Scaling offline goal-conditioned reinforcement learning (GCRL) to long-horizon tasks is difficult because (1) long-range value learning depends on shorter-range estimates that may still be inaccurate, and (2) max-based value backups can amplify overestimation through repeated propagation. We propose DCRL (Divide-and-Conquer RL), which recursively decomposes each trajectory segment into a balanced binary tree and trains the values from leaves to root. Each parent is therefore updated only after its children, using an exact factorization of the observed route rather than selecting among noisy alternatives. Since this objective learns values along demonstrated routes that are not necessarily optimal, DCRL jointly propagates values across trajectories to discover shorter routes. Thanks to the balanced binary tree, DCRL reduces worst-case bootstrap depth from linear to logarithmic, and this shorter dependency structure empirically corresponds to much slower error accumulation. Across diverse goal-reaching tasks, DCRL substantially outperforms prior flat offline GCRL methods, and on the five most challenging long-horizon OGBench tasks, it improves the best prior average score from 55 to 64, surpassing all flat and hierarchical baselines.
As LLM-based human simulators are increasingly used for policy, evaluation, and training, they must faithfully reproduce real behavioral patterns. While prior work has examined behavioral fidelity in survey responses and dialogue, longer-horizon real-world activity remains largely unexplored. We introduce a framework for evaluating behavioral fidelity in long-horizon activity simulations across temporal granularities and levels of analysis. As a case study, we collect a 43-hour multi-camera dataset of in-the-wild office activity and compare trace-derived conditioning mechanisms: persona descriptors, few-shot exemplars, and statistical transition and time-of-day priors. We find that behavioral fidelity is not uniform across metrics: statistical priors bring activity and sequence distributions closest to real behavior, yet over-fragment routines and suppress within-person variability. These findings motivate a more holistic evaluation that spans multiple metrics, temporal granularities, and levels of analysis.
Streaming 3D reconstruction from extremely long videos requires estimating camera motion and scene geometry online under bounded memory and computation. Early streaming models achieve causal, bounded-cost inference using finite context buffers or compact recurrent states, yet their estimates often deteriorate as sequences grow. Recent methods improve long-horizon stability by coupling short-range context with persistent or multi-level long-range memory. We pursue a different route: we keep the learned temporal state strictly local and formulate predictions whose targets remain independent of sequence length. We present ABot-Recon, a simple streaming model that caches KV features from only the preceding 11 frames. It predicts a point map in the current camera coordinate system together with an adjacent-frame relative pose. These predictions remain equivariant under changes of reference frame, and global poses and geometry are recovered through sequential composition. To reduce accumulated drift, a lightweight temporal refiner improves relative rotations using recent visual and motion context, while a composition-aware pose loss supervises multi-step pose composition. Extensive evaluations on challenging long-sequence benchmarks demonstrate the superior long-horizon performance of our local-context approach. On Oxford Spires, ABot-Recon achieves an ATE of 4.35 m and an RPE-R of $0.12^\circ$, reducing both errors by approximately 40\% relative to the best prior results.
Long-horizon GUI agents can retain complete action histories as compact text, but only a few historical screenshots fit in active context. We formulate this as budgeted fidelity restoration: every event remains summarized, while a fixed budget $B$ determines which events regain their archived screenshots. Recent-$B$ assigns all visual slots to the latest events. CausalCache instead scores the complete history and swaps in an older event only when its predicted utility exceeds that of a recent event. A history-gated key/value adapter modifies only restored history-image tokens and is exactly bypassed when no history image is active, preserving current-screen processing. The adapter and selector are trained with matched-budget interventions on desktop trajectories and evaluated zero-shot on mobile. On OSWorld-Verified, activating historical screenshots improves success by about $13$ percentage points over summary-only memory. Under the official $15$-step limit, CausalCache and Recent-$4$ are statistically indistinguishable; in a $30$-step diagnostic, CausalCache achieves $46.7\%$ success versus $42.4\%$ ($+4.3$ points). Zero-shot on $117$ MobileWorld tasks, CausalCache improves over Recent-$4$ from $30.2\%$ to $36.8\%$. The gain is concentrated on a pre-defined cross-app memory-candidate split ($30.6\%$ vs. $19.4\%$, $+11.2$ points), while single-app controls show no detectable difference ($43.6\%$ vs. $42.4\%$). These results show that selecting which past events regain pixels is more effective than spending a fixed visual budget entirely on recency.
