Long-Horizon Tasks

This paper investigates the problem of position estimation of unmanned surface vessels (USVs) operating in coastal areas or in the archipelago. We propose a position estimation method where the horizon line is extracted in a 360 degree panoramic image around the USV. We design a CNN architecture to determine an approximate horizon line in the image and implicitly determine the camera orientation (the pitch and roll angles). The panoramic image is warped to compensate for the camera orientation and to generate an image from an approximately level camera. A second CNN architecture is designed to extract the pixelwise horizon line in the warped image. The extracted horizon line is correlated with digital elevation model (DEM) data in the Fourier domain using a MOSSE correlation filter. Finally, we determine the location of the maximum correlation score over the search area to estimate the position of the USV. Comprehensive experiments are performed in a field trial in the archipelago. Our approach provides promising results by achieving position estimates with GPS-level accuracy.

Long-horizon physical-world agents must reason over distant goals while grounding decisions in reliable closed-loop behavior. Today's foundation models split these capabilities: vision-language models (VLMs) infer missing information and adapt high-level plans but remain brittle and inefficient at repeated navigation grounding, while navigation foundation models (NFMs) robustly execute semantic goals but operate as bounded episodes without persistent task-level reasoning. We introduce NavMCP, an agentic scaffolding framework that couples a VLM reasoning agent with an NFM executor for long-horizon exploration. The VLM decides what evidence to seek, where to search, and when to stop, while the NFM grounds each semantic sub-goal into closed-loop navigation. Three channels structure their collaboration: intent translates evidence needs into navigation calls, observation converts rollouts into source-grounded trajectory evidence, and memory accumulates findings, negative evidence, and unresolved goals across calls. This design turns isolated navigation rollouts into persistent embodied interaction without retraining either model. On Embodied Question Answering, NavMCP achieves state-of-the-art results on HM-EQA, MT-HM3D, and EXPRESS-Bench. Under matched agent and executor backbones, it outperforms an episodic interface by 14.9 percentage points on HM-EQA. On a Unitree Go2, NavMCP reaches 78.3% success, with its margin over the strongest baseline growing from 10 to 45 points as the task horizon increases. These results demonstrate the potential of scaffolding complementary foundation models into long-horizon physical-world agents.

Large language model (LLM) agents are increasingly deployed in long-horizon, interactive, and stateful environments. In these settings, a single wrong action, such as refunding the wrong purchase, can cause irreversible task failure and must be intercepted before execution. Such failures may not appear in every single run, but can emerge across repeated trials, making reliability across steps and trials critical. However, ensuring agentic reliability is challenging: even frontier LLMs struggle to explain why an action may be wrong, especially in long, intertwined trajectories governed by domain-specific policies. Much recent work relies on prompt-based critique agents, while optimization-based methods lack a systematic way to produce rich verification rationales for training. We address this gap with CAST, a critique-aware training framework that converts sparse task outcomes into action-level supervision for critique learning and policy optimization. CAST analyzes agent trajectories to synthesize structured rationales explaining action validity under partial observability. The resulting critique model is used to construct critique-aware training data for optimizing the policy model. Fine-tuning Qwen3-family models on dynamic tool-calling benchmarks, CAST improves reliability across domains, outperforming GPT-OSS-120B by over 10% pass^4 on Retail tasks and yielding an additional 9% improvement on Telehealth in an out-of-domain setting. These results demonstrate that critique-aware training improves the robustness of LLM agents in realistic dynamic environments.

Long-horizon agents need a context manager to compress growing interaction histories into a bounded working context, via passive strategies or active strategies that decide how memory is accessed and reorganized. Meanwhile, prior optical-memory work mainly treats pixels as a dense codec for textualized histories, often presupposing that rendering context into optical memory incurs a significant performance drop relative to text, thus coupling this representation with SFT, self-distillation, or reinforcement learning to close this gap, leaving unresolved (i) how visual rendering performs as a context manager under a fair, controlled comparison, and (ii) whether this carrier offers a native advantage when history is inherently multimodal. In this paper, we formulate context management as a budget-constrained history transformation and introduce Visual Rendering (VR) as a representational context manager. Under a shared harness, policy model, trigger, and task domain, we evaluate VR on four text-centric and three multimodal benchmarks against four baselines (No Compression, Discard-All, Sliding Window, Summarization), finding visual memory is a natural carrier of native visual evidence. Building on this finding, we propose VERA (Visual Evidence-Retaining strategy for long-horizon Agents), a training-free context manager built on deterministic rendering with no exposed memory operations: on text-centric benchmarks it renders textual history as VR does, while on multimodal benchmarks it retains native visual observations instead of translating them into text. Across nearly all benchmarks, VERA cuts cumulative non-cache tokens by 31.5%-63.1% versus No Compression, matches existing managers on text-centric tasks, and achieves the highest accuracy among all baselines on multimodal tasks, supporting a modality-preserving view of long-horizon context management.

Early failure prediction is important for long-horizon agents, as it enables timely intervention and can reduce inference and tool-use costs. Uncertainty quantification, such as verbal confidence and perplexity, offers a promising approach to detecting agent failures; however, it has not been explored whether these signals retain their discriminative power during the intermediate stages of long-horizon execution. We evaluate mainstream uncertainty signals on deep-research tasks and find that verbal confidence reliably distinguishes failures at trajectory completion, achieving a mean AUROC of 0.85, whereas all evaluated signals offer limited predictive value earlier in execution, with none exceeding a mean AUROC of 0.60 at 50% trajectory progress. We identify an underlying mechanism explaining this gap: path switching, where agents frequently abandon their current search direction in-trajectory, breaking the link between early signal and final outcome. These findings challenge the assumption that intermediate uncertainty can reliably guide early intervention. They also motivate a practical recommendation for agent harnesses in deep-research settings: use final-step confidence to decide whether to restart, an approach that our experiments find more effective than in-trajectory intervention.

Long-horizon story-driven video generation requires a production agent to coordinate narrative decomposition, state tracking, shot design, prompt construction, rendering, and revision across interdependent scenes. Existing adaptive video systems primarily refine requests or reusable skills, leaving recurring production failures disconnected from persistent, stage-targeted improvements across stories. We introduce CineForge, a self-evolving video-production agent framework that couples CineForge-Produce for video generation with CineForge-Evolve for cross-story policy evolution. CineForge-Produce organizes each source story into typed narrative, character, spatial, and cinematic states, uses them to coordinate asset and clip generation, and records the process as a canonical production trajectory. CineForge-Evolve applies Case-to-Pattern-to-Policy Evolution (CPPE) to review trajectory evidence, consolidate recurrent findings into bounded stage-local patches, and deploy validated updates through structural replay and confidence-controlled paired evaluation. To measure complete story realization, we introduce CineScope, which combines a 100-script CineScope-Data suite with a human-aligned, multiscale CineScope-Metric spanning causal state, directorial orchestration, pacing and resource allocation, and character arc. Across CineScope-Data and two public benchmarks, the evolved CineForge policy improves CineScope-Metric from 4.024 to 4.380, outperforms three long-video baselines with consistent gains under ScriptAgent, and reduces review LLM calls by 37.0% on new stories. These results establish production trajectories as actionable experience for video agents that improve cumulatively across long-form storytelling tasks.

Humans traverse complex terrain over long distances without losing balance, whereas perceptive humanoid policies become fragile as perception and control errors accumulate. We present SOLO, a unified framework addressing two compounding causes of this long-horizon fragility: dense terrain reconstruction smooths action-critical details, and pointwise imitation lacks temporal credit assignment. Its Query Reconstructor (QR) uses Fourier-encoded cell queries to retrieve spatially specific evidence from depth-proprioception tokens, preserving sharp terrain boundaries. Trajectory-Aware MSE (TA-MSE) Distillation adds next-state teacher-student disagreement to the PPO reward, enabling Generalized Advantage Estimation to propagate future disagreement penalties to preceding actions. In simulation, QR reduces height-map L1 error by factors of 3.3-4.0, while TA-MSE surpasses PPO and MSE+PPO in curriculum progression. On stress-test terrains, SOLO achieves 97.5% mean traversal success and 96% stepping-stone success, versus 75.0-75.6% and 0-3% for dense-reconstructor variants. Deployed zero-shot with only a chest-mounted depth camera and proprioception, SOLO completes a continuous 1.5-km outdoor route and an indoor mixed-terrain course. Project page: https://sunpihai-up.github.io/solo/

World-Action Models (WAMs) build robot control on video-generation backbones, which jointly predict dense future visual trajectories and robot actions. We argue that video generation is an unnecessary intermediate objective for world-action modeling. For robotic manipulation, the goal of a world model is not to reproduce how the world looks at every intermediate moment, but to predict the state that the world will reach after an action is executed. The intermediate frames only describe the visual transition between physical states, which consumes substantial model capacity and computation, but do not directly specify the physical outcome that the robot action is intended to produce. In this paper, we propose DELE-w0.5, which infers robot actions from predicted future states without relying on video generation. Concretely, DELE-w0.5 infers the action sequence from its corresponding compact future latent state. The future latent state captures the action-relevant physical outcome of robot interaction and serves as an explicit bridge between world modeling and action generation. The core design principle of DELE-w0.5 is to model how the physical world changes under robot actions, rather than how its visual appearance evolves frame by frame. This formulation removes the high-dimensional visual redundancy introduced by dense video representations, and it therefore enables cheaper training and low-latency inference. Across 640 real-robot trials on four long-horizon manipulation tasks, our DELE-w0.5 achieves the best performance among all compared policies, attaining 62.5% overall full-task success and 81.3% macro ordered-stage progress. It outperforms the strongest baseline by 32.5 percentage points in full-task success and 20.1 percentage points in macro progress.

Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empirical exploration to drive recursive self-improvement, existing paradigms remain static and fall short of this capability. In this paper, we present AutoDesign, a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. To instantiate and evaluate this framework, we focus on the academic paper-to-poster generation task and introduce PosterBench, comprising a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance, increasing the average PosterBench Score from 54.99 to 67.39 (+12.4%). In a fully autonomous long-horizon loop, it executes 253 tool calls and 11 editing turns within 40 minutes for under $3, reaching average conference-poster quality in human evaluation. A system-blind human study further demonstrates that AutoDesign achieves the highest human preference among evaluated systems.

Video world models simulate future states conditioned on current observations and user actions. Recent systems have demonstrated impressive video consistency and action controllability over long sequences. However, fairly comparing these interactive models remains challenging. In practice, a human player typically evaluates a world model by pursuing long-horizon objectives through interaction. For example, a user may turn around 360 degrees to see whether the environment remains consistent, or walk into the water and inspect whether realistic water ripples are generated. The action sequence required to achieve the same objective may vary substantially between models, making fixed action-conditioned evaluation unsuitable for cross-model comparison. To address this, we employ multi-modal Agent Players to interact with world models toward specified long-horizon objectives. Building on this paradigm, we introduce PlayWorld, a benchmark providing 171 scenarios, each with a specified objective. To evaluate performance thoroughly, we assess models along four core dimensions: geometry consistency, interaction fidelity, out-of-sight evolution, and insight evolution. In addition, we incorporate basic ability metrics for video quality and controllability. Experiments across nine state-of-the-art world models reveal that current models remain unreliable on long-horizon interactive objectives, particularly in maintaining spatial consistency and persistent state evolution. Code and data are available at https://github.com/kxding/PlayWorld.

Autonomous agents are increasingly capable of improving models, systems, and other technical artifacts through long-horizon experimentation. To understand the current state of this capability, however, evaluation must go beyond final scores, which neither reveal where progress is gained or lost nor indicate whether accumulated experience improves later decisions. We therefore present a systematic evaluation of seven frontier models on 36 long-horizon tasks based on a new framework that uses rule-based metrics to characterize within-run behavior through Solution Framing, Execution, and Feedback Control and controlled comparisons to assess experience reuse within and across tasks. The results show that current agents operate more like engineering optimizers than fully autonomous researchers: they can formulate and implement practical solutions, but their performance varies substantially across runs, their strongest solutions mainly adapt or combine established techniques, and genuine methodological novelty remains rare. Detailed analysis reveals that observed performance is shaped by multiple factors, including distinct process bottlenecks behind similar final outcomes, experience reuse that can help or mislead subsequent decisions, and harness designs that affect performance stability. These findings suggest concrete directions for improving model training, inference-time strategies, experience management, and harness design.

We consider the problem of autonomously learning robot skills under a limited practice budget for sequential tasks. We propose an active skill learning algorithm, \emph{Deliberate Practice (DP)}, that computes a provably \emph{budget-optimal} allocation---practicing skills that maximize expected cumulative reward while being learnable within the budget. DP estimates both the time needed to master skills and the cumulative reward of the task plans that the skills unlock. Computing a budget-optimal allocation is challenging as it requires reasoning about combinatorially many skill plans over a large practice budget. Our key contribution is a bilinear program that can compute this exactly using off-the-shelf solvers. Through simulated and real-world experiments on long-horizon manipulation tasks, we show that our approach allows robots to optimally use limited practice time to acquire useful policies and improve long-horizon planning.

Long-horizon surgical robot manipulation is challenging because task rewards are sparse, while meaningful interaction changes occur at irregular intervals. Existing world-model agents typically imagine at primitive-step resolution, leaving variable-duration task progress implicit. Manually specified stages can provide intermediate structure, but their task specific boundaries are difficult to align with state-dependent interaction transitions. We propose S2-HWM, a Sparse Event-Structured Hierarchical World Model that learns sparse event evidence from primitive latent trajectories to coordinate an event-level manager and a primitive-step worker. The event evidence schedules manager goal updates, and each selected latent goal conditions the worker's primitive actions until the next update. The learned event evidence also forms variable-duration segments for an Event Transition Model (ETM), which predicts the next?boundary stochastic state, segment duration, and accumulated segment reward. Chaining these event-level predictions provides a variable-duration continuation beyond the primitive imagination horizon for manager learning, while the worker retains primitive-step actor-critic learning. On a SurRoL-based PegTransfer task, S2-HWM achieves a success rate of 98.7%, outperforming the flat GAS DreamerV3 baseline by 22.7 percentage points.

Long-horizon egocentric memory transforms continuous first-person video and audio into a searchable record of past experiences. We demonstrate two bottlenecks in existing systems: indices built from context-poor captions are unreliable for agentic search, while retrieval ignores a question's temporal intent. To address both bottlenecks, we introduce EgoCITE (Egocentric Context-augmented Indexing and Time-aware Evidence retrieval), a long-horizon agentic memory framework for egocentric QA. EgoCITE comprises three components. EgoScheme uses local multimodal context to turn fragmentary video captions and speech transcripts into self-contained atomic memory indices. EgoIndex organizes complementary action, activity, utterance, and conversation representations into searchable multi-view memory indices at multiple granularities. EgoRetrv combines semantic search with question-conditioned temporal relevance scoring and curation of retrieved evidence. We evaluate EgoCITE on EgoLifeQA, EgoMem, and EgoR1-Bench in terms of answer accuracy and target-event retrieval alignment. EgoCITE improves accuracy over agentic memory baselines by at least 4.4--14.2% while achieving 36$\times$ lower cost than long-context LLM agents.

Strategic reasoning in Large Language Models (LLMs) within long-horizon environments is often limited by inconsistent subgoals. In these settings, finite attention resources prevent the model from maintaining strategic coherence over thousands of steps. This limitation leads to strategic drift, where localized decisions fail to sustain a coherent trajectory across reasoning. To address this, we introduce EpicStar, a framework that enables agents to learn memory as policy to tackle long-horizon reasoning. Specifically, the agent maintains a bank of successful past episodes as a heuristic alongside a working memory to track short-term environmental changes. During inference, a dynamic gating mechanism determines whether to execute a retrieved action directly or to perform new reasoning through a contextual fusion of the retrieved episodes and current working memory. Utilizing StarCraft II as the testbed, we evaluated EpicStar against diverse opponent styles. It significantly outperforms baseline methods, achieving higher win rates while consuming an order of magnitude fewer tokens, and it maintains this advantage consistently across difficulty levels and opponent strategies. Our findings provide compelling evidence that structured cross-episode memory is essential for enabling LLM agents to perform robust, long-term strategic execution in dynamic, autonomous settings.

