap between training and evaluation partitions. The evidence is scoped to this controlled finite-group falsifier rather than to a general architecture ranking. Within that regime, explicit projected non-commutative state composition acts as a useful inductive bias for long-horizon hidden-state tracking.

Multi-operator control of robot teams requires not only access to the same mission information, but also mechanisms for maintaining shared awareness and preventing conflicting interventions. Building on our previous HORUS interface (Holistic Operational Reality for Unified Systems) we present a mixed-reality interface that extends single-operator multi-robot supervision to collaborative multi-operator use. The system supports two complementary modes: a co-located shared workspace, in which operators observe and manipulate the same mini-map in the same physical location, and a private-workspace mode, in which operators work on the same mission through independently placed local workspaces. The architecture combines registration-driven scene construction, lightweight shared-session synchronization, and per-robot control leases to support collaborative monitoring, tasking, and teleoperation while preventing conflicting commands. We evaluated the approach in a human-subject study with 36 participants (18 pairs) controlling three Nova Carter mobile robots in two search environments. The performance of the objective task was comparable across the two modes, indicating that both modes supported effective mission execution. However, the co-located shared workspace significantly improved perceived collaboration, shared understanding, and handoff clarity, and was the preferred collaborative mode. These results indicate that physically co-locating the MR workspace improves how operators coordinate even when the underlying robot-control tools remain unchanged.

3D visuomotor policies offer a promising direction for complex robotic manipulation, as depth maps and point clouds provide rich geometric information for spatial reasoning. However, their success often depends on large-scale real-world demonstrations, which are costly and time-consuming to collect. To this end, existing methods commonly use demonstration generation strategies to improve data efficiency by applying object-centric transformations to human-collected demonstrations, such as varying object poses or scales. While effective for local variation, these transformations largely preserve the original scene structure and skill sequence, limiting their ability to synthesize diverse scene-skill-object combinations for complex tasks. In this paper, we propose Task-Edit, a novel demonstration generation framework that generates diverse trajectories from a task-centric editing perspective. The key insight of Task-Edit is to decompose a task into scene, skill and object components, and flexibly recombine them. In this way, Task-Edit enables scalable demonstration generation and significantly improves generalization for long-horizon manipulation tasks. We evaluate Task-Edit through extensive real-world experiments and demonstrate three advantages: (1) Effectiveness: Task-Edit significantly improves 3D visuomotor policies across various real-world tasks and robot embodiments. (2) Generalizability: Task-Edit improves model generalization across different scenario setups. (3) Applicability: Task-Edit enables models to handle scenarios that are difficult to collect in the real world, including disturbance resistance, obstacle avoidance and unseen cluttered scenes.

Task planning often suffers from severe efficiency bottlenecks when robots must reason over long-horizon action sequences under complex logical constraints, including object affordances, spatial relationships, and sequential action dependencies. Recent neuro-symbolic methods improve planning efficiency by learning object-importance scores to prune task-irrelevant objects, but they typically rely on fixed offline supervision generated from full search spaces. This creates a train-test mismatch: at deployment, the planner operates in pruned search spaces induced by the model's own imperfect predictions, leading to exposure bias and degraded planning performance. To address this challenge, we formulate object-importance learning for task planning as an imperative learning-based bilevel optimization problem. The upper level optimizes a neural scorer, while the lower level solves a symbolic planning problem in the score-pruned search space. To stabilize this learning process, we introduce a 3R strategy into the lower-level planning, using parallel Repair, Restart, and Rollback recovery to provide reliable and adaptive feedback for upper-level learning. Experiments on three challenging benchmarks demonstrate state-of-the-art performance, including an 80.04% reduction in failure rate and a 57.14% reduction in planning time. We further validate the framework on a quadruped-based mobile manipulator in simulation and the real world, demonstrating its potential for efficient and deployable neuro-symbolic task planning.

AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments. Yet current agent benchmarks largely evaluate short-form tasks, such as single pull requests, small tickets, or 5-10 minute exercises, limiting our ability to measure agents' capabilities in planning, long-context understanding, and memory use. We introduce SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains. Each task consists of a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Logged agent attempts average 27.2M total tokens, making SWE-Marathon substantially longer-horizon than existing SWE and command-line agent benchmarks. Current frontier coding agents solve fewer than 30% of tasks. Failures often arise from poor self-verification, self-reported infeasibility, and premature termination. We also observe reward-hacking behavior in 13.8% of rollouts, where agents attempt to exploit the environment or verifier to bypass the intended workflow. SWE-Marathon includes adversarial review of test suites and execution environments, as well as multi-layer checks designed to prevent shortcut solutions. We release SWE-Marathon, evaluation code, and agent trajectories at https://swe-marathon.org/.

Web agents operating over long horizons ingest raw DOM and accessibility trees -- routinely tens of thousands of tokens -- at every action step, causing progressive context degradation that erodes reasoning well before tasks complete. We argue that this coupling of observation frequency to action frequency is an architectural mistake. Drawing on the insight from Recursive Language Models that querying a document outperforms reading it wholesale, we propose Signal-Driven Observation (SDO): a dedicated sub-call reads the full DOM but returns only task-relevant elements and their selectors, and is re-invoked only when a lightweight signal detector fires -- triggered by URL transitions, newly visible interactive elements, action failures, or exogenous browser events. We outline the open problems SDO introduces and call on the community to treat observation compression as a core architectural decision in web agent design.

LLM-based agents increasingly tackle long-horizon tasks with interdependent decisions, where each action reshapes future constraints and intermediate errors can cascade. Existing RAG and agent memory systems organize histories by semantic similarity, retrieving content-relevant entries at decision time. We argue that this design mismatches execution-state dependencies: it fragments decision trajectories and mixes valid and erroneous traces, hindering coherent state reconstruction and error isolation. We propose MAGE (Memory as Agent-Guided Exploration), an active execution-state manager that stores interactions in a hierarchical state tree. The agent derives its state from the active root-to-current path, combining subgoal summaries, recent traces, and hints from prior branches. Four coupled operations maintain the tree: Grow records new traces, Compress summarizes completed subgoals, Maintain validates summaries, and Revise restores a target boundary and resumes on a new branch. This design bounds context growth while preserving state integrity and isolating flawed segments from the active path. Experiments on MemoryArena show that MAGE improves the average task success rate by 7.8--20.4 pp over baselines, while reducing token consumption by 55.1%.

Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.

We propose world-language-action (WLA) models as a new class of embodied foundation models. WLA takes textual instructions, images, and robot states as inputs to jointly predict textual subtasks, subgoal images, and robot actions, conjoining the \emph{world modeling interface} to learn from extensive egocentric videos as in the world-action model (WAM) and the \emph{language reasoning} capacities to solve complex long-horizon tasks as in vision-language-action (VLA) models. At the core of WLA lies an \emph{autoregressive (AR)} Transformer backbone, instead of a bidirectional diffusion Transformer as in WAMs, to predict the \emph{next state}, comprising the \emph{semantic-level} textual intention and complementary \emph{fine-grained} physical dynamics. The physical dynamics are supervised by the world modeling objective based on a dedicated World Expert, and are leveraged to ease the characterization of the state-action correlation for the Action Expert. WLA leverages meta-queries to make the world prediction \emph{implicitly} impact the action generation so that the former can be disabled during inference. The world prediction can also be activated to enable test-time scaling for improved robot control. Our WLA-0 prototype, with 2B active parameters, achieves 40 ms per inference on an NVIDIA RTX 5090. Evaluations across simulated and real-world environments demonstrate that WLA-0 achieves state-of-the-art multi-task and long-horizon learning abilities, e.g., 92.94% success rate on RoboTwin2.0 Clean and 56.5% success rate on RMBench. WLA-0 also holds the promise to learn novel tasks directly from \emph{cross-embodiment robot videos} without action annotations.

AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.

Persistent AI assistants, such as OpenClaw, accumulate large collections of related memories over long-term interactions. As these memories grow, they may reinforce one another, diverge across contexts, or directly conflict, making correct assistance depend on memory relations rather than isolated recall. Existing long-term memory benchmarks rarely probe how agents preserve and utilize such relations during downstream tasks. To address this gap, we introduce SubtleMemory, a benchmark for fine-grained relational memory discrimination in long-running AI agents. SubtleMemory constructs relation-controlled latent semantic artifacts whose variants instantiate complementary, nuanced, or contradictory relations, and embeds them into realistic user-agent histories, requiring agents to recover distributed relational structures during later queries and instructions. The benchmark contains 1,522 evaluation instances over 10 long histories, grounded in 1,090 relation-controlled memory-variant sets and spanning user-related and non-user-related queries. Evaluating six standalone memory systems, two Claw-style agents with native memory modules, and three Claw-style agents with plugin memory modules, we find that current systems remain weak on fine-grained relational memory discrimination. We further introduce diagnostic protocols that reveal distinct capability profiles across memory preservation, retrieval, and downstream reasoning stages.

Embodied AI systems are increasingly expected to reason and act over extended horizons in physical environments. This growing capability brings safety to the foreground, because failures in the physical world can harm people, damage objects, and disrupt workplaces. Although safe embodied AI has attracted substantial attention, the literature remains fragmented across planning, policy design, and runtime execution. Long-horizon robotic manipulation is a particularly revealing anchor domain for this problem because semantic misgrounding, subtask-level error propagation, execution drift, and contact-rich physical risk can accumulate within the same closed-loop system. This survey therefore provides a structured review of safety in long-horizon robotic manipulation from an embodied AI perspective. We organize the literature by intervention locus, covering planning-time, policy-time, and execution-time safety, and we analyze the strength of the evidence that each line of work provides, distinguishing formal guarantees, statistical support, and empirical safety heuristics. This framework clarifies the distinct roles of backbone capability papers, direct safety mechanisms, and benchmark or evaluation studies, while exposing where current safety claims are well supported and where they remain indirect. We identify persistent gaps, including limited evidence for policy-time safety, weak formal support for contact-rich long-horizon manipulation, immature uncertainty-triggered intervention, and a shortage of manipulation-specific safety benchmarks. We conclude by outlining research directions for cross-layer assurance, evaluation design, and safer deployment of long-horizon robotic agents in real-world settings.

Long-horizon autoformalization of research mathematics fails not only at hard lemmas, but at scale: statements drift, dependencies tangle, context decays, and local repairs corrupt distant work. We present LeanMarathon, a multi-agent harness for reliable research-level Lean autoformalization. Its core abstraction is an evolving blueprint: a Lean file that serves simultaneously as formal proof skeleton, natural-language proof graph, and shared system of record. Four contract-scoped agents construct, audit, prove, and repair this blueprint. These agents are coordinated by a two-stage orchestrator that first stabilizes target fidelity through adversarial review and then discharges the proof directed acyclic graph (DAG) from its dynamic leaves upward in parallel CI-gated rounds. LeanMarathon turns one brittle multi-hour run into many local, recoverable, parallel transactions. We evaluate LeanMarathon on two recent research papers spanning four Erdős problems (#1051, #1196, #164, #1217). Across three autonomous runs, it formalizes all seven target theorems with no sorry, proving 258 lemmas and theorems. These results show that reliable AI co-mathematics requires not only stronger provers, but durable harnesses that preserve target fidelity across long mathematical developments. The code can be found at https://github.com/YuanheZ/LeanMarathon.

Scaling robust robot policies requires more than broader randomization, because physical-domain experience must remain organized and learnable throughout training. We study when a policy can benefit from harder physics and identify recoverability as a central constraint in on-policy physical-domain scaling. In on-policy training, new dynamics are useful only insofar as they remain close enough to the current policy to generate corrective on-policy data, rather than collapsing rollouts into unrecoverable failures. Using quadruped locomotion as a physically demanding benchmark for embodied generalization, we introduce HORIZON, a checkpointed frontier curriculum that expands physical domains only within the current policy's recoverable boundary. HORIZON uses rollback and boundary refinement to govern each expansion step, turning fixed randomization into a continual process of physical-domain growth. Experiments reveal three regularities of physical-domain expansion. First, direct domain widening is uneven across physical axes and often unlearnable without staged ordering. Second, domain composition is non-monotonic, and adding more domains beyond a compact core can dilute recoverable joint samples and reduce overall robustness. Third, offline distillation of isolated experts cannot substitute for the joint interaction generated by on-policy curriculum. Together, these results frame physical-domain generalization as a continual growth problem for embodied control, with recoverability as the organizing principle for on-policy expansion.

