arXiv AI

arXiv:2610.28441v1 Announce Type: new Abstract: Large Language Model inference over long contexts is increasingly bottlenecked by the cost of attending to a growing Key-Value (KV) cache. Sparse attention reduces this cost by accessing only a fraction of the cache, but existing methods face a trade-off between deterministic retrieval, which captures high-scoring tokens but discards residual attention mass, and sampling, which improves coverage at the cost of higher variance. Hybrid methods combine top-$k$ retrieval with uniform sampling, yet still exhibit a substantial gap from dense attention at aggressive sparsity. We introduce \emph{Stratified Sparse Attention (SSA)}, a unified framework combining deterministic retrieval with score-guided sampling. SSA deterministically attends to high-scoring tokens while sampling the residual using proxy attention scores and correcting for their inclusion probabilities, thereby recovering attention mass beyond the retrieval set. SSA encompasses top-$k$, top-$p$, and hybrid top-$k$+sampling as special cases. Across RULER, LongBench, LOFT, and AIME, and five models ranging from 3B to 27B parameters, SSA consistently improves over the baselines at matched sparsity. On RULER-HARD with Ministral-3-8B, SSA improves accuracy over the strongest practical baseline from 81.8\% to 84.2\% at 32K and from 63.8\% to 69.3\% at 128K under $50\times$ attention sparsity, recovering 94\% and 85\% of dense-attention performance, respectively. With optimized CUDA kernels, SSA achieves up to $2.96\times$ decode speedup over FlashInfer at 1M context, demonstrating that its accuracy gains translate into practical long-context inference acceleration. We open-source our code at https://github.com/skylight-org/SSA-Paper-Code.

arXiv:2610.28405v1 Announce Type: new Abstract: Recent advances in multimodal large language models (MLLMs) have substantially improved spoken interaction and video understanding. Existing online and streaming benchmarks primarily assess causal multimodal perception through timestamped question answering, whereas real-world interaction requires models to continuously identify task-relevant multimodal cues and decide whether, when, and how to respond. This leaves open how well current models can sustain real-time interaction throughout an evolving multimodal stream. To address this gap, we introduce DIVE, a benchmark for evaluating full-duplex interaction with visual engagement under continuous and realistic multimodal streams. It organizes six tasks into three complementary interaction scenarios. (1) Proactive Intervention evaluates whether models can identify emerging visual or semantic cues and intervene at the appropriate moment. (2) Continuous Engagement evaluates sustained participation through continuous visual tracking and visually grounded query interaction. (3) Interaction Coordination evaluates how models coordinate listening and speaking under overlapping and multi-party interactions. DIVE contains 525 audio/video samples with fine-grained human annotations and precise temporal metadata. Its evaluation protocol separates response quality, participation decisions, and response timing. We evaluate representative real-time models using response quality, window response rates, and reaction latency. Experiments show that leading real-time models achieve sub-second average response latency, yet their average window response rates remain at or below 53.77%. Response quality varies substantially across models and tasks, with adapting ongoing responses to newly arriving information remaining particularly challenging. We hope DIVE facilitates progress toward more natural and capable real-time multimodal interactive systems.

arXiv:2610.28385v1 Announce Type: new Abstract: Head-dimension compression can reduce the memory required by the key-value (KV) cache, but existing methods often reconstruct full-dimensional representations during inference, adding significant decoding FLOPs overhead. We formalize the design considerations for converting pretrained attention into a natively reduced-dimensional form that reduces both KV-cache storage and attention computation using a calibration set. Specifically, we consider reductions that (1) apply a shared projection to queries and keys so that they remain in a common reduced feature space, (2) derive this projection jointly from query and key activations to better preserve their interaction, and (3) are compatible with the pretrained RoPE transformation. To obtain a reconstruction-free realization of this reduction, we restrict it to coordinate selection and theoretically show that RoPE compatibility then holds if and only if complete rotary pairs are retained. Building on this characterization, we introduce RACE, a post-training method that selects rotary pairs using their joint query-key activation energy. We theoretically show that this criterion minimizes a separable, calibration-dependent upper bound on post-RoPE query-key product approximation error. The resulting reductions are folded into the model weights and require neither runtime reconstruction nor specialized kernels. Across three model families and different compression levels, RACE outperforms structured pruning and remains competitive with reconstruction-based methods on language-modeling, zero-shot, and long-context tasks. At 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8%, yielding 1.79x higher throughput, 43.2% lower time-to-first-token (TTFT), and 28.1% lower allocated GPU memory in comparison to an uncompressed model.