Pengyiang Liu, Junbo Niu, Xiaoyang Hu +4cs.CV cs.CL
A long-video answer is evidence-supported only when the frames decoded from the video cover every event the answer depends on. Existing evaluations score final-answer correctness or predicted evidence intervals, but the frames a method decodes before answering are rarely audited, so correct answers can still rest on incomplete observation. We introduce VES-Bench, a 600-question benchmark of Temporal Ordering and Event Counting items over 348 public long videos. Each item carries a jointly necessary set of evidence intervals, letting us audit at three strictness levels whether a method's decoded frames cover every one of them. We also propose TRACE, a training-free agent that grounds answers in raw visual clips, builds an evidence bundle round by round, and stops only when the answer stabilises as the bundle grows and a final pass over the same clips returns the same answer. Under a same-backbone audit, TRACE answers 50.7% of questions correctly with at least two decoded frames inside every evidence interval, at 98.7 frames per question: over 10 points above uniform decoding at 128 frames (40.2%), and within 2.6 points of uniform decoding at 256 frames at 0.39x its frame cost, while reaching the highest answer accuracy in the audit (63.5%). TRACE also stays competitive on Video-MME (86.1), LVBench (75.6), and LongVideoBench (75.1).
Reinforcement Learning (RL) has been promising in single-turn LLM fine-tuning. However, long-horizon agentic reasoning introduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: its heavyweight backpropagation-based training stack makes it impractical to fine-tune larger LLMs, and longer-horizon trajectories make credit assignment in RL substantially harder. This paper argues that evolution strategies (ES) can be a better choice for fine-tuning long-horizon LLM agents. Compared with agentic RL, ES offers three key advantages: 1) Model Scalability: ES enables full-parameter optimization with only minimal, inference-level GPU memory, making it possible to fine-tune large LLMs. 2) Flexibility: its lightweight, black-box feedback interface makes ES fine-tuning easy to compose with prompt-space evolution (e.g., skill optimization & test-time compute); and 3) Long-Horizon Scalability: ES performs trajectory-level parameter attribution without decomposing rewards across horizons, yielding better scalability than Agentic RL as the horizon length grows. Based on this insight, we propose Agentic ESOpt, a full-parameter agentic fine-tuning framework tailored to flexible parameter--context co-evolution. At each step, Agentic ESOpt samples perturbations around the current LLM parameters, evaluates the resulting agents with rewards, and applies an online reward-weighted update. To improve the exploration--adaptation trade-off, Agentic ESOpt further introduces a cosine decay schedule of the perturbation scale $σ$. On WebArena-Lite, full-parameter optimization of Qwen-3.5-27B improves the No Skill baseline by 6.69%. In test-time automatic heuristic design, Agentic ESOpt performs online prompt--parameter co-evolution, improving its matched baseline in 28 of 36 settings.
AlayaWorld Team, Kaipeng Zhang, Chuanhao Li +15cs.AI
This report presents an improved version of AlayaWorld. While the backbone architecture, chunk-wise autoregressive generation scheme, and training data remain unchanged from the previous release, we substantially revise how conditioning signals are represented and integrated into the model. The new design is guided by a simple principle: conditioning signals should match the generated content as closely as possible in both latent representation and temporal structure. To this end, we make two major changes. First, we replace the previous depth-warping-based spatial memory with a streaming 3D point-cache renderer. Second, we redesign the conditioning pipeline so that visual conditions are encoded in the same causal-VAE latent space, with temporal statistics consistent with those of the generated video. Concretely, the new version introduces six modifications: (1) replacing static-frame image conditioning with motion-aware latent conditioning; (2) causally encoding re-rendered spatial memory as a continuous sequence; (3) aligning the temporal-memory window in pixel space; (4) adopting hard memory dropout that removes memory tokens rather than zeroing them; (5) unifying the VAE encoding and decoding protocol across training and inference; and (6) removing the camera AdaLN branch, such that viewpoint control is provided entirely through the re-rendered spatial condition.
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.
