arXiv subjects

Jiaxin Zhang

Publications and source records attributed to Jiaxin Zhang.

At least 19 recordsLinked to original sources

arXiv2026

MTA-Zero: Self-Improving Multimodal Deep Search via Online QA-Solver Co-training

Multimodal large language models (MLLMs) have become strong visual reasoners, yet remain limited in open-world, multi-step search that requires grounding images, selecting tools, and integrating web evidence. Existing approaches typically construct and filter fixed offline corpora whose question distributions cannot adapt as the target deep-search agent improves. Building on MTA-Agent's image-grounded QA construction, we present MTA-Zero, an online multimodal QA-Solver co-training framework for self-improving multimodal deep-search agents that requires no precomputed multi-hop corpus. Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online. A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions. To focus generation on the Solver's learning frontier, we derive a Goldilocks reward for the QA Generator from the Solver's group-wise rollout success rate: questions solved by all rollouts or by none receive zero reward, whereas questions near the current frontier are favored. We optimize the two policies independently with GRPO, without sharing parameters or gradients. Across six multimodal deep-search benchmarks, MTA-Zero improves a Qwen3-VL-8B Solver from 36.91% to 51.58% mean accuracy and achieves performance comparable to GPT-5 under the same agent workflow. The same framework also improves a newer Qwen3.6-27B Solver by 10.82 points. These results show that verified online question generation provides a policy-specific curriculum for continually improving multimodal deep-search agents.

cs.CL↗

arXiv2026

Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?

Aiming to modify textual content while preserving the original style, Scene Text Editing (STE) has significant practical value in many applications. STE typically requires paired images before and after editing for training. However, recent approaches rely heavily on synthetic paired data and complex disentanglement strategies. They struggle to faithfully capture the diversity and complexity of natural scenes, often producing artifacts or hallucinations. In this work, we challenge the common practice and investigate whether STE can be driven by turning each real sample into its own paired supervision. Concretely, we explore multiple ways to disrupt the original image (e.g., Crop, Shuffle, Mask) to construct self-paired training data, and build a strong baseline (TextRIRO}) upon a conditional diffusion architecture. Our experiments show that combining character-column shuffling with image cropping (Shuffle&Crop) effectively destroys text semantics while preserving key style cues, enabling the model to learn editing behaviors directly from the real text image. With this concise yet real-oriented design, TextRIRO achieves promising performance in both accuracy and visual realism across real-world benchmarks, substantially mitigating the synthetic-to-real domain gap. Furthermore, we leverage TextRIRO to build TextRIRO-3M, a large-scale realistic dataset for long-text recognition by concatenating edited and original images. Experimental results show that training recognition models with TextRIRO-3M significantly improves recognition accuracy, demonstrating that TextRIRO provides highly useful training resources for downstream tasks. Overall, TextRIRO lifts STE to realistic, reliable, and widely applicable new levels. Code is available at https://github.com/YesianRohn/TextRIRO.

cs.CV↗

arXiv2026

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and how the agent fails. We ask a complementary question: can hindsight teach an agent what it could have anticipated before acting? We introduce prospective learning, which uses post-hoc experience to supervise foresight predictions from the pre-interaction view, and instantiate it with Self-Retrospection Distillation (SRD). Intuitively, a completed trajectory reveals knowledge that would have been useful and pitfalls that should be avoided; SRD distills this privileged hindsight into trajectory-blind foresight of the same policy. Foresight serves only as a training target and need not be explicitly generated at inference time. Across 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines with gains of up to 24.2 pp. Its advantage is especially pronounced when reward contrast is scarce: when 37--98% of rollout groups are reward-uniform across model scales, yet SRD can still exploit learning signal from sampled trajectories. In the 2B setting, where 98% of groups are all-failure, the RLVR training ends up at 0.0% success, while adding SRD reaches 60.6% under the same rollout budget. Our results suggest that post-hoc agent experience is useful not only for evaluating or improving behavior, but also for shaping predictive representations before available interaction. Training code can be found at https://github.com/SalesforceAIResearch/SRD.