arXiv:2610.28361v1 Announce Type: new Abstract: Long-horizon deep search requires LLM agents to iteratively decompose complex information needs, gather evidence from external sources, and maintain a reliable research state under multiple constraints. A central challenge is that early mistakes can be amplified over the search trajectory: agents may anchor on an incorrect candidate, collect partially supporting evidence, overlook contradictions, and accumulate errors that misguide subsequent reasoning. We propose FALCA, a falsification-aware framework that formulates deep search as a propose-falsify process. FALCA adopts a collaborative Searcher-Falsifier architecture, where the Searcher proposes candidate answers or sub-answers, while independent Falsifiers actively seek counterevidence, check unmet constraints, identify conflicts, and produce concise falsification reports for research-state revision. To strengthen this loop, we introduce falsification-aware training, combining SFT warmup on falsification trajectories with agentic RL based on parallel rollouts and hierarchical credit assignment over research stages, state transitions, and actions. Experiments on five search benchmarks show that FALCA improves long-horizon deep search performance over backbone and harness ablations, while reducing confirmation bias and error accumulation through more effective falsification and research-state revision.

arXiv:2610.28354v1 Announce Type: new Abstract: Spatial reasoning asks a model to work out where objects are in 3D space, how they relate to one another, and how they move. Agentic VLMs have recently brought large improvements on such tasks by letting the model actively gather evidence with 3D perception tools instead of answering in a single pass. Even so, on many questions the tools return the right evidence and the model still does not arrive at the correct answer: the gap between model and tools, where raw readings become usable conclusions, is left unfilled. We present S-Agent 2.0, a spatial reasoning agent that equips an off-the-shelf general coding agent with a command-line perception toolkit, and we design the interface around one principle: help the VLM close the evidence loop: acquire scene evidence, remember it in a persistent scene memory, and interpret each reading in plain spatial language. This design keeps the full flexibility of a coding agent, and every reading it acts on stays grounded in the tools' geometry. On MMSI-Bench and ReVSI, the toolkit improves a strong open-source model, Qwen3.8-27B, by +13.9 and +20.5 points over the same agent without tools, and a strong closed-source model, Gemini-3.8-Flash, by +11.2 and +9.6 points. With the open backbone S-Agent 2.0 reaches 66.5 on MMSI-Bench and 82.3 on ReVSI, ahead of the specialized system SpatialClaw on the same backbone (62.6 and 74.5), and with Gemini-3.8-Flash it reaches 74.1 and 90.5 against its 67.0 and 79.7. Both tool gains shrink as the backbone grows more capable, but SpatialClaw's shrinks far more on ReVSI, from +13.2 to +2.7, while ours stays at +9.6. S-Agent 2.0 also works across different general coding agent harnesses, such as Codex and Claude Code.

arXiv:2610.28325v1 Announce Type: new Abstract: Factual sycophancy, a language model's deference to a user's false beliefs, is usually understood as a bid for the user's approval at the expense of truthfulness. We argue that this framing over-emphasizes the user and propose viewing sycophancy as a failure within the broader and generally useful process of in-context updating, by which models revise their answers based on information supplied during an interaction. We support this view with behavioral and mechanistic evidence. First, models' responses to user claims sit within a broader pattern of source-sensitive updating. They revise their answers in response to information attributed to external sources as well, more so when it appears better supported, and their deference to the user persists even when no user approval is at stake. Second, the internal components that support deference to the user are also causally important for in-context updating more broadly. Our results suggest that factual sycophancy arises when in-context updating gives a user's assertion excessive weight relative to its evidence. Because sycophancy is embedded in this process, reducing it risks reducing useful updating as well, a trade-off that calls for evaluations measuring both behaviors and mitigations that decouple them.

arXiv:2610.28282v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) have substantially advanced chart understanding and reasoning, yet their performance remains highly uneven across chart types. While modern MLLMs perform well on common charts, their accuracy drops sharply on long-tail charts with diverse visual structures and reasoning requirements. Tool-augmented agents can improve complex chart reasoning by introducing specialized visual capabilities, but manually engineering strategies and tools for the heterogeneous and evolving long tail is difficult to scale. We introduce \textbf{ChartEvo}, a recursive self-improvement framework that enables chart agents to adapt to underperforming chart types through experience-driven harness-level self-improvement. ChartEvo jointly evolves reusable chart-specific Skills and executable Tools, while learning reusable local tool-call chains to compose evolved capabilities and control the inference overhead of tool-augmented reasoning. On long-tail charts from ChartBench and ChartX, ChartEvo improves accuracy over its base MLLM, Qwen3.7-Plus, by \textbf{45.6\%} and \textbf{32.5\%}, respectively, demonstrating the effectiveness of recursive self-improvement for long-tail chart adaptation.