Long-term agent memory is usually treated as select--store--retrieve, but retrieval does not decide whether contradictory, superseded, retracted, deleted, or stale records may support an outgoing claim. We introduce Governed Persistent Memory (GPM), an auditable bitemporal state-transition model with source-bound admission, derived lifecycle state, current public barriers, and fail-closed structured release. Five executable clauses cover ledger integrity, source binding, conflict isolation, non-revival after retraction or deletion, and exact claim closure over a fresh view at one verified head. On a prespecified hash-frozen 3,600-case GPM-ReleaseBench, GPM matches all complete outcomes; the strongest of three intentionally simple complete policies matches 1,800/3,600 and makes unmatched releases on 50% of violation cases. A separate sealed end-to-end service evaluation exercises real ingestion and release across eight query families. In its publicly disclosed V3 arm, the governed lane is correct on 2,400/2,400 clusters versus 600/2,400 for ungoverned local Qwen2.5-7B; it repairs all 1,800 baseline failures with no regression (one-sided 95% lower bounds 99.875% and 99.834%). A later V5 reseal over Chinese- and English-command arms, with generation-date pinning and no post-freeze reducer amendment, again obtains 2,400/2,400 per arm. A production-code-independent finite model explores 331,776 semantic and 1,990,656 query states without a full-contract counterexample, and a 100,000-trace three-engine differential yields zero mismatches. These are bounded contract and implementation results, not open-world model accuracy or evidence of world truth. Governed answers in the sealed service evaluation are deterministic service outputs; the 7B result is the ungoverned comparison, not a claim that a language model itself became perfectly accurate.

Large language model agents increasingly rely on long-horizon reasoning to solve complex tasks involving planning, tool use, and memory. A critical capability in such settings is reflection: assessing trajectory progress, identifying missing evidence and unreliable intermediate states, and deciding whether to continue, revise, or abandon the current branch. Learning effective reflection, however, is challenging because reflection is performed locally within the current branch, whereas its utility can only be determined by its contribution to the final trajectory outcome. This local-global mismatch makes outcome-based reinforcement learning provide only local, sparse and delayed supervision for reflective decisions. To solve these, we propose LoongReflect, a training framework that formulates reflection as a memory-control policy. The agent operates over a reversible trajectory tree using explicit reflect and backtrack actions. Reflection consolidates verified facts, missing evidence, and branch-specific risks into working memory, while backtracking removes an unreliable branch from the active context and preserves a concise corrective lesson. To learn this policy, LoongReflect combines two complementary signals through a look-ahead, extragradient-style coordination mechanism. A fast channel distills globally informed reflective behavior from a privileged teacher, with supervision restricted to reflection and backtracking tokens. A slow channel optimizes complete trajectories using outcome-based GRPO, aligning local control decisions with final task success. Experiments on multi-hop retrieval-augmented generation and mathematical reasoning benchmarks demonstrate consistent improvements over outcome-only reinforcement learning and self-distillation baselines.

AI agents are increasingly used in mathematics research, but it is often unclear how to use them effectively. Towards this, we present an extensive case study of how AI was used to improve bounds on the Grothendieck constant $K_G$, which captures the hardness between combinatorial problems and their continuous relaxations. Specifically, while the precise value of $K_G$ is not known, we recently tightened the best known bounds to $$\frac{6π}{11} \;\le\; K_G \;\le\; \fracπ{2\log(1+\sqrt2)} - 10^{-4}.$$ Crucially, these improvements were achieved using an AI research system that could arrive at insights deemed novel by domain experts. We give a detailed discussion of our experience using AI for mathematics research, particularly touching upon its strengths and weaknesses, as well as our experience with creating ideal conditions for AI to arrive at breakthrough insights.

Large language model (LLM) agents are increasingly deployed as personal assistants. Existing evaluations, however, mostly use short, self-contained requests in static environments. Everyday life assistance is different. A task runs for weeks rather than minutes. The world keeps changing while the agent is not being prompted. Many constraints are never stated outright. An agent that merely answers the request in front of it will fail at such a task. What is needed instead is an agent that stays proactive and consistent. It decides on its own when to act, when to ask, and when to stay silent. It notices changes that nobody announced. It keeps one plan coherent from the first day to the last. No current benchmark measures this. We introduce VibeLifeBench, a benchmark of 200 long-horizon tasks across ten everyday-life domains. Each task is a scripted multi-week timeline in a simulated world of 22 mock services. The world advances on its own clock, and many of its changes are silent, so only an agent that re-inspects the world discovers them. Every task is graded by fine-grained, weighted checks that read only what the agent actually left behind, covering the end state, the timeliness of its actions, and whether it upheld the implicit constraints. We evaluate seven frontier models. All of them score low, which shows how far current agents are from assisting with real life. We will open-source all tasks, environments, and the evaluation framework.

We study the approximation of dynamical systems by semi-autonomous neural ordinary differential equations (SA-NODEs) over long time horizons. For a single network trained on the whole horizon, the available error bound deteriorates double exponentially in the horizon length. We develop two training strategies that avoid this barrier, each built on a reset of the state. The model predictive strategy partitions the horizon adaptively and restarts every window from observed data: when training meets a prescribed tolerance on every window, the composite model meets it uniformly in time, with a parameter budget linear in the horizon for targets with a bounded, uniformly regular reachable tube. The Floquet strategy addresses autonomous targets with a stable limit cycle and uses no data at deployment: a certified contraction of the learned return map confines the error to linear growth in the number of elapsed periods. For the time-periodic architecture we deploy, the scalar certificate degenerates; we prove instead a uniform-in-time orbital guarantee whose hypotheses are measured on the trained model, and an obstruction showing that, for an exactly periodic learned field, small one-period error and a contracting stroboscopic map cannot hold at once. Numerical experiments on four benchmarks confirm the predicted error laws and measure the hypotheses of every guarantee.

Large language model (LLM)-based search agents answer questions through multi-step interactions with external environments. However, providing complete execution trajectories to the LLM causes unbounded context growth and introduces noise. Existing compression methods reduce context at the cost of important details and often replace erroneous facts without repairing downstream reasoning derived from them. To address this problem, we propose ReTree, a self-correcting tree-structured memory mechanism for search agents. ReTree constructs a bounded per-step reasoning context while preserving source-linked evidence. It models search as an evidence tree whose nodes store bounded summaries, evidence, and revision histories. When newly retrieved evidence contradicts an earlier claim, ReTree traces back to the node where the claim was introduced, replaces outdated evidence, regenerates summaries, prunes affected branches, and resumes search. Source-grounded evidence provenance supports reliable conflict localization and keeps final claims traceable to retrieved passages. Experiments on four public question-answering and search benchmarks show that ReTree consistently outperforms Full-Trajectory ReAct, improving answer accuracy by up to 25.6 percentage points (pp); the average maximum per-step reasoning context of Full-Trajectory ReAct is $1.27$--$1.51\times$ that of ReTree. These results establish ReTree as an effective self-correcting memory abstraction for long-horizon search.

Long-horizon service robots require persistent world models that can be built autonomously in unseen environments and revised as task-relevant objects change. Existing methods rely on online mapping, which accumulates localization and observation errors, static scene representations that cannot capture persistent object changes, or holistic vision-language predictions that lack verifiable 3D geometric evidence. We present PBD-AG, a persistent baseline-delta active graph framework that decouples robot-verified stable fixtures from revisable dynamic object events. Under our framework, the robot autonomously bootstraps the structural baseline from onboard exploration and inspects discovered fixtures to ground hierarchical object beliefs. PBD-AG maintains reliability-weighted object states over geometry, semantics, identity, existence, and support relations, utilizing a geometric visibility gate to mitigate false deletions under occlusion. Inspection viewpoints are selected by a graph-conditioned policy that balances target coverage, travel cost, collision risk, and redundant observation. Simulation experiments in multiple environments and under controlled dynamic evaluation show higher aggregate coarse-fixture F1 than capability-matched controls, as well as stronger identity continuity and event recall. A qualitative physical-robot demonstration further illustrates integration with onboard sensing, providing a traceable world model for long-horizon robotic perception. The project page of PBD-AG is available at https://shuobao214.github.io/PBD-AG/

Long-horizon tool-using agents must reason over user goals, domain policies, tool calls, simulator state, and delayed verifiable rewards. Reinforcement learning (RL) is a natural fit for this setting, but multi-turn on-policy rollouts create long contexts, while model-specific attention layers may require custom masks and learned sink normalization. We present SINKFLEX-RL, a modular training system for RL in dual-control tool-use environments. The system combines a Gymnasium-compatible environment wrapper, a VERL-style rollout dataflow, group-relative policy optimization without a separate value model, and a sink-aware FlexAttention path designed to preserve model-specific sink scaling under causal and sliding-window masks. In a preliminary Tau2Bench retail run, validation reward (mean@1) rises from 0.25 early in training to $0.44$ later in the observed training window, while training-score and trajectory-reward proxies also trend upward. In a fixed-configuration memory benchmark, the optimized attention path reduces peak VRAM from 28.06GB to 22.52GB at 4096 tokens, a $19.7\%$ reduction, and runs the measured 8192-token configuration using $25.53$~GB where the eager baseline runs out of memory. These results illustrate the value of integrating environment interfaces, RL dataflow, and attention-kernel design for memory-feasible long-horizon agent training.