cs.AI↗

arXiv2026

Agentic Self-Awareness: Aligning Knowing, Acting, and Holding

Frontier AI agents increasingly act in environments where seemingly solvable tasks may lack prerequisites, contain conflicting requirements, or reveal blockers only during execution. We formulate agentic self-awareness (ASA) as the alignment of task-grounded feasibility judgments (know), warranted behavior (act), and responses to reconsideration (hold), distinguishing what a task permits from what an agent can solve. We build paired feasible and blocked tasks in Terminal-Bench 2.1 and retail $τ^2$-bench and introduce an authorship-aware M0-M3 rubric that separates recognizing a blocker from taking the certified deferral action for the correct reason. Separate probes assess stated feasibility beliefs and responses to a single challenge. In a Terminal-Bench protocol baseline, blocker recognition reaches 84.4%, but strict self-awareness (SA) reaches only 40.6%. Harness search without model-weight updates increases strict SA from 45.3% to 93.8% on reused Terminal-Bench search tasks; selection for abstention alone, however, can reduce success on feasible tasks. Independently initialized joint-objective searches improve both blocked-task SA and feasible-task solve on their own search sets. Retail held-out gains remain limited, and Terminal-Bench blocked-task transfer is not measured. This evaluation pipeline makes the recognition-action gap measurable and sets a two-sided target for improving agent decisions under changing task evidence.

cs.AI↗

arXiv2026

CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution

Terminal-agent capability depends jointly on model weights and the runtime harness that formats prompts, binds tools, and handles error recovery. Existing harness-model co-evolution approaches improve both components, yet often treat trajectories produced during harness search as an undifferentiated replay buffer. This practice overlooks that a trajectory's value for model training depends on the harness under which it was generated. To systematically analyze this interface, we establish an alternating co-evolution framework that decouples harness search and policy training through component-wise promotion decisions. Within this framework, we introduce CoTrace, a harness-aware data recipe that explicitly governs trajectory routing, provenance matching, and curriculum refresh. Under CoTrace, recurring execution failures guide harness synthesis, while policy training is strictly conditioned on verified rollouts matched to the adopted runtime for supervised fine-tuning (SFT) or fresh online interactions for reinforcement learning (RL). On the Tmax promotion split, CoTrace advances Qwen3.5-9B from 78 to 88 solved tasks under supervised fine-tuning while an online reinforcement variant reaches 90. Specifically, a compact harness-matched corpus produces steady model gains at substantially lower compute than much larger corpora pooled across sibling harnesses. Furthermore, evaluations on Terminal-Bench 2.1 and SWE-bench Lite show that out-of-distribution transfer depends fundamentally on harness compatibility, where maintaining consistency between training and evaluation runtimes prevents procedural execution breakdowns observed under foreign scaffolds.

cs.CL↗

arXiv2026

Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps

Reinforcement learning for long-horizon agents relies on purely retrospective training signals: credit is assigned only after observing environmental consequences, leaving the agent's belief at action time invisible to the gradient. We introduce Prospective Hindsight (PH), a self-calibrating training principle that augments any retrospective base method with a signal derived from the gap between the agent's prospective prediction (before feedback) and the retrospective evaluation (after feedback). This per-rollout surprise identifies samples where the agent's self-model is most inaccurate and amplifies their gradient contribution through a stop-gradient surprise-weighted advantage. Since the prospective predictor shares parameters with the policy, the two co-evolve, progressively shifting focus to the agent's remaining blind spots. We connect this principle to a privileged-information gap and show that minimizing the surprise residual provides a descent pathway on the agent's miscalibration rate; calibration thus emerges as a byproduct of optimization rather than from an added objective. On single-turn verifiable tasks and a multi-turn personal-agent task (under GRPO, on-policy distillation, and their combination), PH improves both task performance and calibration, with consistent gains across model scales. Notably, the dominant miscalibration mode shifts structurally between regimes, overconfident failures in single-turn, underconfident successes in multi-turn, yet the same training principle addresses both successfully.

cs.LG↗

arXiv2026

SpatialSpeak: QA-Native Reconstruction with Local and Global Context for Spatial Chain-of-Thought Reasoning