Scientific and engineering progress is fundamentally a long-horizon iterative process: proposing changes, running experiments, measuring outcomes, and continuously refining artifacts. Yet existing benchmarks for frontier models primarily evaluate either single-turn responses or short-horizon agent trajectories, failing to capture the challenges of sustained iterative improvement over extended time horizons. To address this gap, we introduce AutoLab, a new benchmark for ultra long-horizon closed-loop optimization. AutoLab consists of 36 realistic, expert-curated tasks spanning four diverse domains: system optimization, puzzle & challenge, model development, and CUDA kernel optimization. Each task begins with a correct but deliberately suboptimal baseline and challenges agents to improve it within a strict wall-clock budget. Evaluating 17 state-of-the-art models reveals the dominant predictor of success is not the quality of an agent's initial attempt, but its persistence in repeatedly benchmarking, editing, and incorporating empirical feedback. While claude-opus-4.6 exhibits strong long-horizon optimization capabilities, most frontier models, including several proprietary ones, either terminate prematurely or exhaust their budgets with minimal progress. These results underscore the importance of time awareness and persistent iteration in autonomous agents. We open-source the full benchmark, evaluation harness, and task artifacts, to accelerate research toward truly capable long-horizon agents.

Long-horizon online visual mapping requires continuous camera-motion and scene-geometry estimation under bounded computation. Recent feed-forward 3D reconstruction models provide strong geometric priors, but streaming variants often predict poses in a fixed or historically maintained coordinate system, leading to train--test mismatch, early-anchor attention bias, and accumulated drift. We propose \emph{Anchor3R}, a current-centric streaming 3D reconstruction framework that predicts window-relative poses and local geometry in the current-frame coordinate system. Overlapping predictions form a dense relative-pose graph, supporting online pose updates and loop-aware motion averaging for global reconstruction. Experiments on indoor, outdoor, driving, and RGB-D benchmarks demonstrate improved long-horizon pose accuracy and dense reconstruction quality over existing streaming baselines. Despite being trained only on 48-frame sequences, Anchor3R directly generalizes to streams exceeding 10,000 frames while maintaining bounded GPU memory during online inference. Code is available at https://github.com/polar-explorer/Anchor3R.

Recent progress in Large Language Model (LLM) agents has enabled promising advances in automated data science. However, existing approaches remain fundamentally limited by their static action sets and lack of principled long-horizon context management, hindering their ability to accumulate reusable experience across tasks and operate reliably in multi-stage, iterative data science pipelines. To address these challenges, we introduce EvoDS, a self-evolving autonomous data science agent that learns to expand its skills and adaptively managing long-term context through agentic reinforcement learning. Specifically, EvoDS introduces two key strategies: (1) Autonomous Skill Acquisition (ASA) mechanism, which enables agents to synthesize, validate, and reuse executable skills; and (2) Adaptive Context Compression (ACC) strategy, which treats context management as a learned control problem rather than passive truncation. These strategies are orchestrated within a two-stage multi-agent training scheme, enabling EvoDS to autonomously improve over time. Theoretically, we prove that EvoDS's hierarchical design reduces tool-selection error, and its optimization objective aligns with an information bottleneck principle, ensuring efficient context use. Empirically, EvoDS outperforms state-of-the-art open-source data science agents by an average of 28.9% across four diverse benchmarks while eliminating out-of-token failures. Our code and data are available at https://github.com/usail-hkust/EvoDS.

A central goal of large language model (LLM) research is to build agentic systems that can plan, act, and adapt through sustained interaction with dynamic environments. While recent LLM-based agents exhibit impressive contextual reasoning, their long-horizon decision-making remains fragile, often suffering from objective drift, where goals and plans drift over extended interactions. We introduce Multi$^2$, a hierarchical multi-agent decision-making framework that explicitly decomposes agent behavior into complementary roles. A high-level agent (System 1) focuses on context-aware sub-goal generation using supervised fine-tuning (SFT), while a low-level agent (System 2) executes atomic actions through offline-to-online reinforcement learning (RL) in interactive environments. This separation enables stable long-horizon control, mitigates objective drift, and allows efficient adaptation. Across diverse interactive environments, Multi$^2$ consistently outperforms strong agentic baselines, demonstrating improved robustness and coordination in multi-turn interaction. Beyond performance, we introduce and release three hierarchical benchmark datasets, filling a long-standing gap in training and evaluating hierarchical decision-making for LLM-based agents.

Building memory is essential for long-horizon planning in zero-shot embodied navigation. Detector-centric scene graphs often compress observations into sparse nodes, discarding fine-grained visual evidence and accumulating noise, while 3D reconstruction-based methods remain computationally prohibitive. We present EvoMemNav, an efficient, self-evolving, fine-grained memory framework for zero-shot embodied navigation. EvoMemNav constructs a Visual-Semantic Memory Graph (VSMGraph) that keeps raw views as first-class memory and organizes them with lightweight semantic cues and topological relations into a room-view-object hierarchy, preserving fine-grained details for disambiguation and Stop verification. To scale to growing memory, we introduce a budgeted coarse-to-fine policy: a coarse stage compresses the search space into promising regions, and a fine stage invokes a VLM only for targeted verification and decision. Beyond static memories, EvoMemNav performs reflection-driven write-back after each subtask, updating graph-attached priors that encode accumulated environmental knowledge to refine future decisions without retraining. Experiments on GOAT-Bench and HM3D across object, text-description, and image-goal modalities show consistent gains in SR/SPL, with better multi-instance disambiguation, fewer premature stops, and stronger zero-shot generalization.