Long-horizon agents accumulate trajectories spanning hundreds of interleaved reasoning, action, and observation steps, where answering a query may depend on evidence buried far back in the history. External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. Existing multi-memory systems either read a fixed set of structures for every query, inflating context and introducing noise, or route each query to a single structure, preventing the composition of complementary evidence. A controlled analysis on AMA-Bench shows that the optimal memory configuration is typically neither a single structure nor the full union, but a tailored composition of multiple structural memories that varies with query and task demands. Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. We propose MESA (a Multi-structure Evidence Selection framework for long-horizon Agent), which builds five complementary structure views of each trajectory and learns from end-to-end answer-level feedback to select and fuse a query-specific subset for a frozen answer model. To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.

Long-horizon complex tasks require agents to repeatedly observe states, formulate plans, invoke tools, verify results, and recover from failures in continuously changing environments. However, such control experience often remains confined to a single context or a fixed prompt, and is difficult to accumulate and reuse across historical traces. This paper presents OpenLoopEvolve (OLE), a self-evolution framework centered on the Loop Policy. OLE represents an agent's observation, planning, memory, action, verification, recovery, stopping, and budget control as portable policy assets with versions and lineages, and provides online and offline evolution modes that can be selected according to practical needs: the online mode triggers candidate generation based on feedback from continuous operation, whereas the offline mode searches for candidate policies from archived traces and failure evidence. Both modes share an evolution mechanism consisting of autonomous proposals by a large language model, Champion--Challenger paired evaluation, and robust release. Policies released online are activated at a subsequent task boundary, monitored using subsequent feedback, and rolled back to their parent versions when degradation conditions are met. On the simulated business benchmark YC-Bench, both modes improve aggregate task performance, task success rate, and risk metrics relative to a fixed initial Loop Policy. The results indicate that treating the Loop Policy as a governable asset can support the accumulation, comparison, release, and reuse of control experience and improve agent performance on long-horizon complex tasks.

Electrical circuit analysis requires more than recognizing components in an image. A solver must ground symbols and labels, recover latent topology, select a physical model, formulate coupled equations, propagate intermediate quantities, and preserve units, signs, directions, and phase conventions. We introduce \benchmark, a benchmark of 1,000 authentic textbook problems for evaluating this complete long-horizon visual-to-symbolic reasoning process. Each problem pairs one or more circuit diagrams with a self-contained question, a typed or semantically specified answer, and a reference worked solution. An evidence-first construction pipeline aligns questions, figures, and solutions, while a reasoning-oriented taxonomy organizes problems by circuit type and dependency depth. Evaluation combines conservative typed scoring with identity-blinded multi-model semantic consensus, retaining every problem in the denominator. Across three commercial chatbot systems and six open-source multimodal large language models, the highest-scoring system reaches 84.8% accuracy. However, performance consistently deteriorates on long-horizon problems, and qualitative analysis exposes persistent failures in topology-to-target binding, physical conventions, and late-stage output propagation. \benchmark{} provides a focused testbed for measuring whether multimodal models can transform technical visual evidence into sustained, physically valid symbolic reasoning. Code are available at GitHub - CircuitReason/CircuitReason1K.

Tool-use LLMs are increasingly asked to act on users' behalf, but existing benchmarks usually focus on profile recall, style imitation, generic tool use, or response-level personalization. We introduce UserToolBench , a benchmark for personalized decision making in tool-use LLMs. UserToolBench tests whether a model can infer latent user preferences from interaction history, recognize when clarification is needed, and produce user-aligned tool-call trajectories under incomplete information. The benchmark is built from privacy-sanitized real interaction traces and combines structured persona profiles, public API-style tool ecosystems, and long-horizon multi-turn trajectories. It includes 10 user profiles, 36 tool sets, 1,065 turns, 170 unique tools, and evaluation-focused task types covering lack-of-information, single-tool, and multi-tool settings. Experiments with strong tool-use LLMs show that current models still have difficulty with personalized delegation. Multi-tool coordination, missing-constraint inference, and long-horizon behavioral consistency remain major bottlenecks. These results suggest that personalization evaluation should move beyond asking whether outputs sound user-specific and instead ask whether LLMs make correct decisions for the users they represent.

Long-horizon multi-part object disassembly requires robots to compute feasible sequences of collision-free removal motions, even in the presence of tight, narrow escape corridors. To efficiently solve such disassembly problems, we propose Parallel Extraction for Long-Horizon Disassembly (PEEL), an algorithm which efficiently computes disassembly motions for object assemblies and feeds them to a robot manipulator for execution. PEEL uses sampling-based motion planning to compute single-object motions through the use of a scale-invariant sampling scheme, where the object scale is estimated in a burn-in phase and a subsequent directional sampler exploits the scale. This sampling scheme is integrated into a multi-arm bandit rapidly-exploring random tree (MAB-RRT) planner, which switches between different samplers depending on the reward signal received. Using MAB-RRT, the PEEL algorithm runs a batch of planners in parallel to obtain an ordered graph specifying the sequence in which object parts have to be removed. We show that MAB-RRT can efficiently solve single-part disassemblies with 100 percent success rate on 76 assemblies, and that it is robust to its parameters. By integrating MAB-RRT into PEEL, we solve four long-horizon disassembly problems using the Fetch manipulator robot involving 10 to 17 individual object parts.

Long-horizon research agents solve open-ended tasks through iterative retrieval, aggregation, and synthesis, but context grows rapidly while the marginal value of additional evidence often declines. This leads to unnecessary token cost, higher latency, and noisier inputs for final report generation. We study marginal value estimation for context management in deep research agents and present the first systematic stage-aware comparison of pruning strategies across the pipeline. We evaluate lightweight heuristic criteria and a learned value model at pre-retrieval, post-retrieval, and pre-synthesis stages. Our results show that pruning effectiveness depends more on where pruning is applied than on the specific scoring rule: early pruning yields the largest end-to-end savings, while later pruning mainly refines the final synthesis context. Lightweight heuristics reduce token usage by up to 73% with little quality degradation, learned pruning remains competitive on selected trade-offs, and no single method dominates across quality, efficiency, and faithfulness. These findings provide practical guidance for designing efficient long-horizon agentic systems.

Ultra-long egocentric video understanding requires reasoning over temporally sparse evidence distributed across hours or days, challenging current multimodal models with limited context and the grounding of key video segments. While Chain-of-Tool-Thought (CoTT) agent systems enable iterative retrieval and inspection, they suffer from error propagation due to rigid zoom-in strategies that lack recovery mechanisms. In this work, we address these challenges through SCOUT (Self-Checking Chain-Of-Tool-thought), a recovery-aware agentic framework introducing an adaptive policy that evaluates intermediate tool observations and dynamically trades off exploitation (zoom-in) and exploration (region switching), enabling robust multi-hop reasoning over extremely long horizons. However, training such multi-turn tool-using agents remains challenging, as existing RL methods rely on sparse outcome-level rewards and lack supervision over extended decision trajectories, resulting in suboptimal credit assignment for long-horizon reasoning. To address this, we develop UPS-GRPO, an uncertainty-prioritized policy optimization method that concentrates exploration on high-uncertainty post-tool states while preserving sample efficiency. We further introduce a turn-level advantage decomposition that integrates outcome rewards with tool-grounded temporal alignment rewards for improved credit assignment. Experiments show that SCOUT achieves state-of-the-art results on ultra-long egocentric benchmarks, while remaining competitive on shorter-horizon long-video settings.

Embodied agents using LLM-based planners often struggle with physical hallucinations, poor generalization to long-horizon tasks, and lack of environmental awareness. We propose GraphThink, a novel framework that integrates a task graph to provide structured knowledge for robust planning and a scene graph to maintain environmental memory for event-driven replanning. Specifically, the task graph guides LLM thinking through contextual prompting and iterative refinement, effectively mitigating planning hallucinations. Furthermore, within the GRPO framework, the task graph offers delicate reward design to train the LLM planner, enhancing long-horizon planning capabilities and improving generalization. Finally, an event-driven replanning module, powered by the scene graph, enables closed-loop environment awareness and error correction. GraphThink achieves state-of-the-art performance on the ALFRED benchmark. In particular, our high-level planner surpasses leading API-based LLMs on both the validation set and held-out long-horizon tasks, underscoring its robust zero-shot and few-shot capabilities. Additional evaluations further demonstrate strong out-of-distribution generalization to novel tasks and environments.