arXiv:2610.28201v1 Announce Type: new Abstract: Recursive self-improvement (RSI) of agent harnesses has emerged as an effective approach to enhancing the capabilities of large language model (LLM) agents. By searching the harness space for a specific LLM on a given dataset, RSI can discover a harness tailored to the capabilities and behavioral characteristics of that model. However, such optimization often creates a strong coupling between the learned harness and its target LLM, limiting generalization across models and requiring substantial computational resources to adapt to new LLMs. Moreover, existing RSI methods typically optimize a harness using feedback from a single model and evaluate its generalization to other models only after training. This paradigm treats model differences primarily as a criterion for evaluating generalization rather than as valuable evidence for optimization. In this paper, we introduce a new perspective that treats models as data and places the harness at the center of optimization. Building on this perspective, we propose \textsc{MaD-RSI}, a framework that directly optimizes a shared harness across multiple heterogeneous LLMs by incorporating their diverse failure trajectories into the training process. Rather than assessing model differences only after optimization, \textsc{MaD-RSI} leverages these differences as training evidence to guide the self-improvement of a shared harness. Furthermore, we introduce a new evaluation framework with two complementary metrics designed to assess the effectiveness of harness optimization across models. Experiments across multiple datasets and LLMs demonstrate that \textsc{MaD-RSI} improves generalization over existing approaches while reducing training costs.

arXiv:2610.28121v1 Announce Type: new Abstract: Safe autonomous navigation of UASs requires tactical deconfliction to address conflicts that strategic coordination cannot prevent. This task is challenging under partial observability, when sensor or traffic data are missing, delayed, or degraded. Geometry-based methods can offer separation assurances but face limitations in dense encounters. Recurrent reinforcement learning has therefore been proposed to learn navigation policies that integrate observation-action histories to optimise long-term safety-efficiency trade-offs. However, operational safeguards may require a remote pilot, onboard detect-and-avoid system, or runtime assurance mechanism to temporarily override the learned policy. A research gap remains in understanding how these interruptions affect recurrent policies when the agent regains authority. Benchmark evidence shows that interruptions make task variables less decodable from the recurrent hidden state and increase post-interruption collision risk. The findings suggest that altered recurrent context prolongs post-handback degradation through an action-observation feedback loop, in which changed actions alter the flight path and subsequent recurrent updates. To address this vulnerability, we propose three interruption-aware architectures that condition recurrent processing on interruption context through separate feature streams, dedicated interruption memory, or gated environmental features. Across simulations, all three architectures preserve more decodable task information than a standard recurrent policy. Moreover, the strongest architecture reduces post-interruption collisions by 74% and interruption-induced mission failures by 69% relative to the recurrent baseline, without degrading interruption-free performance. These results show that safe interruption does not guarantee safe handback, and support interruption-aware recurrent design for resilient UAS autonomy.

arXiv:2610.28110v1 Announce Type: new Abstract: Spiking Neural Networks (SNNs) have shown great potential for low-power neuromorphic computing due to their event-driven processing and spatiotemporal sparsity, but they remain vulnerable to input perturbations, hardware-intrinsic noise, and adversarial attacks in real-world deployment. We attribute this vulnerability partly to input-specific hidden representations learned by standard SNNs trained with conventional cross-entropy loss, as our information-plane analysis empirically shows that substantial input-related information can still be retained after label-predictive information becomes stable. To address this issue, we propose SCIB, a simple yet effective robust learning objective for SNNs based on the Conditional Information Bottleneck. Different from conventional variational IB methods, SCIB replaces the mutual information compression term $I(X;S)$ with the conditional mutual information $I(X;S|Y)$ and derives a tractable variational upper bound. Moreover, it introduces a learnable class-conditional prior instead of a label-agnostic isotropic Gaussian prior, improving the expressiveness of variational compression and encouraging class-structured representations when combined with the supervised classification objective. Systematic experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 under external environmental noise, neuromorphic hardware noise, and adversarial attacks show that SCIB improves robustness over standard SNNs and IB-based baselines under most evaluated noise conditions, improves adversarial accuracy under the evaluated attacks without using adversarial training, and reduces the SynOps-based single-sample inference energy consumption by 24.9% relative to an unconstrained SNN baseline.