In robotic manipulation, the tight coupling between grasping and motion planning often obscures the true source of failure, leading to inefficient trial-and-error. To enable efficient long-horizon manipulation, we propose GTP-FA (Grasp-Then-Plan with Failure Attribution), a task-oriented two-stage grasp-then-plan framework that generates grasp candidates and performs downstream motion planning conditioned on the selected grasp. Given a failed manipulation trajectory, we learn a failure attribution model that generalizes to unseen grasps and produces a stable distribution over failure modes for diagnosis-guided optimization. Based on these attribution results, we then optimize both modules in a diagnosis-driven manner: on the grasping side, we inject task-level priors and risk penalties into grasp candidate scoring and optimization to suppress unstable or task-incompatible grasps; on the planning side, we target high-risk initial states through data collection and fine-tuning to address genuine planning bottlenecks. We evaluate the proposed framework in both simulation and real-robot experiments, and show that GTP-FA improves the corresponding base learners across RL, IL, diffusion-policy, and VLA-based settings, achieving substantially higher overall task success rates.

Real-world professional desktop workflows in specialized creative and engineering software unfold over long horizons and often require human-in-the-loop coordination, where agents proactively seek necessary information and users provide additional instructions, clarifications, feedback, or corrections as the task progresses. Yet existing desktop GUI benchmarks mostly reduce this setting to short, simplified tasks with all user instructions provided upfront. To address this issue, we introduce DeskCraft, a desktop GUI benchmark targeting long horizon creative and engineering workflows and proactive human-agent collaboration. DeskCraft organizes tasks into a multilevel difficulty taxonomy, with long horizon tasks requiring over 50 execution steps, and covers professional creative software across design, video, audio, and 3D creation. Furthermore, DeskCraft formalizes human-agent collaboration into an interaction protocol covering mid-turn and post-turn exchanges. Mid-turn interaction captures both agent-initiated clarification under uncertainty and user-initiated interruption during execution, while post-turn interaction accommodates user-driven feedback after the agent signals completion, together spanning the full space of realistic collaboration patterns. We evaluate 18 proprietary and open source agents on 538 tasks and find that GPT-5.4 reaches 31.6% on standard tasks and 27.6% on interactive tasks. Further analyses reveal persistent failures in long horizon workflow delivery and proactive clarification. We will open-source all evaluation codes, tasks, and data at https://github.com/mrwwk/DeskCraft.

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 ${\sim} 1.5 \times$, 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.

Large language agents are increasingly used for social simulation, yet it remains unclear whether they can sustain coherent behavior in structured organizations, where goals must propagate through hierarchy, tasks depend on prior execution, and artifacts accumulate over long horizons. We formulate long-horizon organizational simulation as a memory-centered coordination problem and introduce TaskWeave, a hierarchical agentic framework that maintains planning states through a Formulate-Partition-Diagnose-Align cycle and grounds execution through dependency-aware trace memory. We evaluate TaskWeave in a year-long IT company simulation and compare it with other multi-agent frameworks on organizational coherence, execution grounding, and downstream enterprise NLP utility. Experiments show that TaskWeave supports coherent and long-horizon organizational dynamics while producing grounded artifacts and adapting to external environments. These findings suggest that structured simulation memory is a key mechanism for building reliable LLM-based organizational simulators.

Imitation learning enables robots to learn how to execute tasks via observation. However, real-world environments like homes and offices are often severely partially observed due to their large spatial scales. In addition, many tasks involve executing a series of subtasks requiring autonomous robots to reason over extended time horizons. To address these challenges, we propose using scene graphs as an explicit and structured memory mechanism in imitation learning. By maintaining a dynamic scene graph that captures object-centric relationships and their evolution over time, our method allows the agent to retain relevant historical context during task execution to efficiently reason over incrementally accrued scene information. Our experiments on simulated mobile manipulation and real-world tabletop manipulation demonstrate that our approach substantially improves policy performance, particularly in settings that demand long-term reasoning and robust generalization under partial observability.

Vision-Language-Action (VLA) models have shown remarkable progress for mobile manipulation, but their performance on long-horizon tasks remains poor. These tasks are especially challenging because (1) progress toward high-level goals must be maintained across extended sequences of spatially distributed subtasks, and (2) early execution errors compound rapidly over the task horizon. These challenges persist despite finetuning on large human teleoperated mobile manipulation data, indicating that more data alone may not resolve the problem. To address these challenges, we propose MPVI: Motion Planner / VLA Interleaving, a framework that integrates model-based motion planning with VLAs to improve robustness without further training. The proposed integration enables localization and navigation to distant or occluded target objects through cluttered scenes using open-vocabulary object detection, frontier exploration and motion planning. However, such integration is non-trivial, requiring reliable switching between modules; we show one way forward via VLM-based completion checking with proprioceptive triggers. We evaluate our approach on the BEHAVIOR-1K benchmark and demonstrate 113% improvement in task progress over a top end-to-end VLA baseline. Additional details are available at the project page: https://mpvi.netlify.app/.

Diffusion models provide strong priors for generating structured data, but many tasks require outputs beyond the scale on which these models are typically trained. Compositional generation addresses this by composing overlapping local plans from a pretrained short-horizon prior into a long-horizon output. However, standard composition primarily enforces agreement between neighboring local plans, yielding local consistency without directly specifying the global structure of the full composition. As a result, locally compatible plans may still form an implausible route, task sequence, or temporal evolution. Existing methods improve global coherence by repeatedly propagating local consistency signals or by adding inference-time optimization, but these procedures become expensive as the number or dimensionality of local plans increases. We propose Coarse-to-Fine Compositional Diffusion (CoFi), an inference-time sampler that separates global structure formation from local detail refinement. CoFi first aligns local denoised estimates around a shared coarse structure, producing a global scaffold that captures the long-range task-level arrangement. It then diffuses this scaffold to an intermediate noise level and denoises it with the same pretrained local prior, restoring local fine structure while preserving the scaffold-induced global coherence. Across long-horizon robotic planning, panoramic image generation, and long video generation, CoFi not only improves both global coherence and local sample quality over prior compositional baselines, but also requires 2-8x fewer denoiser evaluations.