Vision-language models (VLMs) can benefit from geometric priors for multi-view spatial reasoning, yet answer-only training does not directly supervise the intermediate geometric estimates and their use in deriving quantitative spatial answers. We hypothesize that spatial chain-of-thought (CoT) supervision becomes more effective when the VLM first jointly learns complementary local geometry and global scene context through multi-view reconstruction. We introduce SpatialSpeak, a two-stage framework that connects QA-native reconstruction pretraining with spatial CoT learning. In Stage I, QA-Native Reconstruction Pretraining (QA-RP) combines marked-point 3D queries for fine-grained local geometry with object-center queries for global scene context across views. Both tasks are formulated as text-based question answering, allowing geometric estimation and subsequent reasoning to share the same autoregressive output interface. In Stage II, spatial CoT with Visual Compensation (CoT-VC) trains the model to express question-relevant geometric estimates and use them to derive answers, with reliability assessment and visual compensation supporting answer refinement when needed. On ReVSI, QA-RP increases the gain from CoT-VC from 2.6 to 6.9 points, and ablations show that both local and global reconstruction supervision are beneficial. SpatialSpeak achieves state-of-the-art results on ReVSI, VSI-Bench, and SPAR-Bench, with a ReVSI score of 62.8 that exceeds the strongest compared baseline by 8.7 points.

cs.CV↗

arXiv2026

SameFact: The Same Safety Facts Lead to Different Responses Across Interfaces

Safety evaluations often ask whether a model recognizes that an action is unsafe, whereas agent evaluations ask what the model chooses to do. Using safety judgments as evidence about action selection therefore raises a measurement question: does the influence of the same safety-relevant fact persist across response interfaces? We introduce SameFact, a matched-counterfactual benchmark that tests this question directly. SameFact contains 300 safe/unsafe pairs that hold the task, prior observations, candidate action, identifiers, and non-target facts fixed while changing a single state-grounded safety fact. Across six LLM backbones, we measure the effect of this matched intervention through three interfaces at the same candidate-action boundary: explicit safety judgment, checkpoint candidate admission, and open first-action selection. All six backbones show lower aggregate sensitivity under open first-action selection than under judgment, but the change is not a uniform attenuation: across 24 model-factor cells, Spearman agreement falls from 0.817 between judgment and checkpoint admission to 0.470 between judgment and open first-action selection, while pairwise ordering disagreement rises from 18.5% to 32.6%. A follow-up 2x2 first-response experiment shows that a checkpoint-style protocol increases measured sensitivity in all six backbones by 8.4-29.3 percentage points, whereas action-space effects and their interactions with protocol vary in magnitude and direction across models. These results show that the response interface is part of the measured quantity: judgment and action interfaces share safety signal, but do not provide interchangeable measurements of how safety-relevant facts shape model responses.

cs.CR↗

arXiv2026

GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games

Post-training large language models (LLMs) on two-player zero-sum games provides a verifiable reward, and the gains have been shown to transfer to general reasoning. Existing pipelines, however, rely on a handful of hand-written games and on advantage estimators tied to them, so both the supply of games and the advantage estimate on a large corpus limit scaling. We present GameSmith, which scales both. An LLM synthesises games inside templates whose reward, observation, and hidden-information code is fixed. Each game is verified by seed-paired replay and by an independent re-implementation of its specification, then expanded into encoding, rule, and size variants, giving 26 new games in 29 families and 336 verified variants from the three SPIRAL templates. The model is distilled from the self-play of a larger teacher (Qwen3-32B) on these games and then trained by self-play reinforcement learning with CRAB, an advantage estimator that credits each decision against a shrinkage estimate of the expected return of its position and is defined on every variant. On Qwen3-4B-Base, distillation raises the 13-benchmark mean pass@1 from 45.5 to 62.5, above the three SPIRAL games at the same budget, and self-play with CRAB reaches 64.9. The ordering holds across models, and transfer grows with the number of game families.

cs.LG↗

arXiv2026

All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts

Multilingual scene text recognition (STR) remains challenging due to the scarcity of training data for most languages and the difficulty of serving diverse scripts within a single model. Existing solutions either deploy one recognizer per language, inflating cost and introducing error accumulation, or rely on massive vision-language models (VLMs) that are expensive and still inaccurate on many scripts. In this work, we pursue an all-in-one multilingual recognizer that is simpler than per-language experts, lighter than VLMs, and more accurate than both. First, we construct TextMuSS-10M, a large-scale synthetic scene text dataset spanning 10 scripts and 229 languages. It provides balanced and sufficient supervision where real data is unavailable. Second, we propose ScriptMoE, a script-aware Mixture-of-Experts (MoE) architecture. It shares a single visual encoder and replaces the dense decoder with a sparse MoE block, which consists of an image-level router dispatches each image to the top-2 script-aligned experts and a shared expert absorbs cross-script knowledge. Extensive experiments on our assembled TextMuSS-Bench (10 scripts, 10,899 images) show that ScriptMoE achieves the highest accuracy of 82.06%, outperforming the strongest STR baseline by 1.31%. On the CC-OCR end-to-end multilingual task, replacing only the recognizer in PP-OCRv5 with ScriptMoE lifts F1 score from 65.71% to 80.89%, slightly surpassing the best VLM (80.73%) at a fraction of the parameter count.