arXiv:2610.28103v1 Announce Type: new Abstract: Structural Causal Models (SCMs) provide a principled framework for observational, interventional, and counterfactual reasoning, all of which rely on probabilistic inference. Existing frameworks either transform SCMs into alternative tractable representations or rely on approximate inference with expressive generative models. We instead perform inference using a message passing scheme directly on the SCM by representing structural mechanisms and probability densities in a common finite-dimensional function space. Forward propagation combines a numerical realization of the Law of the Unconscious Statistician, while propagated densities are recovered by spline projection from LOTUS moments. The same representation naturally yields backward message passing for conditioning, providing a unified framework for causal inference. Using tensor-product B-spline bases, the resulting algorithm expresses inference through numerical quadrature and linear algebra without explicitly constructing conditional distributions. Experiments on randomly generated continuous SCMs demonstrate accurate approximation of Monte Carlo reference solutions and systematic improvement with increasing basis resolution.

arXiv:2610.28102v1 Announce Type: new Abstract: Scenario planning is a foresight practice in which analysts construct a small set of plausible futures and prepare an organisation for each. Each scenario is a narrative that must be plausible, internally consistent, relevant and challenging, and the set as a whole must be differentiated. The practice remains manual: a workshop (i.e., a facilitated meeting in which a small group of experts writes the narratives) considers only the factors it has time for, carries the biases of the experts who run it, and cannot be redone whenever conditions change. In this paper, we automate scenario planning by casting it as a diversity planning problem, in which two scenarios differ when they end in different narratives on the analyst's drivers. The result is a set of scenarios, one for each narrative that the topic can reach within a budget. Our implementation, Polyscene, is a generic engine that generates scenarios from a topic. We use geopolitical conflicts and shipping-route disruptions as case studies. For the first, we build the planning problem's initial state from the GDELT Project so it reflects current relationships among the countries involved. Polyscene then generates scenarios showing how relationships among those countries change. With a board of six countries as of 1 January 2021 and a budget of five events, the conflict set puts five of the 15 country pairs at war, and each of the three wars that break out among these countries by 1 January 2026 ends a scenario. For the second, we build the initial state from ship traffic, a supply-pressure index, and the news, and Polyscene generates scenarios showing what a closed or diverted route could do to a manufacturer's parts stock. With four routes per board and the same date and budget, each set includes three outcomes, and the Suez Canal set includes the March 2021 blockage and the December 2023 Red Sea container diversion.

arXiv:2610.28085v1 Announce Type: new Abstract: The rapid growth of scientific literature makes survey writing increasingly challenging, motivating automatic survey generation. Yet, existing systems must search an open space of papers and an open space of concepts with no trace of how the community itself has navigated them, leading to incomplete scope, weak taxonomies, and unreliable citations. Our core insight is that author-written Related Work sections contain the human judgments needed to address both limitations: authors curate the citations relevant to their contributions and explain how the cited works relate, yielding paths for scope reconstruction and relational evidence for organization. We propose InsightRAG. Its insight-guided retrieval follows author-supported citation relations to reconstruct a verified research scope and structure the resulting relational evidence. It aggregates these judgments under a minimum-conflict objective into a taxonomy that reflects recurring author-written perspectives in the retrieved literature, guides generation, and bounds each section's evidence. Results on ten topics show gains in scope reconstruction, taxonomy quality, citation support, and reference reliability.

arXiv:2610.28083v1 Announce Type: new Abstract: Electrocardiography (ECG) is a cornerstone of cardiac diagnosis, yet most machine learning methods for ECG analysis lack interpretability, limiting their clinical adoption. We present IECG, a weakly-supervised ECG foundation model that jointly performs diagnostic classification and abnormality localization without requiring manual abnormal region annotations, overcoming a long-standing bottleneck in interpretable ECG learning. Central to IECG is a weakly-supervised attention strategy that explicitly incorporates clinical priors by deriving abnormality heatmap pseudo-labels from established ECG clinical knowledge. These pseudo-labels guide the model's CLS-token attention through a KL-divergence loss, encouraging it to focus on diagnostically relevant waveform regions. During pretraining, IECG is jointly optimized with clinically guided attention supervision, ECG-text contrastive alignment, and report captioning objectives, enabling it to learn rich multimodal representations while producing localized waveform evidence. By explicitly aggregating features from clinically guided attention regions, IECG grounds its diagnostic predictions in interpretable waveform evidence without sacrificing classification performance. Extensive experiments demonstrate that IECG achieves state-of-the-art abnormal region localization while maintaining strong classification performance, advancing trustworthy foundation models for ECG analysis.