Recent advances in agentic AI have enabled agents to complete complex tasks through tool use, reasoning, and multi-step planning. Yet existing benchmarks evaluate agents within a single session, ignoring past actions, stated preferences, and prior decisions that agents must integrate to fulfill personalized user goals. We introduce Momento, a benchmark for persistent agentic task completion in multi-session service environments, requiring agents to take consequential, tool-mediated actions while resolving temporal dependencies and evolving user goals across sessions. Experimental results reveal that current agents fail primarily through misestimation of user state, treating prior session history as a reliable proxy for current context rather than stale information requiring re-validation, highlighting a substantial gap between current agent capabilities and realistic long-horizon human-agent interaction.

Deploying lightweight Large Language Model (LLM) agents on edge servers can reduce latency and move agentic services closer to users, but resource-constrained edge models often struggle with long-horizon tasks that require persistent memory, subgoal tracking, and reflection. Fine-tuning edge models after deployment is costly and difficult to scale across heterogeneous nodes, while purely local memory leaves agents with isolated experience and growing prompt context. We propose \textsc{CoMIC}, a parameter-update-free cloud-edge framework for Collaborative Memory and Insights Circulation. \textsc{CoMIC} follows a \textit{Centralized Reflection, Decentralized Execution} design: edge agents execute locally using subgoal-oriented hierarchical memory and selective re-expansion of relevant histories, while a cloud-side LLM critic asynchronously evaluates completed trajectories, filters reusable experience, and aggregates cross-agent guidance keyed by semantic subgoal identifiers. Across five long-horizon agent tasks spanning symbolic planning and text interaction, \textsc{CoMIC} improves progress rate and action grounding for weak edge agents and yields task-dependent success-rate gains without updating model parameters.

Long-horizon autonomous agents require memory systems to retain historical information, track evolving states, and reuse relevant knowledge beyond finite context windows. Existing agentic memory systems typically follow a memory construction-retrieval (MCR) pipeline, but often adapt mainly the memory bank while keeping the surrounding pipeline fixed after deployment. This fixed-pipeline design struggles to handle heterogeneous task-specific failure modes and can become misaligned with memory banks that evolve in scale and structure over time. To address these limitations, we propose MemPro, a system-level evolution framework that treats the entire MCR pipeline as an evolvable program rather than adapting only the memory bank or prompt text. MemPro maintains a version tree of runnable memory-system implementations, where an Evolving Agent iteratively selects promising versions, diagnoses recurring failures, and creates improved child versions through failure-mode-guided edit-debug refinement. Experiments on LongMemEval, LoCoMo, HotpotQA, and NarrativeQA show that MemPro consistently outperforms strong static and prompt-level evolving baselines within a few iterations, continues to improve with evolution, and achieves a favorable performance-cost trade-off. Code is available at https://github.com/wanghai673/MemPro.

Long-horizon LLM agents produce safety evidence across long trajectories, where sparse, delayed, and compositional risk signals often escape local moderation. Existing turn-level or short-context detectors struggle to reliably retain and aggregate such evidence over extended horizons. We reframe long-horizon agent safety detection as trajectory-level evidence compression and propose Trajectory Risk-Aware Compression for Long-Horizon Agent Safety (TRACE). TRACE uses a Compressor-Reader design: the Compressor encodes the full trajectory into a compact latent evidence state under trajectory-level supervision, and the Reader judges the raw trajectory with this latent evidence state as a safety reference. This design helps aggregate dispersed risk cues and reduce premature evidence loss. Across ASSEBench, Pre-Ex-Bench, and R-Judge, TRACE achieves the best accuracy on all evaluated backbones, improving over strong baselines by up to 12.6 percentage points. On LongSafety, TRACE shows smaller performance degradation as context length grows. Attention visualizations and case studies suggest that the compressed reference helps the Reader focus on risk-critical segments and recover cross-step evidence. Code is available at https://github.com/Peregrine123/TRACE_official.

Large language models are increasingly deployed as tool-augmented agents to acquire information beyond parametric knowledge. While recent work has improved long-horizon tool-use reasoning, most approaches focus on tasks with a single correct answer. In contrast, many real-world queries require discovering a comprehensive set of valid answers, a setting known as Multi-Answer QA. This setting raises two challenges: fine-grained credit assignment over long search trajectories and reward alignment for sustained exploration beyond easy high-frequency entities. We propose SPADER, a reinforcement learning framework for long-horizon tool use in Multi-Answer QA. SPADER includes Step-wise Peer Advantage (SPA), a critic-free step-level credit assignment mechanism that aligns parallel trajectories by decision step and estimates advantages from peer returns. It also includes a diversity-aware exploration reward that promotes long-tail entity discovery by upweighting rare findings and downweighting redundant ones. Experiments on QAMPARI, Mintaka, WebQSP, and QUEST show that SPADER generally improves recall and overall F1 over prompting-based agents, outcome-supervised RL methods, and recent step-level supervision approaches. Our code and model weights are available at https://github.com/KhanCold/spader.

For agents to learn continuously from interaction with the world at test time, they must be able to explore effectively, acquire new world knowledge and skills, retain relevant episodic experiences, and plan over long horizons. To evaluate these key abilities of test-time continual learning agents, we introduce AgentOdyssey, a novel evaluation framework that procedurally generates open-ended text games with rich entities, world dynamics, and long-horizon tasks. Critically, AgentOdyssey goes beyond the conventional machine learning assumption that learning does not occur at test time by placing agents in a continuous, long-horizon setting that interleaves learning and inference throughout deployment. We further propose a multifaceted evaluation methodology that measures not only game progress but also offers diagnostic tests on world knowledge acquisition, episodic memory, object and action exploration, action diversity, and model cost. We evaluate diverse agent paradigms in the generated games. Our experimental results reveal critical limits in agents' key abilities, as well as factors that influence their meaningful horizon. Although performance scales with stronger base models, even the top agent remains far below human performance, leaving substantial headroom for improvement. Among agent mechanisms, we find that short-term memory benefits multiple agent paradigms and is an important component of agent test-time training.

The diffusion based robot navigation world models are typically trained using parallel supervision, while autoregressive inference is employed during path planning. This results in a distribution shift between training and inference, which destabilizes the performance over long-horizon prediction. We propose AR Forcing, an autoregressive training strategy, which integrates the standard diffusion loss into the autoregressive training loop. At each step, the model uses its own predictions to update the context and optimize the single step noise prediction objective, thereby explicitly exposing the model to the inference state distribution during training. Our method does not require additional discriminators or distribution-matching losses, retains the original diffusion framework and sampler, and is easy to integrate. Experiments on multi-domain navigation datasets (RECON, SCAND, HuRoN, TartanDrive) show that compared with strong baselines, AR Forcing improved the consistency of generated images during long-horizon navigation and the accuracy of predicted trajectories, enhancing robustness of the model in complex known and unknown environments. We will release the code soon.