Long-horizon humanoid loco-manipulation requires composing versatile whole-body skills and reliable high-level decision making. Existing methods often coordinate pretrained skills with scripted planners, finite-state machines or task-specific model-free policies, restricting their ability to handle complex task sequences. To address this limitation, we propose \textbf{LUCID}, a hierarchical model-based reinforcement learning framework that plans over reusable skills through imagined rollouts of a learned dynamics model. LUCID first trains a structured latent-conditioned low-level policy via adversarial imitation and then freezes it while jointly learning a high-level policy and macro-dynamics world model. The world model predicts the temporally extended state transitions induced by latent decisions, enabling high-level policy optimization through imagined rollouts. We evaluate our framework across various simulated multi-object rearrangement scenarios. Experimental results show that LUCID improves the full-task success and partial-completion rates compared to prior baseline methods, demonstrating its effectiveness in complex sequential loco-manipulation tasks.

World models are expected to support imagination over extended temporal horizons, yet most are still trained through local few-step prediction objectives and deployed by recursively rolling out their own predictions. This creates a fundamental mismatch: few-step losses optimize local transition fidelity, while long-horizon prediction depends on how errors and gradients propagate through the entire trajectory. As a result, transitions with different downstream influence on the endpoint are treated uniformly during training, and small local errors are amplified through recursive inference. We argue that long-horizon accuracy is better achieved by optimizing directly, through an end-to-end endpoint prediction objective. To instantiate this paradigm, we introduce the Direct Prediction World Model (DPWM), a non-recursive architecture that compresses an action sequence of arbitrary length into a single embedding and predicts the endpoint observation in a single forward pass. This design avoids recurrent rollout in both prediction and gradient propagation, making long-horizon end-to-end training practical at horizons where unrolled autoregressive training becomes unstable. Empirically, DPWM substantially improves long-horizon endpoint prediction over recursive world-model baselines on continuous-control and pixel-based benchmarks, with larger gains as the prediction horizon increases. We further show that recurrent baselines benefit similarly when retrained with the same long-horizon endpoint objective, supporting our central claim that the training objective, rather than the particular backbone choice, is the main driver of long-horizon prediction accuracy. Our results suggest that world models can benefit from being trained and evaluated at the temporal scales where they are ultimately used, shifting the focus from local transition modeling toward long-horizon predictive accuracy.

Long-horizon agents accumulate growing contexts during interaction, impairing performance and stability. Compact memory mitigates this problem by compressing and rewriting the history retained between model invocations. Learning what to retain typically relies on proximal policy optimization (PPO) with final task rewards, but sparse rewards provide little guidance for individual memory updates. This limitation motivates on-policy distillation (OPD), which supplies dense teacher supervision on student rollouts. For such supervision to be valid, the teacher must evaluate each sampled action under the same state in which it was generated. However, the context rewriting performed during memory compression can break this alignment. When sampled responses are retained and re-encoded for later invocations, flattening the interaction into a persistent history may cause the teacher to score the action under a state that the student never visited during rollout. The action therefore remains on-policy by provenance, but not necessarily by state. We therefore propose Memory-Aligned On-Policy Distillation (MemOPD). MemOPD records the inputs and sampled outputs of each model invocation, restores its original token positions and causal visibility, and packs the reconstructed invocations for efficient teacher scoring. The teacher provides full-vocabulary supervision at the sampled action positions, while PPO preserves the final task objective. Experiments verify state alignment across several context updates and show that it improves F1 by 7.0% over persistent-history teacher scoring in a matched control. Overall, MemOPD-3B improves F1 over PPO by up to 416.2%, while packing yields up to a 1.63x speedup in actor computation during training. The code for this work is publicly available at: https://github.com/TPssp/MemOPD.

Horizon detection in images of ice-covered waters is a challenging problem for maritime navigation due to low contrast between water and sky, cluttered ice structures, and varying illumination conditions. This paper presents a comparative evaluation of six horizon detection algorithms, including four classical computer vision methods and two hybrid approaches combining deep learning with classical line detection. A new bespoke IceHorizon dataset consisting of 30 ship-based and 8 drone-based videos is used to evaluate detection accuracy, horizon coverage, and computational performance. The results show that hybrid methods achieve the highest accuracy and most reliable horizon estimates. In contrast, purely classical methods exhibit reduced robustness, particularly in visually ambiguous scenes. Performance on ship-based imagery was consistently higher than on drone-based imagery, indicating a strong dependency on acquisition characteristics. The created dataset and codes used in this study are made publicly available to support further research on this topic. The code is available at https://github.com/allythe/HorizonDetection. The dataset is available at https://doi.org/10.5281/zenodo.20411867

Long-horizon role-playing demands that characters remain recognizable as they evolve with the narrative. Yet existing work falls short on two fronts: representations are typically static profiles that cannot be updated locally without destabilizing unchanged traits, and benchmarks mainly test persona preservation and memory recall rather than whether a model speaks from a character's currently evolved state. We address both. PHASE-Tree is a multi-timescale character-state tree with an immutable identity root and mutable persona, session, and moment layers, making each mutable field an addressable target for localized within- and cross-episode updates. It conditions generation through explicit textual provision or implicit parametric adaptation. To measure evolved-state generation, we introduce LongEvoRoleBench, which pairs four long-dialogue corpora for cross-episode evolution with four short-dialogue corpora as within-scene state-tracking checks, under a unified next-utterance protocol. On the long-dialogue core, textual PHASE-Tree ranks first in 11 of 12 dataset-metric cells against internal variants and all 12 cells against external textual baselines, improving character-level, semantic, and embedding scores by 19.7%, 12.4%, and 15.1% respectively. In a blinded 200-response study, human ratings correlate with the GPT-4.1 judge (Pearson r= 0.65); on descriptive n= 10 PT and NR prompt subsets, the Overall difference is +0.20. The long-dialogue Sem advantage persists across LLM judges and generation backbones.

Long-horizon agents rely on memory to reuse experiences, yet existing memory systems often assume that evidence can be directly consumed through a fixed representation. This leads to representation mismatch, where relevant information is available but not organized for the current decision. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and render transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.

Lifelong Multi-Agent Path Finding (LMAPF) requires generating collision-free paths for large agent fleets under strict real-time constraints. Reactive frameworks such as PIBT and Enhanced PIBT (EPIBT) scale effortlessly to thousands of agents through rule-based, step-by-step coordination but suffer from severe temporal myopia, making them ineffective in scenarios where long-horizon reasoning is essential. RHCR plans windowed paths over multi-step horizons but incurs substantial planning overheads that hinder scalability. TP tackles both challenges by planning only subsets of agents at each timestep, yet its applicability is restricted to highly structured maps. To achieve long-horizon planning at scale across general maps, we propose Path Updates over Staggered Horizons (PUSH), a LMAPF planner capable of coordinating thousands of agents in under a second while planning over multi-step horizons. PUSH combines the key advantages of PIBT, RHCR, and TP. Like TP, PUSH reduces computational complexity by planning only a subset of agents at each timestep using staggered planning windows. Unlike TP, however, PUSH plans RHCR-style windowed paths in general maps without relying on restrictive map assumptions. To maintain high throughput in congested environments, PUSH further integrates EPIBT-inspired priority inheritance, backtracking, and anytime improvements into its windowed planning. Empirical evaluations across two realistic MAPF scenarios requiring long-horizon reasoning show that PUSH scales to the same massive agent loads as EPIBT (e.g., 10k agents) while achieving significantly higher system throughput than all baselines.

Frontier language models solve reasoning problems in a single forward pass that would have been research contributions years ago, yet fail at multi-hour tasks: losing track of earlier decisions, declaring half-finished work done, or drifting from goals. We call this the horizon gap and survey 1,547 arXiv papers (2024-2026) collected via systematic seed harvest with a disclosed 26.8% bleed filter, extended by targeted supplementation. We disambiguate three routinely conflated properties: long-horizon (task property: required steps), long-context (model property: token capacity), and long-term memory (system property: persistence across steps/sessions). We organize the corpus into six categories tracking a long-horizon task's lifecycle -- planning, memory, execution, training, evaluation, and foundations/safety -- crossed with an axis capturing where horizons are carried (within-context, within-task-beyond-context, or cross-task-persistent). Across all categories, we find the same pattern: outcome-only signals grow uninformative as horizons lengthen, and the field's response -- whether process reward models, credit assignment, or trajectory-level diagnostics -- manufactures denser step-level signals. We treat critical and diagnostic literature as first-class threads throughout, arguing that segregating critique from method would routinely split single papers across chapters. We close by naming open measurement problems: decomposing model versus harness capability, managing correlated bias in process-level signals used for both training and evaluation, and whether long-horizon reliability admits general predictive theory.