arXiv:2610.28074v1 Announce Type: new Abstract: As a prototypical latent world model, Joint-Embedding Predictive Architectures (JEPAs) provide strong semantic representations. However, dense patch representations expose a redundant interface to downstream predictors, which must implicitly infer spatial organization and temporal correspondence, presenting three challenges. First, object information is distributed across patches, forcing each downstream predictor to redundantly recover object structure that could otherwise be reused across tasks before modeling interactions. Second, the lack of explicit entity correspondence across time makes historical transitions expensive, restricting the predictor's flexibility in utilizing context. Third, without explicit transport structure, existing JEPA models typically use pointwise l1/l2 distances, which are insensitive to spatial displacement when predicted and goal supports are largely disjoint. We propose the Slot-based Joint-Embedding Predictive Architecture (Slot-JEPA), which decomposes dense representations into stable semantic slots and time-varying spatial masks. Pretrained once and frozen downstream, this decomposition exposes a reusable object-structured interface. Building on it, our Transition Memory Predictor (TMP) stores visual transitions as compact masks and retrieves useful context using statistics of corresponding slots. Its two branches separate slot-specific dynamics from shared cross-slot dynamics. We further use Unbalanced Optimal Transport (UOT) to supervise mask prediction and evaluate candidate plans by transporting mass between aligned predicted and goal masks, providing a signal better aligned with spatial progress. Across seven visual-understanding benchmarks, the slot-reconstructed features remain transferable. Slot-JEPA also improves goal-conditioned planning across six planar environments and four robotic arm planning tasks on the REALM-JEPA benchmark.

arXiv:2610.28060v1 Announce Type: new Abstract: The Reinforcement Learning (RL) paradigm has achieved remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). However, when applied to long-horizon tasks, it faces severe challenges: i) Due to the high task complexity, the model struggles to sample enough successful trajectories, resulting in slow convergence. ii) Relying solely on sparse final rewards causes all steps to be updated indiscriminately, encouraging redundant step generation. When both the quantity and quality of successful trajectories become difficult to guarantee, the performance is severely constrained. This dilemma stems from the limitations: existing RL paradigms rely entirely on the reasoning capability of base models, lacking the ability to actively optimize trajectories. To address the above issues, we propose a novel trajectory evolution paradigm with hindsight credit assignment, termed TrajEvolve. TrajEvolve not only relies on the trajectory inferred by the model but also generates ``evolutionary trajectories'' by refining low-quality steps in historical trajectories, thereby exposing the policy to high-quality compositions that lie within the base model's support but are rarely sampled within practical training budgets. Furthermore, TrajEvolve reformulates step-level continuous credit assignment as a binary classification task, whether each step is necessary for task completion. This simplification not only reduces the learning difficulty but also enables step labels to be obtained through counterfactual verification, achieving precise identification of redundant steps. Extensive experiments on two datasets demonstrate the superiority of the proposed method.

arXiv:2610.28003v1 Announce Type: new Abstract: Diffusion models generate samples through a multi-step denoising trajectory, yet condition every step on the same instruction, so the trajectory cannot be addressed at the level of individual operations. We introduce Dancing Diffusion, which assigns each span of timesteps a dedicated continuous condition block and thereby turns conditioning into a latent program. A compiler maps text features into this program, and a fixed pretrained generator executes its blocks over successive sampling intervals. The trajectory can then dance, keeping its executed history while its continuation is redirected. A shared prefix is executed once and reused, and a later operation is rewritten while content fixed by earlier operations is preserved. The compiler is trained through the images it produces, with visual feedback on both the requested change and the retained content. On SANA, reusing an executed prefix preserves markedly more of the scene than regenerating from noise at a similar strength of edit, learned conditions preserve more background than native prompt switching from the same saved state, and the learned program improves generation quality over native SANA. Later branches weaken the edit, which leaves semantic control over individual operations limited. By converting a single-instruction trajectory into a program of addressable operations, Dancing Diffusion lays a foundation for structured intermediate computation in diffusion models.