Context management enables agentic models to solve long-horizon tasks through iterative summarization of previous interaction histories. However, this process typically incurs substantial decoding overhead for the extra summarization tokens, which significantly affect the end-to-end response latency at deployment. In this paper, we introduce CoMem, a novel framework that decouples memory management from the primary agent workflow, enabling these processes to execute in parallel. We propose a $k$-step-off asynchronous pipeline that overlaps the memory model's summarization with the agent's inference, effectively masking the latency of context processing. To ensure robustness under this asynchronous setting, we introduce a reward-driven training strategy that aligns the memory model to capture sufficient statistics for the agent's decision-making. Theoretical analysis confirms that CoMem offers a superior efficiency-effectiveness trade-off compared to coupled architectures. Our extensive experimental results on SWE-Bench-Verified show that CoMem provides 1.4x latency improvements upon vanilla long-context solutions while preserving most of the performance. Furthermore, we demonstrate that these latency gains scale favorably with increased system throughput, offering a modular path forward for the independent optimization of agent reasoning and memory compression.

LLM agents increasingly face long-horizon tasks such as web search and deep research in real-world applications, where accumulated context can cause long-context degradation and reasoning failures. Prior work mitigates this through context management with agent-side context control or fixed strategies such as summarization, which require training the agent itself for adaptation - making it impractical for closed-source agents and ignoring that different agents may require different strategies. We introduce Adaptive Context Management (AdaCoM), which trains an external LLM to manage the context of a frozen agent through flexible modification actions and end-to-end reinforcement learning. Across diverse agents on web search and deep research benchmarks, AdaCoM substantially improves performance by preserving task constraints and progress while pruning stale content. The learned strategies reveal a Fidelity-Reliability Trade-off: agents with higher vanilla ReAct performance benefit from higher-fidelity context preservation, whereas lower-performing agents require more aggressive compression to stay within a reliable reasoning regime. Transfer experiments show that AdaCoM generalizes most effectively across agents with similar capability (measured by vanilla ReAct performance), suggesting a practical path toward reusable context managers for agent systems.

Non-prehensile robot manipulation is challenging due to discontinuous, long-horizon interactions between the robot and the objects it manipulates. Sampling-based model predictive control methods are effective with discontinuous contact but face challenges with finding promising trajectories in long-horizon planning. We propose a closed-loop object-informed (CLOI) method that splits the problem into object-level planning to find long-horizon object poses that lead the object to its goal, and robot-level planning to select robot actions that follow those poses. We use model predictive path integral (MPPI) control to solve the subproblems and couple their solutions through consensus on the object poses using the alternating direction method of multipliers (ADMM). The object plan is revised toward robot-realizable object trajectories, while the robot plan is aligned with the object poses the task requires. In planar pushing tasks with obstacles using an xArm6 manipulator, CLOI increases the success rate by 35% in simulation and 43% on hardware, compared to standard MPPI given the same computational budget.

Extending humanoid traversal to the open world is key to practical deployment in human environments, but remains challenging. The robot must use vision to ensure safe and reliable foot placement on heterogeneous terrain under highly dynamic motion, while producing coordinated, natural whole-body behaviors. We propose SSR, an efficient end-to-end framework for egocentric vision-based humanoid traversal that jointly learns these capabilities. SSR introduces imagined foothold guidance, which learns to model forthcoming swing-foot contacts and evaluates their support to guide pre-touchdown swings toward stable regions, reducing edge slips. It further employs equivariant latent-space symmetry augmentation to efficiently induce bilateral coordination under high-dimensional visual observations, and uses terrain-specific multi-discriminator motion priors to encourage human-like behavior across scenes. Extensive experiments show that SSR achieves safe, stable, and high-quality locomotion on diverse real-world terrains, including stairs with varied structures and extreme challenges such as wide gaps and high platforms, while enabling reliable long-horizon traversal in open outdoor environments.

Data selection is increasingly used to reduce the cost of large language model (LLM) fine-tuning, with recent methods prioritizing samples by current utility, diversity, quality, or influence. This paper studies a different question: when fine-tuning occurs over multiple stages, can selection strategies that look optimal now make the model less adaptable later? We introduce a long-horizon view of LLM data selection in which a selector is evaluated not only by immediate task performance, but also by future adaptation speed, forgetting, capability imbalance, and out-of-distribution robustness. We compare representative random, loss-based, gradient-based, diversity-based, quality-based, and utility-diversity selection families under a unified multi-stage protocol. Through controlled experiments designed to instantiate this protocol, we show how short-term selectors can exhibit rank reversal: they improve the current stage while slowing subsequent learning and increasing forgetting. We formalize this behavior as \emph{myopic selection}, provide a simple local analysis of why it can occur, and propose a diagnostic Long-Horizon Aware Selection (LHAS) objective that augments immediate utility with coverage, future-proxy transfer, and anti-concentration terms. The study argues that data selection should be evaluated as a training intervention that shapes the model's learning trajectory, rather than only as a local data-efficiency mechanism.

Real-world data analysis is inherently iterative, yet existing benchmarks mostly evaluate isolated or short interactive tasks, leaving agents' ability to track evolving analytical context over long horizons untested. We introduce LongDS, a benchmark for long-horizon, multi-turn data analysis where agents must maintain, update, restore, and compose evolving analytical states. LongDS comprises 68 tasks constructed from real-world Kaggle notebooks, spanning 2,225 turns across six domains including Geoscience, Business, and Education. Tasks are designed around state-evolution patterns (e.g., counterfactual perturbation, rollback, multi-state composition), with an average dependency span of 11.3 turns. Evaluating five state-of-the-art models, we find that the best model reaches only 48.45% average accuracy, performance drops nearly 47 points from early to late turns, and long-horizon errors account for 52%--69% of failures. Further analysis shows that additional agent steps do not necessarily improve performance, suggesting that the key bottleneck is maintaining a correct analytical state rather than increasing interaction budget. We release LongDS to support research on reliable long-horizon agentic data analysis. Code and data will be released at https://github.com/zjunlp/DataMind.