Recurrent context compression controls context growth in long-horizon agents, but its behavioral effects remain poorly understood. In this preliminary empirical study, we show that compression can weaken the influence of recent interactions, increasing blocked actions, repeated exploration, and instability across runs. Motivated by these observations, we introduce TRACE, a verifier-guided framework that evaluates individual compaction events through paired closed-loop continuations from the same environment state and uses summary preferences to optimize a natural-language compression prompt while keeping all models frozen. Initial results on AppWorld show improvements over existing compression baselines in task performance, multi-run reliability, and context--execution efficiency. These findings provide early evidence for boundary-local evaluation as a promising direction for reliable agent context compression.

LLM-based agentic systems have shown remarkable capabilities in complex domains, while suffering from cascading errors and difficulty in debugging. Critical error detection aims to locate the earliest error step in a failed trajectory that is responsible for the final failure. However, progress faces two main challenges. First, long trajectories make it difficult to identify individual errors, since the evidence for judging a step may be scattered across distant instructions, observations, and prior context. Second, failed trajectories often contain multiple local errors with different downstream effects, only some of which remain responsible for the final failure. In this work, we propose TrajDebug, an error-lifecycle tracing framework that addresses long-trajectory error discovery with multi-granularity history compression and evidence-based error identification, and supports critical attribution by tracing each error's resolution status and terminal impact. We further construct TrajErrBench, a benchmark of 486 manually annotated failed trajectories from Tau2Bench and SWE-Bench Pro, covering realistic tool-use and coding scenarios. Experiments across diverse agent benchmarks show that TrajDebug achieves the best overall performance over existing baselines, and application studies further demonstrate that its diagnoses provide actionable feedback for improving downstream agent success. We will release the codes and data to facilitate further research.

Despite advances in artificial intelligence (AI) across multiple sectors, today's AI tools, including deep learning and generative AI, still fail when embedded into physical systems, such as robots and vehicles operating under real-world physical laws. This stems from their inability to maintain reliable world models for long-horizon planning under uncertainty and generalize to unseen scenarios. In this context, wireless networks, through pervasive sensing and communication, can orchestrate physical intelligence. However, current architectures optimize throughput, latency, and reliability and cannot support real-time physical AI coordination, requiring agents to maintain shared spatiotemporal context. To address these challenges, a network of holonic digital twins (HDT-Nets) framework is proposed to deliver real-time physical AI inference through holonic agents that actively reason about their environment rather than passively mirror physical assets. Each HDT is realized as a hierarchical structure spanning the physical agent and network edge, reasoning autonomously at the local level while cooperating with neighboring HDTs to form collectively intelligent units. In HDT-Net, causal Markov blankets spanning sensing, communication, and control determine which agents must coordinate and enable counterfactual reasoning over multi-domain interventions. Active inference within these boundaries unifies perception, action, and learning by minimizing expected free energy while deciding which beliefs to transmit based on their cognitive value to the receiver. Category theory ensures that transmitted beliefs preserve semantic structure across heterogeneous agents with incompatible representations. Finally, integrated information theory quantifies when collective intelligence exceeds independent operation and how network intelligence evolves through coordinated learning and information exchange.

Vision-language-action (VLA) models have demonstrated remarkable capabilities in robotic manipulation by leveraging pretrained vision-language models. However, existing post-training methods predominantly optimize VLA models as flat policies, making it difficult to explicitly model task progression and perform robust long-horizon manipulation. Although hierarchical approaches introduce task decomposition, they mainly rely on supervised learning from offline demonstrations and cannot effectively improve execution through online interaction. To address this limitation, we propose Hierarchical Robotic Control (HiRoC), a hierarchical post-training framework that decouples high-level task planning from low-level action execution. The planner decomposes complex tasks into executable subgoals to provide explicit semantic guidance, while the executor continuously improves subgoal-conditioned action generation through reinforcement learning. To enable effective collaboration between the two modules, we further align the executor with planner-generated subgoals before reinforcement learning, mitigating the distribution misalignment between planning and execution. Extensive experiments across diverse robotic manipulation benchmarks demonstrate that HiRoC consistently outperforms strong baselines. Comprehensive analyses further validate the effectiveness of hierarchical post-training and the contribution of each key component.

Mobile GUI agents can operate apps through pixel perception and touch actions, making them a promising interface for collecting and improving long-horizon mobile interaction policies. However, real trajectories are difficult to obtain for sensitive apps and privacy-critical operations. At the same time, existing simulated environments are costly to scale up, and GUI world models still suffer from unstable generation, limited modality coverage, and inconsistent action-transition logic. To address these limitations, we propose AppDeltaWorld, a transition-grounded delta code world model that predicts the next GUI as a reachable code update rather than as an unconstrained image or text description. AppDeltaWorld retrieves app-specific Level-1 HTML references under an action-transition constraint, generates Level-2 executable HTML conditioned on the current screen, action, predicted next-screen text, and retrieved structure, and inserts generated visual assets into image slots before browser rendering. As a world model, AppDeltaWorld achieves the highest fidelity on CMGUIBench-500 under Code2World evaluation, with clear gains in structural layout and UI element reconstruction over image-only and code-only baselines. As a training environment, AppDeltaWorld supports filtered closed-loop SFT data construction that, when combined with public supervision, enables AppDeltaAgent to achieve state-of-the-art performance on AndroidLens and consistent gains on MobileGym and MobileWorld. Moreover, world-model-based test-time reinforcement learning enables policy adaptation and shows further improvements without additional interaction with real apps.

Task difficulty dictates an agent's likelihood of success, and estimating it without rollouts means forecasting this directly from a task description before executing costly simulations in stateful environments. Reliable estimates would therefore allow environment designers to calibrate evaluation benchmarks and construct progressive training curricula. This becomes increasingly important as agents move into long-horizon domains, where empirical trial-and-error is a severe computational bottleneck. Prior work on early prediction is limited to static tasks or isolated coding environments, often relying on narrow features and inaccurate evaluation metrics. We study \textit{ex ante} difficulty prediction across 17 agentic benchmarks spanning coding, mathematics, machine learning, web navigation, function calling, and other domains. We show that AUC can mask poor difficulty estimates, identify token-level entropy as a useful predictive signal, and show how residuals between expected and observed difficulty can expose hidden environment flaws such as contamination and infeasibility.

High-quality long-horizon training data for terminal agents is expensive to produce, often costing hundreds to thousands of dollars per task, because each task must keep the instruction, environment, reference solution, and verifier mutually consistent. Human authoring does not scale, and direct generation with large language models (LLMs) often breaks these dependencies. We present Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for constructing long-horizon terminal-agent tasks at scale. Starting from verified seed tasks, RST extends the reference solution, realigns the verifier and instruction to the new workflow, validates the result in a fresh sandbox, and reuses accepted tasks as seeds for subsequent rounds. Across fifteen recursive rounds, RST produces 37,484 synthesized terminal-agent tasks at roughly $0.05 per task. Task difficulty increases substantially over rounds: the median reference solution grows from 67 to 374 lines, the median number of executed commands grows from 40 to 244, and DeepSeek-V4-Pro pass@4 drops from 90% at R1 to 2.5% at R15. To demonstrate training utility, we collect rejection-sampled Qwen3.5 trajectories on the synthesized tasks and use them for supervised fine-tuning. Fine-tuning on these trajectories improves Qwen3.5-27B and Qwen3.5-122B-A10B by up to 10 points on Terminal-Bench 2, Terminal-Bench Hard, and Long-Horizon Terminal Bench, while agentic PPO lifts Qwen3.5-27B to 49.44%, 32.00%, and 22.07% on the three benchmarks, corresponding to relative gains of 20.0%, 41.2%, and 21.9% over the base model. Moreover, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable as difficulty keeps climbing, indicating that the process can continue well beyond the scale reported here.