Large language models (LLMs) power educational applications from tutoring to essay scoring, but each is a point solution to a single task, and only recently have these point solutions been integrated into agents operating over a learning management system (LMS). Yet tutoring is long-horizon, since a learner improves over days and weeks rather than in a single turn, and no benchmark evaluates an agent tutor across a sustained relationship. We introduce EduClaw-Bench, a benchmark that places an agent tutor in a continuous 30-day relationship with a simulated learner grounded in knowledge tracing (KT), whose knowledge-concept mastery, from a KT model trained on real-student data, drives its answers and is probed for learning gain across 55 scenarios. Each agent is scored on three primary axes (learning gain, responsiveness, and helpfulness) and two curriculum-design axes (Gagné and Rosenshine), with helpfulness and the curriculum axes judged by a cross-family panel of three LLM judges. Evaluating 10 agent adapters over three base-model tiers yields two findings that single-tier, single-session evaluation cannot reach. First, tutoring quality belongs to the base model and the agent harness together rather than either alone. Second, almost no combination sustains good tutoring over the full horizon. A calibration check ($\text{ECE}=0.049$) and a live-classroom field study confirm that the simulated learner and its measurements track reality. Our work is a step toward trustworthy AI tutors for future education.
We study what happens when a single general-purpose large language model acts as the sole researcher on a long-horizon neural architecture design problem. The agent receives a scientific question, an initial hypothesis and motivation, a compute budget, and research affordances (source and experiment management, experiment tracking, literature access, and persistent memory), then autonomously proposes, implements, evaluates, and records experiments over an extended period. The study comprises three phases, separated by human-declared transitions, that progressively expand the agent's tool surface or problem scale. Across approximately 100 sequential experiments, the agent improves a non-standard Vision Transformer from a weak baseline to a stronger, efficient model on small benchmarks and a usable but sub-SOTA model on ImageNet-1K, while producing a dense behavioural trace. We report four findings.(i)Productivity exhibits a clear phase structure: rapid early gains, a multi-dozen-hypothesis saturation wall, and recovery, with recovery triggered by expanding the action surface rather than changing the underlying model.(ii)A single early hypothesis contributes more to accuracy gain, with later improvements long-tailed.(iii)The preference for greedy, incremental hypotheses is largely workflow-induced: a commit-or-discard evaluation rule is isomorphic to greedy hill-climbing; the remainder reflects risk aversion after bold failures and anchoring on familiar literature. (iv)The agent independently rediscovers established results and, in the unfamiliar regime of pure channel attention, overturns a standard design choice. We conclude that workflow design was at least as influential as agent capability in this study and propose diversified search, budgeted moonshot hypotheses, explicit forks, and regime-aware re-validation as testable directions for future autonomous research.
Deep search agents answer difficult information-seeking questions by iteratively issuing search queries to gather supporting evidence, but it remains unclear whether and how greater search effort leads to better answers. We study these questions through a trajectory-level diagnosis of long-horizon search agents. Using human-annotated document-level relevance judgments, we evaluate the evidence retrieved at each search step and separate two stages of agent behavior: what evidence an agent retrieves and how effectively it uses that evidence. This distinction further allows us to decompose failures into retrieval gaps, where the necessary evidence is never found, and utilization gaps, where relevant evidence is retrieved but not used correctly. With the retrieval model and evaluation harness held fixed, we compare six agents on BrowseComp-Plus and further validate our findings on BrowseComp with an open-web search API. Across settings, we find that search effort and answer quality are only weakly aligned. Answer accuracy is better correlated with the quality of retrieved evidence, especially cumulative retrieval recall, than with the number of searches or the amount of context consumed. Useful evidence often appears early in the trajectory, yet agents tend to continue searching, producing a long tail of low-yield retrieval steps. At the query level, exploratory reformulations remain useful, but the best-performing agents issue far fewer redundant queries. Overall, by systematically characterizing the search behavior and failure modes of long-horizon search agents, this work points to practical directions for building better deep research systems, including stronger query formulation, more effective evidence selection and context management, and stopping criteria based on whether sufficient supporting evidence has been retrieved.