Memory-augmented LLM agents tackle complex long-horizon tasks by recursively summarizing interaction trajectories into compact memory. However, existing approaches typically train these memory policies using outcome-based reinforcement learning, failing to localize where intermediate memory quality degrades. As interactions unfold, ambiguous recursive summaries progressively discard task-relevant information and introduce semantic noise. This exacerbates belief deviation, obscuring the agent's estimate of the latent task state and ultimately derailing long-horizon reasoning. We therefore argue that memory optimization should focus not merely on trajectory-level success, but on the clarity of the belief induced by intermediate summaries. To this end, we introduce Belief Entropy, a self-supervised proxy that probes how uncertain the model remains about the latent task state given its current memory. Based on this proxy, we propose Metacognitive Memory Policy Optimization (MMPO). Instead of relying only on sparse outcome-based signals, MMPO provides fine-grained, memory-specific supervision via explicitly penalizing summaries that induce high epistemic uncertainty. Experiments show that MMPO consistently outperforms existing methods on diverse long-horizon tasks, maintaining 97.1% performance even when scaled to 1.75M-token contexts.

Multi-step robot manipulation requires acting under uncertainty about how the scene will evolve, making exploration and policy adaptation challenging. We study whether short-horizon, task-consistent future videos can provide useful structured priors for control and reinforcement-learning fine-tuning. We formalize this idea through Future-Experience Conditioning (FEC), a simple interface that conditions closed-loop policies on a latent representation of a short future video. In our simulation setup, future clips are generated in three stages, an LLM reasoner operating over a task ontology initialized from the current scene state, a robot-free digital-twin rollout of the intended object motion, and a mask-free video diffusion model that synthesizes a robot-consistent future clip without requiring segmentation at inference. We instantiate this future-conditioning interface primarily with BC and BC+RL, and compare against a future-conditioned Streaming Flow Policy (SFP) baseline on RoboCasa and CALVIN under NoFuture, GTFuture, GenFuture, and WrongFuture. Generated futures improve performance over no-future conditioning, while mismatched futures degrade it, and our BC+RL instantiation achieves the strongest overall results. An average BC+RL learning-curve analysis across 8 CALVIN tasks further shows that GTFuture improves fastest, GenFuture improves earlier and to a higher level than NoFuture, and WrongFuture remains at zero throughout training. These results suggest that short-horizon future videos can serve as useful structured priors for exploration and policy adaptation under imperfect future predictions. https://enact2026.github.io/

LLM-based multi-agent systems (MAS) have emerged as an effective paradigm for complex and long-horizon tasks. However, in real-world tasks, MAS often exhibit various failures during execution and such failures are difficult to eliminate during design. This motivates experience-driven MAS evolution, where a system improves based on its own execution experience. Yet such evolution is challenging because MAS experience is prolonged and intricate, interleaving multiple agents' execution chains and communication messages, which makes it difficult to identify what should be improved. To address this challenge, we propose Meta-Team, an experience-driven MAS evolution framework based on collaborative self-evolution. Meta-Team preserves the execution context of each agent and coordinates post-task communication, enabling agents to exchange distributed evidence for evolution. Building on this design, Meta-Team conducts multi-scale self-evolution, transforming execution experience into reusable improvements to agent behaviors, inter-agent coordination, and team-level organization. Across six long-horizon agent benchmarks, Meta-Team consistently outperforms single-agent systems, hand-crafted MAS, and prior MAS evolution methods; further analyses demonstrate that Meta-Team enables more reliable and scalable MAS self-evolution.

While GUI agents have advanced rapidly, they often lack the robustness to recover from their own errors, hindering real-world deployment. To bridge this gap at both the evaluation and data levels, we introduce GUI-RobustEval and propose Robustness-driven Trajectory Synthesis. GUI-RobustEval contains $1,216$ executable test cases that systematically measure error recovery capabilities across a broad and realistic spectrum of error modes. At the data level, RoTS is a scalable synthesis framework that creates $800k$ high-quality data via a tree-based pipeline that proactively discovers diverse error modes and synthesizes corresponding recovery steps. Our two models, RoTS-7B and RoTS-32B, fine-tuned on our dataset, both demonstrate significant gains on GUI-RobustEval and traditional GUI benchmarks. Notably, RoTS-32B achieves state-of-the-art performance on OSWorld, with a $47.4\%$ success rate and a $33.8\%$ All-Pass@4 score, suggesting that improved long-horizon error recovery ability contributes to both robustness and overall performance. Our code is available at https://github.com/AlibabaResearch/RoTS.

Mobile GUI agents excel at immediate reactive control but frequently fail in realistic, long-horizon tasks that require memory. This failure stems from a fundamental conflict between limited context windows and token-heavy screenshots. To save the limited context, agents must progressively discard older visual history, permanently losing crucial transient information. Furthermore, existing action-centric datasets fail to teach agents what or when to explicitly memorize, and augmenting static real-world data is prohibitively expensive and lacks interactive verification. To resolve this, we present STAMP, a framework that trains explicit memory in mobile agents through controllable virtual environments, where deterministic memory variables are programmatically injected into synthesized tasks to control what must be memorized, when it should be encoded, and when it must later be retrieved, thereby producing verifiable supervised data at scale and enabling online reinforcement learning through environment-driven reward feedback. Evaluated on our newly introduced Memory-World benchmark, the resulting Stamp-GUI agent achieves state-of-the-art performance among GUI-specialized models and sets a new high watermark on our Memory-World benchmark, demonstrating exceptional memory accuracy and task resilience while maintaining strong general mobile navigation capabilities.