Long-horizon LLM agents increasingly rely on external execution support to maintain state, track progress, invoke tools, verify outcomes, and reuse experience across interactions. However, effective harness use raises two coupled challenges: state formation from noisy interaction traces and runtime control over external-state access. Existing agents usually handle both through prompts, heuristics, or domain-specific conventions, leaving the external workspace and its usage policy manually engineered. To address this, we study the problem of harness policy learning, where agents learn harness policies offline and deploy them to construct and update external harness state online during runtime task execution. We introduce EvoHarness-RL, which exposes Belief, Progress, and Experience (BPE) as policy-facing harness state. Supervised harness fine-tuning teaches the base agent the harness action space and how to construct useful external state, while cost-aware GRPO explores coordination policies to selectively read, update, and consolidate that state during long-horizon interaction. Instantiated on ALFWorld with a Qwen3-8B LLM, EvoHarness-RL reaches 96.9% success and reveals two key dynamics: harness annealing, where training internalizes recurring harness-use patterns into the model policy and shifts the agent from frequent harness calls toward selective external-state access, and harness evolution, where progress updates and experience consolidation refine the harness into a compact, task-adaptive state substrate. These results suggest that long-horizon agents benefit from trainable policies for constructing and coordinating with external harness workspaces, beyond simply adding stronger tools or larger memories.

Long-horizon reasoning requires an agentic runtime that can persist when evidence supports its current approach and pivot when measurements reveal failure, hidden constraints, or a misspecified objective. We present Argus, a persistent, self-evolving runtime in which Manager, Planner, Engineer, and Reviewer execute bounded missions over durable project state. Argus separates stable user intent from operational objectives, constraints, and verification criteria, and admits memories, skills, procedures, verifiers, routing decisions, and rejected routes only after role-owned review and, when available, task-native verification. Model weights remain fixed; self-evolution occurs through persistent runtime state and control policy, with autonomous execution between operator-owned escalation points. Across seven GPT-5.5 benchmark arenas, Argus achieves about 78% on SWE-Bench Pro versus 59% for Direct Copilot while using 1.41 times the aggregate tokens. After verification-gated self-evolution, mature SWE-Bench waves use 21% fewer solve-input tokens and 15% less active workflow time per task than startup waves, while recording 34 verifier recoveries and 22 strict review-loop rescues. Argus also reaches 76.8% on AARRI-Bench and a 28.0-point gap on mathematical data synthesis, with competitive GPU-kernel and language-model-training results. Beyond benchmarks, an optimized RWKV6 kernel was merged upstream; a multi-day mathematics campaign retained falsified routes and proof-backed frontier updates; and six paper pipelines completed 254 missions with 16 stage rollbacks. These results show that a fixed-weight, self-evolving harness can revise, recover, and accumulate verified approaches while producing structured trajectories for future supervised and reinforcement learning.

Long-horizon reasoning in recent LLMs demands that the model switch between distinct skills inside a reasoning chain, such as first doing a math derivation, then using the result to plan a schedule. We call such problems cross-skill long-horizon tasks: multi-step tasks whose steps require different reasoning skills and depend on earlier outputs. Existing benchmarks often evaluate individual skills, lacking a principled way to measure how well a model switches between skills. We address this gap from both the evaluation and training sides. We introduce Skill Entropy, a measure of the difficulty of switching from one skill to another. We then propose Skill^2-Bench, a benchmark of cross-skill long-horizon tasks built over 558 skills across 9 verifiable and open-ended domains. Each task is assigned a task-level skill-entropy score and grouped into three difficulty levels. Evaluating 8 frontier and 4 open-source models on Skill^2-Bench reveals a skill-switching gap: accuracy decreases on higher-entropy tasks. We then turn skill entropy from a benchmark scale into a training signal. We propose Skill-Entropy RL, an RL framework where the model predicts not only the answer at each step but also the skill used to produce it. The reward combines step-level correctness with a skill-entropy reward that measures the alignment between the model-predicted skill sequence and the gold skill sequence. On Qwen3-4B-Instruct and Qwen3-1.7B, Skill-Entropy RL improves the Skill^2-Bench score from 34.4% to 68.4% and from 14.6% to 40.1%, respectively, outperforming competitive baselines. The same pipeline can be applied to off-the-shelf training data such as OpenR1-Math, indicating that skill entropy is a reusable training signal. Code available at: https://github.com/Gen-Verse/Skill-Entropy-RL

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. In this paper, we propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions. Specifically, given a potentially obscure query and its corresponding ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces back from the answer to recover intermediate clues required to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against these clues, converting sparse binary outcome supervision into dense step-level rewards. Based on these rewards, we develop ABC-SFT, which reweights the loss of each turn, and ABC-GRPO, which uses the step-level scores as rewards in GRPO. Building on this framework, we train ABSeeker based on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores further improve to 55.3% and 52.9%, respectively, significantly outperforming same-scale (4B) agents and even matching the performance of larger ones (approximately 30B). These results demonstrate the effectiveness of answer-backtracked step-level credit assignment for training long-horizon search agents.

Interactive video world models are essential for long-horizon planning and exploration, yet they suffer from compounding errors. Post-training methods such as reinforcement learning (RL) can improve these models, but they hit a verification bottleneck: for arbitrary action sequences, no ground-truth future state exists to measure long-term drift. Our key insight is that reversible action cycles make this verification possible: a sequence composed with its inverse must analytically return to the initial state, yielding annotation-free supervision on long-horizon correctness. Building on this, we introduce WorldCycle, a self-verifiable RL framework that constructs closed action cycles and their repeated executions from ordinary action sequences, and optimizes two complementary rewards: a spatial closure reward enforcing symmetry between mirrored forward and reverse segments, and a temporal consistency reward aligning states across repeated cycle executions. These rewards force the model to learn actions as consistent state operators rather than memorized temporal patterns, and extend naturally to out-of-distribution composite action cycles that the base model handles poorly. We further release CycleBench, a diagnostic benchmark for state-returning ability under complex action structures. WorldCycle reduces state returning drift by up to 44% and lifts composite-action accuracy nearly 4x over the base model, providing a vital foundation for physically grounded world models.

Vision-language-action (VLA) models provide a unified paradigm for connecting visual perception, language understanding, and robotic control. However, existing VLA models still face major challenges in long-horizon tasks: sparse expert demonstrations constrain cross-task compositional generalization; the non-Markovian nature of long-horizon tasks makes it difficult for policies conditioned only on current observations to maintain temporal consistency; limited closed-loop error correction allows execution errors to accumulate; and end-to-end action fine-tuning may weaken the high-level semantic representations of vision-language model (VLM) backbones. To address these issues, we propose a hierarchical long-horizon VLA architecture with an explicit language-memory module. The central idea is to convert discrete temporal observations into a coherent textual memory sequence with temporal logic. The system is decoupled into a high-level VLM and a low-level VLA: the high-level VLM performs semantic reasoning through a visual question answering training paradigm, while the low-level VLA executes precise continuous control conditioned on subtask instructions and visual observations. The high-level VLM recursively updates both language memory and subtask instructions using the previous memory as a contextual anchor, enabling persistent temporal tracking and dynamic correction during long-horizon execution. We evaluate the proposed method in multiple simulation environments and conduct sim-to-real experiments on a real robotic platform. The results demonstrate that explicit language memory improves the success rate and robustness of VLA models on complex long-horizon tasks while providing an interpretable semantic account of the decision process.

Subtask labels decompose a long-horizon manipulation demonstration into shorter semantic segments for policy training and evaluation. Natural language descriptions are easy to read, but their linguistic variability makes automatic verification difficult. Rigid template formats, such as BEHAVIOR-1K's skill_annotation, are linguistically over-segmented, hindering both readability and annotation consistency. We propose the Structured Subtask Chain (SSC), a state-transition representation that bridges these extremes. A demonstration is a sequence of Structured Subtask Template (SST) entries. Each SST stores core action components (subject, predicate, object), flexible conditions (adverbial modifiers such as spatial or instrumental phrases), a base-motion field separate from arm actions, and an after-state scene graph. Built on this format, SSC supports three vision-language assisted functions: rendering SSTs as natural language, checking the assembled chain against four state-transition rules, and completing underspecified fields through a query resolution cascade. We instantiate the pipeline on BEHAVIOR-1K (50 tasks, 3 episodes per task, 2,357 annotated action cells) for logic verification and content completion, evaluating 13 selected state-of-the-art VL models as candidate verifiers and reporting labelling anomalies.