Jiabin Lou, Haopeng Wang, Yuanshuai Wang +6cs.RO cs.AI
Vision-and-Language Navigation (VLN) enables embodied agents to follow natural-language instructions. However, route-level instructions commonly encode spatial priors, such as orientation, distance, and layout, that are not explicitly available from onboard sensing at deployment in open, GPS-denied environments. Benchmark performance under such interfaces therefore jointly reflects visual navigation ability and the use of route structure explicitly supplied by the task description. As a complementary formulation, we propose Vision-Only Long-Horizon Navigation (VoLN), which shifts route-relevant information from externally supplied instructions and global guidance to locally observable in-scene cues. In VoLN, goal views specify the destination, while route-relevant information is available only through locally observable in-scene cues that the agent must detect, interpret, and select online. We instantiate VoLN for aerial navigation through VoLN-UAV, a 7,210-episode benchmark that combines long-horizon goal-directed flight, continuous 3D motion, large viewpoint changes, and context-dependent beacon selection. We further provide VoLN-MLLM as an initial reference baseline. It aligns self-supervised visual features with a structured semantic space and predicts short-horizon waypoint segments from observation history, goal views, retrieved visual--semantic tokens, and proprioception. On the five-environment Test-Unseen split, it obtains success rates of 7.4%, 4.5%, and 1.8% on Easy, Normal, and Hard episodes, respectively. These results provide an initial evaluation of VoLN and reveal substantial remaining challenges in long-horizon evidence integration, cross-view goal matching, and closed-loop stability. Project page: https://admire-ljb.github.io/VoLN-UAV/
Yuan Xiong, Linji Hao, Shizhu He +2cs.AI cs.CL cs.CR
Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.
LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or thousands of actions. In these settings, outcome-only rewards provide too sparse guidance, failing to inform the model about the goodness of intermediate actions. Dense supervision methods aim to solve this problem by scoring intermediate steps, from intrinsic confidence to self-distillation and embedding similarities. However, it is common practice to evaluate them by measuring the downstream performance of a training pipeline that integrates them. This is expensive, conflates supervision quality with training engineering confounders, and renders different methodological families requiring distinct training setups incomparable. As a result, dense supervision methods are rarely benchmarked on common ground. We introduce QVal, a training-free testbed for directly evaluating dense supervision signals. Given a state-action pair, QVal measures how well a method's score is Q-aligned: whether it orders actions according to the Q-values of a strong reference-policy. This lets us compare signals before any training run and separate signal quality from other engineering choices. We instantiate QVal as QVal-v1.0, benchmarking 21 dense supervision methods across four diverse environments and seven methodological families, with over 1.2K evaluation experiments across six open-weight model backbones. We find that simple prompting baselines consistently outperform recent dense supervision methods from the literature, and that performance clusters strongly by family. These findings hold across model sizes, environments, and observation modalities. QVal is designed to be easily extensible to new environments and methods, enabling researchers to iterate on dense supervision methods before any training run.
We introduce Agents-A1, a 35B Mixture-of-Experts Agentic Model that reaches trillion-parameter-level performance by scaling the agent horizon. We investigate agent-horizon scaling from two perspectives: scaling long-horizon trajectories and scaling heterogeneous agent abilities. To support this goal, we build a long-horizon knowledge-action infrastructure that connects external knowledge, actions, observations, and verifier outcomes, producing agentic trajectories with an average length of 45K tokens. Based on this, we train Agents-A1 with a three-stage recipe. First, we perform full-domain supervised fine-tuning to align the base model with broad agentic behaviors. Second, we train domain-level teacher models to capture specialized expertise in each domain. Third, we propose a multi-teacher domain-routed on-policy distillation with salient vocabulary alignment to improve knowledge transfer efficiency across different domains, unifying six heterogeneous domains into one deployable student model. Agents-A1 achieves strong and broad performance for long-horizon agent benchmarks. Compared with 1T-parameter model such as Kimi-K2.6 and DeepSeek-V4-pro, Agents-A1 achieves leading results on SEAL-0 (56.4), IFBench (80.6), HiPhO (46.4), FrontierScience-Olympiad (79.0), and MolBench-Bind (56.8), and remains highly competitive on SciCode (44.3), HLE (47.6) and BrowseComp (75.5). We hope this work provides the community with a practical path for scaling the horizon using a 35B agent that can reach or match the performance of 1T models on long-horizon tasks.