Web agents, which couple language models with browsing and tool-use capabilities, show promise as open web assistants. Yet progress is increasingly limited by the lack of scalable, process-level supervision. Existing benchmarks are largely manually constructed, providing only coarse start-goal annotations without intermediate trajectories, while recent automatic generation efforts remain expensive, biased, and shallow. These limitations prevent reliable training and evaluation of agents that must generalize to realistic, multi-hop, cross-page tasks. We introduce a scalable framework, GTA, that integrates crawling, retrieval-based seeding, in-context generation, and automated quality control to produce realistic tasks paired with executable trajectories. This design decouples crawling from generation for greater efficiency, grounds tasks in the site graph to enforce compositionality, and ensures dense supervision through deterministic replays and systematic validation. We instantiate the pipeline on over 50 websites covering e-commerce, government, forums, and news, with multilingual and multi-hop coverage. The resulting benchmark reveals a significant human-agent performance gap and enables detailed diagnostics. Our contributions are three-fold: (i) formalizing multi-hop web-agent task generation, (ii) proposing an efficient and validated pipeline for automatic data creation, and (iii) releasing a dynamic benchmark with reproducible evaluation.

We present ProgVLA, a compact vision-language-action (VLA) model designed for reliable robot manipulation under tight compute and memory budgets. The model specifically focuses on efficiently processing long multi-modal sequences by maintaining an explicit representation of task progress over extended horizons. To this end, ProgVLA integrates two key components. First, a multi-modal encoder with a two-stage Perceiver resampling scheme compresses variable-length visual, language, and proprioceptive streams into a fixed set of control-ready context tokens, substantially reducing sequence length while preserving cross-modal grounding. Second, an auxiliary set of progress heads is trained with offline reinforcement learning (RL) objectives to jointly learn critics over normalized remaining-horizon targets. This provides the policy with an internal estimate of task progress and enables advantage- and success-weighted flow-matching imitation learning. On two well-established multi-task robot manipulation benchmarks, a 0.1B-parameter ProgVLA model reaches success rates that are competitive with, and on long-horizon and harder task tiers exceed, substantially larger pretrained baselines. Ablations indicate that the learned context resampler and task-adaptive visual fine-tuning are the largest single contributors, while progress-aware training provides a consistent additional gain that is concentrated on long-horizon and multi-object tasks. We further validate the approach in real-world toy-kitchen environments.

Offline goal-conditioned reinforcement learning (GCRL) is challenging in long-horizon tasks, where distant state--goal pairs provide weak supervision and value estimates become vulnerable to accumulated bootstrapping errors. Hierarchical methods mitigate this difficulty by introducing intermediate subgoals, but fixed temporal abstractions or fixed hierarchy depths can be mismatched to state--goal pairs with different reachability horizons. We propose Coarse-to-Fine Hierarchical Goal Reinforcement Learning (CFHRL), a fully offline GCRL framework that adaptively refines distant goals before execution. Starting from the final goal, CFHRL recursively proposes intermediate targets, trained from replay-supported candidates, and stops refinement once the current target is estimated to be locally executable by a learned reachability cost. The key idea is that a subgoal need not be an exact midpoint or globally optimal waypoint; it only needs to provide reliable progress and reduce the remaining reaching difficulty, enabling subsequent refinement over shorter horizons. A stylized analysis further supports the robustness of approximate recursive contraction. Experiments on OGBench show substantial gains on several long-horizon tasks, with ablations validating the proposed refinement and stopping mechanisms

We present PEAM, a Parametric Embodied Agent Memory framework in Minecraft that transforms agent memory from inference-time retrieval into parameter-resident skills internalized through experience. PEAM pairs a slow deliberative LLM for open-ended reasoning with a fast parametric module for reflexive execution of consolidated skills. The fast module is a multimodal Mixture-of-Experts LoRA architecture with per-category physically isolated adapters, enabling parameter-level continual learning without catastrophic forgetting. We treat failure as a first-class training signal: failure--correction trajectory pairs are internalized through a joint behavioral-cloning and contrastive objective, so the agent learns not only what succeeds but also how corrected actions differ from failed ones. To govern consolidation, PEAM introduces a parameterization-worthiness score for deciding which experience should be internalized, and a scale-free self-triggered consolidation mechanism for deciding when to internalize without task-specific hand-tuned thresholds, making the agent self-evolving as the trigger transfers across task distributions without re-tuning. Experiments in Minecraft show that PEAM improves long-horizon task performance, mitigates forgetting on previously consolidated skills, and improves parametric-versus-retrieval efficiency over retrieval-based embodied agents and parametric memory variants.

Transformers process images and videos by flattening space and time into long token sequences. While attention and KV caching preserve past features, their memory grows with sequence length and they lack an explicit, persistent spatial state, making long-horizon video understanding and occlusion-sensitive reasoning difficult. We propose Tensor Memory, a lightweight module that augments Transformer blocks with a fixed-size recurrent 3D memory tensor: tokens write into a voxel grid via a differentiable soft write that deposits content as a Gaussian-weighted volume around a predicted continuous 3D location, the memory is updated with an efficient local interaction operator and gated recurrent dynamics, and tokens read back context via continuous sampling with gated residual fusion. Because the memory tensor has a constant size, Tensor Memory decouples state capacity from input length while preserving a spatial inductive bias. We evaluate the module on standard language, image, and video benchmarks and on a controlled toy diagnostic suite designed to isolate when persistent state is beneficial; it integrates with standard Transformer training pipelines and can be attached to or removed from existing blocks without other architectural changes.

Planning with world models is bottlenecked by compounding prediction errors and the difficulty of defining optimizable goals. Visual targets provide precise local gradients but poor distant guidance, while language is flexible yet limited by noisy cross-modal alignment or dependence on large generative models unsuited for the high-sampling nature of model-based planning. To address these challenges, we introduce Latent Goal Prediction from Language (LAGO), a framework that predicts both sequences of intermediate goal states from language instructions and action-conditioned rollouts, all within the same latent space. Rather than optimizing toward a single global objective, LAGO dynamically decomposes instructions into explicitly predicted, locally tractable latent subgoals. By updating these subgoals online and using a soft minimum trajectory cost during planning, LAGO enables an agent to follow coherent latent trajectories over long horizons. Evaluation across multiple environments planning horizons shows that LAGO avoids the sharp degradation of prior methods. By achieving robust and precise long-horizon planning purely from language, LAGO bridges the precision of visual goals with the flexibility of text-guided control.