Swap regret governs the rate at which uncoupled learning dynamics converge to correlated equilibria in multiplayer general-sum games. Under full-information feedback, the best previous guarantee when every player follows the same dynamics grows logarithmically in the horizon $T$. We construct uncoupled dynamics under which every player incurs only $O(nm^2\sqrt{\log m\log T})$ swap regret, where $n$ is the number of players and $m$ bounds the number of actions per player. To our knowledge, this is the first sublogarithmic individual guarantee in this setting, and it implies that the time-averaged product distribution of play is an $O(nm^2\sqrt{\log m\log T}/T)$-approximate correlated equilibrium. The key algorithmic choice is to combine the Blum--Mansour reduction with optimistic follow-the-regularized-leader using a hybrid regularizer that separately weights negative Shannon entropy and the log-barrier: the entropy controls the optimistic prediction error, whereas the log-barrier controls the transition-matrix movement through its Bregman divergence. A new sensitivity theorem for stationary distributions of Markov chains, which involves neither mixing parameters nor the smallest transition probability, transfers this control to the played strategies and yields a simpler analysis without local-norm or self-concordance arguments. The guarantee is preserved by an adversarially robust variant that additionally ensures $O(nm^2\sqrt{\log m\log T}+\sqrt{mT\log m})$ swap regret against arbitrary utility sequences, and by a horizon-free variant that requires no prior knowledge of $T$.

We study linear representations of temporal horizon in the large language model Qwen3-32B and use them to change the model's time-related preferences, recommendations, and capabilities. We train contrastive linear probes on teacher-forced temporal-choice answers to find a short-term versus long-term direction in the model's residual stream, and evaluate contrastive activation-addition steering on a held-out binary temporal-choice task, an out-of-distribution monetary intertemporal-choice task, and a TravelPlanner capability benchmark. The central result is that temporal-horizon directions can be identified with simple contrastive linear probes and then used for steering to induce large, bidirectional preference changes. On an out-of-distribution monetary choice task that varies reward size and delay, steering strongly shifts the model's indifference threshold between smaller-sooner and larger-later rewards in both directions. We further show improvements on a planning-related capability metric under moderate temporal steering. These results suggest that model intertemporal preferences are measurable and steerable, which is relevant for AI systems that give advice involving delayed costs and benefits, and for safety questions about long-horizon planning.

How do you verify a long-horizon agent when its own state and self-reports are exactly what you cannot trust? We present an agent instrument built so that verification is structural rather than post-hoc. A deterministic Executive owns all belief; a language model may only file typed proposals, and a claim is admitted only when a prediction pre-registered before acting is matched against observation by code. Two properties make the instrument a verifier of its own science, not just of the agent: every run invalidates itself when per-organ write-error, render-size, or salted-canary-echo floors are breached (four of the first eight architecture runs were invalidated, each localizing a real defect); and a render-invisible shadow reference compiles the plan the full system would have committed in every ablation cell, so drift metrics are defined even where the mechanism under test has been removed. Using this instrument we report a clean, single-variable result on a failure every long-horizon agent suffers: ablating the commitment mechanism flips goal-abandonment from 0.00 to 1.00 while binding error stays flat at 0.00 (three seeds per cell, up to 394 reference beats per run, every run gated valid). The binding channel, by contrast, does not reappear as per-beat drift when its repair is ablated -- because binding is code-owned, the failure class is structurally absorbed, its only residue appearing one layer upstream as a collapse in hypothesis formation. We report these under full disclosure that task efficacy is null (zero level completions across 52 gated runs on ARC-AGI-3), pre-registered as a structural defeater. The contribution is a verification methodology for agent development and the drift decomposition it makes measurable.

Learning long-horizon manipulation skills with reinforcement learning remains challenging due to the complexity of reward design, the limited guidance of sparse rewards, and the high cost of manual subtask annotation. Visual demonstrations can provide supervision for reward learning, but rewards learned from raw pixels can be brittle and sensitive to visual variation, background appearance, and robot motion. In this work, we propose GORDON, a graph-based object-centric reward learning framework that learns dense rewards from action-free video demonstrations. Each visual scene is represented as a graph of detected objects and spatial relations, and a graph neural network is trained in a self-supervised manner to embed these graphs into a task-aligned latent space. To align the representation with semantic task progress, we introduce an activity-aware weighted pooling mechanism that emphasizes task-relevant objects while masking robot-dominated motion. The dense reward is then computed as distances in the learned latent space of the current state to demonstrated goal configurations, providing a measure of task progress. In long-horizon tasks, the temporal profile of this reward reveals stage-wise object-state transitions, enabling automatic subtask discovery without manual segmentation. The discovered segments are then used to train subtask-specific rewards and specialized policies that are composed sequentially. Experiments on seven manipulation tasks on MAGICAL and ManiSkill3 benchmarks show that our object-centric reward improves reinforcement learning in short-horizon settings and enables successful policy learning in complex long-horizon tasks through automatic decomposition, achieving an average success rate of 74.4% across the long-horizon tasks (on average approximately +35 p.p. vs. best learned baseline and approximately +25 p.p. vs. oracle).

Persistent memory helps long-term agents retain knowledge, yet a single update error can repeatedly distort future retrieval and reasoning. Most existing systems reduce memory updating to a binary Write/Hold decision, which cannot distinguish whether new information should be added, ignored, used to revise an outdated belief, rejected as unreliable, or deferred for verification. These choices may share the same binary label while producing fundamentally different memory states. We introduce TARL, a memory state update framework that maps each statement to one of five executable actions. TARL identifies the affected memory, resolves its temporal scope, compares source reliability, and updates accepted, pending, and rejected ledgers. It is further trained by comparing the memory states produced by alternative update operations, encouraging the model to select the operation that leads to the correct result. We also introduce TARL-Mem, a benchmark with fine-grained action labels and next-state targets. Across in-domain, cross-source, temporal, counterfactual, and sequential evaluations, TARL improves action prediction and state recovery, reduces memory pollution, preserves conflicting evidence, and limits cumulative corruption.

Forecasting precise future motion of surrounding agents is essential for reliable autonomous vehicles. However, as the demand for longer prediction horizons increases, existing endpoint-completion or iterative-refine methods increasingly struggle with weak guidance and compounding errors. To tackle the long-horizon prediction challenge, we propose Pivot-Centric Trajectory Prediction (PCTP). By introducing pivots'' and focusing on predicting pivot points along extended trajectories, we divide the long-term prediction task into short-term sub-tasks at various scales. Specifically, PCTP decouples the long-term trajectory predicting process into two processes: pivot prediction and pivot-based trajectory refinement. The pivot prediction process aims to utilize global map context and agent-to-agent interactions to identify these pivot points'', while the pivot-based trajectory refinement process focuses on local map details and refines the short-term trajectory based on predicted ``pivot points''. Compared with existing methods, PCTP provides more intermediate guidance while reducing compounding errors. Moreover, PCTP is a flexible approach that can be integrated into most state-of-the-art trajectory prediction models. Experimental results show that PCTP improves the prediction accuracy of leading models on both Argoverse I and Argoverse II datasets with minimal impact on model size. Specifically, PCTP combined with QCNet outperforms all published ensemble-free methods on the Argoverse II leaderboard at submission.

Existing chunk-based Vision-Language-Action (VLA) models execute a fixed number of actions (i.e., execution horizon) before replanning, turning replanning into a task-agnostic periodic schedule that is independent of task progress. As a result, when no replanning boundary falls before a critical manipulation stage, it is executed from a stale chunk rather than a freshly replanned one. To address this limitation, we propose Bernoulli-Continuation Policy (BCP), a lightweight, plug-and-play framework for adaptive horizon execution that keeps the base VLA frozen. Given a fixed-length action chunk, its continuation head decomposes execution-horizon selection into a sequence of continue-or-replan decisions, which imposes an ordinal, prefix-sharing inductive bias over candidate horizons rather than treating them as independent classes. Since the optimal horizon for each chunk is not observable, we train this head with reinforcement learning from trajectory-level outcomes and introduce a Replanning-Efficiency Reward that jointly rewards task success and efficient VLA usage, discouraging the policy from collapsing to unnecessarily short horizons. On RoboTwin 2.0 with LingBot-VLA as the base policy, BCP improves the average success rate by +11.08% on 13 low-success tasks and from 89.88% to 93.94% (+4.06%) across all 50 tasks. Although trained only under the Clean setting, BCP generalizes to the Randomized setting, raising the average success rate by +4.06%. It also transfers to a different base policy $π_{0.5}$, achieving a better result on LIBERO (+1.7%) and, notably, on the harder LIBERO-PRO (+6.8%). On a real robot, BCP lifts success from 74% to 92% and from 44% to 84% on two manipulation tasks. Meanwhile, its negligible overhead, combined with higher success, makes BCP's overall runtime even lower than the fixed-horizon baselines.