Zishang Jiang, Tingyun Li, Jinyi Han +7cs.LG cs.AI cs.CL
Reinforcement learning (RL) has become a widely adopted technique for improving large language models (LLMs) on complex tasks. Despite this progress, existing RL methods still face challenges in training agents with longer-horizon interactions. One major bottleneck is distinguishing the contribution of different actions in long-horizon interaction, leading to high optimization variance. To address this, we introduce a novel policy gradient method, Hindsight Policy Optimization (HPO), that projects both the current policy distribution and the hindsight distribution into an intent space and extracts low-variance learning signals from the Wasserstein distance between them. We theoretically and empirically show that aggregating semantically similar states and actions in the intent space yields a bounded-variance estimator and improves policy performance stably. Our code is available online.
LLM agents increasingly act as personal assistants that must remember a user's profile over months: who they are (attributes), what they routinely do (habits), and what they prefer (preferences), and keep it updated as jobs, routines, and tastes drift. Existing benchmarks evaluate this "memory" ability through short, simplified interactions, missing three core properties of real behavior: the profile is heterogeneous, with attributes, habits, and preferences evolving on different timelines; changes are driven by external context such as seasons and life events; and evidence is rarely stated explicitly, instead scattered across many small actions in different apps that a memory system must infer from. We introduce DynamicMem, a synthetic benchmark that constructs 15 months of activity per user, providing long-term multi-app data that real users' privacy keeps out of reach. It provides user-consistent trajectories averaging 2.2M tokens and 1,772 grounded events per user across 16 applications such as e-commerce, fitness, and social platforms. The profile evolves over this period and is never given explicitly: each attribute, habit, or preference must be inferred from small signals scattered across apps. We evaluate at five quarterly checkpoints to track how systems scale as history grows. Benchmarking five representative systems exposes problems a single accuracy score hides: (i) profile reconstruction degrades with history length while service-task accuracy stays flat, despite both drawing on the same memory; (ii) no system both keeps facts that stay true and replaces facts that change, with errors clustering on preferences and on naming the exact referent; and (iii) over 93% of failures trace to what the memory retrieves, not to the model writing the answer, so the largest room for improvement lies in memory itself. Code: https://wenyaxie023.github.io/DynamicMem/
LLM agents increasingly operate in large tool ecosystems, where real-world tasks require discovering relevant tools, inferring implicit sub-goals, and adapting to dynamic environments over long horizons. However, existing benchmarks rarely evaluate planning under retrieval-limited tool visibility. To address this gap, we introduce PlanBench-XL, an interactive benchmark of 327 retail tasks over 1,665 tools that tests whether agents can iteratively retrieve usable tools, invoke them to uncover intermediate evidence for subsequent calls toward the final goal. PlanBench-XL further features an optional blocking mechanism that simulates real-world unpredictability through missing, failing, or distracting tool functions, forcing agents to detect disrupted paths and adapt at runtime. Experiments on ten leading LLMs show that massive-tool planning remains challenging: while GPT-5.4 achieves 51.90% accuracy in block-free settings, it collapses to 11.36% under the most severe blocking condition. Further analysis shows that agents are especially vulnerable when failures lack explicit error signals or when recovery requires longer alternative tool-use paths. These results establish PlanBench-XL as a testbed for diagnosing agentic planning failures and highlight the need for robust adaptive planning in long-horizon tasks with large, imperfect tool environments.
Philippe Weinzaepfel, Christian Wolf, Bülent Mert Sariyildiz +2cs.CV cs.LG
Transformers are AI's workhorse with strong performance in modeling sequential data, but their computational cost becomes prohibitive when processing long sequences. We target long-horizon streaming vision and robotics applications like map-free pose estimation, where it is particularly impractical to store and maintain a history of observations. Recurrent Transformers address this limitation by maintaining fixed-size memory but their performance lags behind that of transformers operating over the full observation history. We argue that this gap does not stem from architectural limitations, but from differences in how these models learn to compress past information. Without access to an observation history, recurrent models must explicitly decide what to retain in memory at each step, a significantly harder learning problem. In this work, we propose a distillation approach that transfers the compression strategy of a classical full-history transformer to a recurrent variant. We enable this by designing a teacher model that explicitly compresses its observation history into a fixed-size bottleneck representation. By directly supervising the student's memory with this bottleneck representation, we align the two compression mechanisms. We show that this approach allows to train a recurrent latent robotic memory with linear-time complexity while substantially narrowing the performance gap to full-history transformers.