cs.CV↗

arXiv2026

RegVGGT: Sustainable Visual Geometry Grounding for Streaming via Regulated Memory

3D reconstruction from a lengthy video stream input poses a dilemma for feed-forward reconstruction models (FFRMs), that a whole-stream inference context cannot be retained under limited GPU memory.Recent studies seek to resolve this problem via a trade-off between the integrity of inference context and GPU memory usage, which either suffer from a rapid memory inflation or degraded context integrity due to artificially capping memory usage.Driven by our key observation that the initial saliency of a token reliably dictates its long-term importance across the stream, we propose RegVGGT, a training-free token regulation method which aggressively regulates the tokens of incoming frames.By admitting at most 1% of tokens per frame to update the context memory, our method dramatically suppresses memory inflation as the stream progresses.Equipped with a FlashAttention-compatible token saliency estimation scheme, RegVGGT is capable of processing thousands of frames on a consumer-grade GPU with negligible compromise to reconstruction quality.Extensive experiments demonstrate that RegVGGT achieves state-of-the-art performance on long-horizon benchmarks across diverse FFRM prediction tasks, surpassing prior FFRM-based stream reconstruction baselines by a large margin.

cs.CV↗

arXiv2026

TecoPrompt: Temporal-Conservative Prompt Learning for Vision-Language Models

Prompt learning adapts vision-language models, such as CLIP, by adjusting a small set of context tokens. However, under few-shot supervision, even moderate label noise can disrupt prompt optimization. To address this issue, we propose TecoPrompt, a closed-loop robust prompt-learning framework that revisits optimal transport (OT) pseudo-labeling from a temporal perspective. TecoPrompt employs an entropic OT plan in the CLIP semantic space to obtain globally consistent label candidates. It verifies the reliability of these candidates by examining trajectory stability: a noisy label is only rewritten if the OT candidate remains unchanged within a K-epoch temporal stability window and passes a confidence gate based on Exponential Moving Average (EMA). This approach helps reduce confirmation bias. The rewritten labels are then integrated back into prompt training using a tri-group objective that includes three loss functions aligned with clean, mid, and noisy subsets. Experiments on seven datasets with synthetic symmetric and asymmetric noise, as well as Food101N, demonstrate significant performance improvements. For example, on the OxfordPets dataset, with 50% asymmetric noise, TecoPrompt achieves an accuracy of 0.843, up from 0.775.

cs.CV↗

arXiv2026

Fork Where the Model Changes Its Mind: Belief-Shift Branching for Tree-Structured Reinforcement Learning

Tree-structured rollouts give critic-free reinforcement learning with verifiable rewards (RLVR) step-level credit: fork a chain at an intermediate point, and sibling outcome differences estimate step value. Each fork adds sampling cost, so realistic budgets typically allow only a few forks per chain. A fork placed where the outcome is already largely settled yields siblings that mostly agree and provide almost no credit signal; hence, for a given tree size, where forks are placed largely determines how much step-level RL can gain. Most existing mainstream methods place forks by structure, such as fixed lengths, midpoints, and delimiters, or by next-token entropy. We formalize fork placement as locating the \emph{pivots} of the chain's value curve, where the expected outcome turns. We propose \emph{belief-shift branching}: read the model's answer belief at candidate boundaries and fork just before the step where consecutive beliefs diverge most. Three instantiations, none needing step-level supervision, span access levels: a black-box probe, a logit-lens depth profile, and a learned activation direction, which is fit offline and therefore used only in the validation before RL training. The signal only \emph{places} forks, and the probe costs about $1\%$ of step compute on mathematics and under $5\%$ on code when it runs inside the rollout engine. In that validation, against Monte-Carlo value curves, a belief-shift signal ranks first in each of the eight model$\times$benchmark panels, ahead of entropy, structural, and LLM-judge baselines. In RL across three model families and two domains, belief-shift forking leads every mathematics aggregate, on OLMo-3-7B by $+2.6$ aggregate and $+2.9$ on AIME 2026 over the strongest baseline, and sweeps every OLMo code column, by $+6.5$ on LiveCodeBench-medium.