arXiv:2610.27993v1 Announce Type: new Abstract: Telecom language models must combine broad networking knowledge with specialist behavior across routing, broadband services, access technologies, and operational diagnostics while remaining practical to deploy as a single model. We present an end-to-end framework that builds source-grounded synthetic training data from structured networking taxonomies and expert graphs, materializes 212,559 routed training records across fourteen networking routes, trains six frozen specialist LoRA teachers over one shared continued-pretrained backbone, and consolidates their behavior into one deployable student adapter using Routed Specialist On-Policy Distillation (RSOD). In RSOD, the committed student generates each trajectory, the route-assigned specialist teacher-forces the exact native token sequence, compact teacher hidden states are reconstructed through the shared frozen output head, and only the student LoRA is updated. A controlled reconstruction validation shows mean absolute logit error of 7.26 x 10^-6 with 100% top-1 agreement between native and reconstructed teacher logits. On four RTX 6000 Ada GPUs, the completed 2,000-step run produced 16,000 trajectories with zero observed student-version or token- ID mismatches. Under the primary route-conditioned evaluation over 2,850 held-out questions, the raw Meta-Llama-3.1-8B-Instruct model scores 12.8661/25 micro (13.0172/25 macro), the combined routed six-teacher system scores 21.7641/25 micro, and the consolidated RSOD student, released as Llama-CalixCLM v1, reaches 21.7112/25 micro (21.8006/25 macro) on the same five dimension LLM-judge rubric. Under a separate neutral deployment-like prompt, the same student reaches 20.0028/25 micro (20.0778/25 macro). The consolidated student remains within 0.0529 micro-score points of the combined routed-teacher system while retaining a numerically validated and auditable train-serve execution path.

arXiv:2610.27965v1 Announce Type: new Abstract: Can humans and artificial intelligence cooperate such that their combination yields greater accuracy than either agent alone? This result, termed \emph{synergy} or \emph{complementarity}, is notoriously difficult to achieve, even when both agents possess significant independent accuracy. In this paper, we present a formal framework to model human-AI interaction through the lens of reliance patterns, which characterize the joint correctness structure of human and algorithmic decisions. We derive mathematically necessary and sufficient conditions for synergy, demonstrating that its emergence depends on the ability of either agent to identify and correct the errors of the other, what we call Human and Algorithmic Correction. Beyond its theoretical contribution, the proposed framework provides a foundation for the design and evaluation of knowledge-based decision support systems, where interaction policies must dynamically integrate heterogeneous sources of expertise. We demonstrate this through the analysis of seven clinical case studies. Our empirical results reveal a ``dominance trap,'' where a strong algorithmic advantage can paradoxically inhibit synergy by eroding human corrective intervention. We also show that synergy can persist even in the presence of cognitive biases, such as automation bias, provided there is functional de-correlation between human and AI errors. These results provide a formal foundation for key concepts in the human-AI systems literature and offer actionable insights for designing interaction-centric AI systems that foster complementarity.

arXiv:2610.27954v1 Announce Type: new Abstract: Recent advances in large (vision) language models have enabled computer-use agents that automate digital workflows via fine-grained keyboard and mouse operations. However, such agents remain fragile in long-horizon settings, where errors accumulate and executed actions leave unintended side effects, making the ability to recover from failures critical. Yet, existing remedies remain inadequate: environment resets are infeasible in proprietary settings, action replay is prohibitively expensive over long horizons, and API-based rollback is limited to web environments, leaving model-based recovery largely unexplored. To address this gap, we present ReTrace, a scalable evaluation framework for systematically studying recovery behavior in computer-use agents. We formalize Transition-Consistent Recovery: identical downstream actions from the recovered and target states must yield identical outcomes. We operationalize this criterion through goal-oriented action replay following controlled error injection. Using ReTrace, we present the first systematic study of autonomous, model-based recovery, revealing that (i) ideal recovery can unlock model's latent task completion capabilities, (ii) current computer-use agents fail to exploit their latent recovery capabilities, and (iii) accurate model-based recovery yields measurable end-to-end gains. Together, these findings establish recovery as a fundamental capability for building reliable computer-use agents. Our code and dataset is available at https://github.com/BangProx/ReTRACE.