Long-horizon large language model (LLM) agents accumulate interaction trajectories that quickly exceed any practical prompt budget, and existing memory methods either truncate aggressively and lose non-local evidence or retain boilerplate that degrades decision quality. We ask a mechanism question rather than claiming a better general-purpose memory system: when does organizing trajectory memory into overlapping semantic units (OSUs) -- groups of related steps in which one step may belong to several units -- help retrieval over flat or disjoint alternatives? We instantiate this in OSU-Mem, which retrieves from an overlapping OSU pool via budgeted coarse-to-fine expansion, and show its benefit is conditional: overlapping memory helps when the evidence steps a query needs share tool calls or entities, but hurts when those steps are fully heterogeneous and share neither. On a synthetic benchmark where evidence carries such shared structure by construction, OSU-Mem improves over the strongest baseline as the theory predicts; yet on a concatenated, constructed unaugmented $τ$-bench setting its aggregate advantage over flat retrieval vanishes. Splitting queries by whether their evidence shares tools and entities shows this near-tie to be an artifact of mixing query types rather than a property of either method, and ToolBench, a controlled probe built to carry shared structure by design, corroborates the same mechanism via an overlap-vs.-disjoint construction contrast (under a coverage-guided variant), isolating the construction principle rather than validating the full default system. Because the relevant sharing is cheaply estimable from metadata, the analysis yields a metadata-based heuristic for predicting when overlap is likely to improve retrieval. We deliberately isolate the retrieval layer, assessed by retrieval quality and an LLM-mediated evidence-selection stage.
Zhenyu Yang, Kairui Zhang, Bing Wang +2cs.CV cs.AI
Despite the remarkable progress of Video Large Language Models (Video-LLMs), current online architectures still struggle to simultaneously process continuous video streams, decide autonomously when to respond, and preserve long-horizon contextual memory. These obstacles undermine real-time responsiveness and cause severe forgetting throughout prolonged interactions. In this work, we introduce LiveStarPro, a live streaming assistant that is designed for proactive video understanding over long-horizon streams. The design of LiveStarPro rests on three complementary components. The first component is Streaming Verification Decoding (SVeD), an inference framework that identifies the appropriate response timing through single-pass perplexity verification, thereby eliminating the dependency on explicit silence tokens. The second component is Streaming Causal Attention Masks (SCAM), a training strategy that enforces incremental video-language alignment over variable-length streams. The third component is Tree-Structured Hierarchical Memory (TSHM), a recursive memory architecture that organizes evicted historical information into event chains and consequently enables efficient retrieval from effectively unbounded video streams. To facilitate a comprehensive evaluation under realistic online conditions, we further present OmniStarPro, a large-scale benchmark that spans 15 diverse real-world scenarios and that extends to hour-scale streams for the assessment of long-term recall. Extensive experiments demonstrate that LiveStarPro consistently surpasses existing methods, attaining a 28.9% improvement in semantic correctness and an 18.2% reduction in timing error, while its streaming key-value cache further yields a 1.58x inference speedup over the same model without caching. The model and the code are publicly available at https://github.com/sotayang/LiveStarPro.
Recent vision-language models (VLMs) have shown promising progress in generating webpages from visual inputs, yet existing evaluations mainly focus on short, single-screen, and largely static webpages. We introduce LongWebBench, a benchmark for evaluating long-horizon webpage generation from both structural and functional perspectives. LongWebBench contains 490 real-world long webpages for structural fidelity evaluation and 507 goal-oriented interaction tasks over 129 webpages for functional evaluation. It employs two complementary protocols: a multi-dimensional VLM-based metric for assessing long-range structural coherence, and a DOM-augmented agent-based pipeline for end-to-end functional verification. We further examine the automatic evaluation protocols through human agreement analysis. Experiments with state-of-the-art open-source and proprietary VLMs under single-image and multi-image settings reveal that structural fidelity degrades as webpage length increases, while visually plausible generations often fail to support executable multi-step interactions. These results highlight the need to evaluate long webpage generation beyond visual similarity, with executable interaction as a core criterion. Our code and data are available at https://github.com/zheny2751-dotcom/LongWebBench.
Issa Sugiura, Daichi Hattori, Kazuo Araragi +5cs.AI
As LLM agents become capable of increasingly long-horizon tasks, evaluating their performance in economic systems is becoming increasingly important. Unlike existing benchmarks that primarily evaluate a single agent interacting with a passive environment, economic systems are inherently multi-agent, requiring autonomous agents to communicate, negotiate, and transact while pursuing their own objectives over extended periods. We introduce CoffeeBench, a benchmark for evaluating LLM agents in a long-horizon multi-agent economy composed of heterogeneous firms. In CoffeeBench, two farmers, two roasters, and two retailers autonomously operate their businesses over a 90-day simulation, each seeking to maximize cumulative net income through communication and transactions while managing cash, inventory, and pricing. The evaluated model controls one coffee roaster, while the remaining firms are controlled by fixed reference agents. Across several recent open-weight and proprietary LLMs, all models outperform a passive baseline that takes no actions, with most achieving positive net income. Analysis of agent behavior reveals substantial differences in long-horizon economic interaction: higher-performing models communicate more actively with other firms, whereas Claude~Haiku~4.5 exhibits an idle-drift failure mode, repeatedly choosing inaction despite producing coherent assessments and plans. We release our code and agent trajectories to support future research.
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.
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.
In recent years, large language models have shown substantial potential in psychological support tasks. However, existing psychological counseling data mostly rely on single-turn question answering or short multi-turn dialogues, making it difficult to characterize how college students' psychological distress accumulates, interacts, and gradually evolves over long periods within campus life events. To address this issue, this paper proposes Psy-Chronicle, a structured data-generation framework for synthesizing long-horizon campus psychological counseling dialogues. We generate a semester-spanning temporal stress event graph to model the chronological order and evolutionary dependencies among campus stress events. Through interactive simulation between a student agent and a counselor agent, together with a structured memory integration mechanism, Psy-Chronicle generates long-horizon dialogues with continuity across counseling sessions. Based on Psy-Chronicle, we construct and open-source CPCD, a Chinese long-horizon dialogue dataset for college psychological counseling, containing 100 student profiles, 90,000 counseling dialogues. We further build CPCD-Bench to evaluate models' long-horizon campus counseling capabilities from three dimensions: session-level response, long-horizon memory recall, and temporal-causal reasoning. Experimental results show that CPCD effectively improves session-level response generation and long-horizon memory recall for models with the same base architecture. Meanwhile, improvements in temporal-causal reasoning remain limited, indicating that event-chain organization and causal explanation are key challenges in long-horizon psychological counseling modeling. The related code and data are available at: https://github.com/EdwinUSTB/Psy-Chronicle
Language models are saturating benchmarks for procedural tasks with narrow objectives. But they are increasingly being deployed in long-horizon, non-stationary environments with open-ended goals. In this paper we introduce KellyBench, an environment for evaluating sequential decision-making in sports betting markets. Agents are placed in a sequential simulation of the 2023-24 English Premier League season and tasked with maximising their long-term bankroll growth. They are given detailed historical data, including advanced statistics, lineups, and public odds. To succeed they must build machine learning models, identify edge in public markets, and adapt as the environment changes over time. We find that all frontier models evaluated lose money on average over the course of the season for five seeds. The best performing model achieves an average return of -8%, and many models experiencing ruin across seeds. To judge strategy sophistication, we use a human expert rubric to grade each model and find their approaches to be unsophisticated compared to human baselines; Claude Opus 4.6 achieves a rubric score of 26.5%, which means there is significant room for improvement. KellyBench is available as an open-access API endpoint at https://openreward.ai/GeneralReasoning/KellyBench.
Autonomous LLM agents increasingly operate in long-horizon, interactive settings where success depends on reusing experience accumulated over extended histories. However, existing agent memory systems are fundamentally constrained by text-context budgets: storing or revisiting raw trajectories is prohibitively token-expensive, while summarization and text-only retrieval trade token savings for information loss and fragmented evidence. To address this limitation, we propose Optical Context Retrieval Memory (OCR-Memory), a memory framework that leverages the visual modality as a high-density representation of agent experience, enabling retention of arbitrarily long histories with minimal prompt overhead at retrieval time. Specifically, OCR-Memory renders historical trajectories into images annotated with unique visual identifiers. OCR-Memory retrieves stored experience via a \emph{locate-and-transcribe} paradigm that selects relevant regions through visual anchors and retrieves the corresponding verbatim text, avoiding free-form generation and reducing hallucination. Experiments on long-horizon agent benchmarks show consistent gains under strict context limits, demonstrating that optical encoding increases effective memory capacity while preserving faithful evidence recovery.