cs.AI↗

arXiv2026

On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification

Memory-based self-improving agents--those that learn from an online stream of tasks and improve over time by maintaining a textual memory bank--have shown great promise in recent literature. However, the reliability aspects of these methods have been critically overlooked. In this work, we conduct a comprehensive re-evaluation of two memory-based methods, broadening the scope of evaluation along two axes: (1) including multiple self-improving runs to quantify variance, and (2) shuffling the tasks to investigate the effect of task order. Through these experiments, we make two observations that expose the fragility of current methods: First, agent evaluation is inherently noisy in complex environments and on multi-step tasks, and stacking a self-improving loop on top can further amplify this noise. Second, the agent's improvement is highly dependent on task order. Prior works often adopt default orderings that impose an implicit curriculum, acting as a hidden prerequisite for success. To better understand this fragility, we manually examine the agents' memory and hypothesize that task and environment underspecification contribute to this fragility. We validate this hypothesis by incorporating information that enables better specification, such as detailed rubrics and environment feedback, into the memory construction process. While this added information partially closes the performance degradation in previous experiments, significant gaps still remain, suggesting that other uncharacterized factors contribute to this fragility. Looking ahead, our work advocates for more rigorous evaluation protocols for self-improving agents by reporting results across multiple runs and stress-testing them under challenging conditions. Moreover, our findings on underspecification call for systems and interfaces that enable effective human oversight, preventing agents from failing in unforeseeable ways.

cs.AI↗

arXiv2026

FANS: Federated Adaptive Network Search Learning for Heterogeneous Devices

Heterogeneous Federated Learning (HFL) aims to train models across devices with diverse resource budgets while preserving data privacy. Existing HFL methods typically bind training to a small predefined menu of model configurations, which limits architectural coverage. To address this bottleneck, we introduce Federated Adaptive Network Search (FANS), a hypernetwork-based framework that learns a shared architecture space rather than a fixed set of client models. To optimize this shared space efficiently, we propose the Federated Parallel Scaling (FPS) algorithm, which jointly trains multiple sampled subnetworks in parallel with self-distillation so that larger sampled subnetworks can supervise smaller ones during local updates. We evaluate FANS on CIFAR-10, CIFAR-100, and MNLI using ResNet-18, DenseNet-121, and BERT-base, respectively. Across all benchmarks, FANS expands the feasible subnetwork pool by orders of magnitude (e.g., 4,680 candidates for ResNet-18 vs. 4 in existing methods) and improves the average accuracy-efficiency trade-off relative to representative HFL baselines. Device heterogeneity is emulated through resource tiers, and evaluation covers accuracy, parameter count, and MACs.

cs.LG↗

arXiv2026

Few-Shot Video Recognition via Hierarchical Metric Learning

Few-shot action recognition (FSAR) aims to recognize unseen action categories with only a small number of annotated video samples. Recent works typically apply single-prototype supervision at the network output and fail to sufficiently exploit rich cross-frame global spatial information in videos. Even existing multi-level metric schemes only impose parallel prototype constraints on intermediate layers, without progressive supervision along the full feature pipeline, which results in limited generalization ability of the learned class prototypes. Inspired by this, we present a novel method, hierarchical metric learning for few-shot action recognition (HML-FSAR). First, a spatial-enhanced module is developed to capture cross-frame global spatial representations. Combined with temporal MHA, heterogeneous alignment, spatial-temporal feature fusion and dictionary learning modules, it constructs the complete feature processing pipeline. Second, a hierarchical metric learning (HML) strategy is embedded into HML-FSAR. Composed of center metric, alignment metric, contrastive metric, dictionary metric and prototype metric, HML imposes progressive multi-stage complementary constraints from frame-level representations to final class prototypes, so as to jointly optimize feature compactness, heterogeneous spatial-temporal alignment, inter-class discriminability and anti-noise robustness. The proposed HML-FSAR method is validated on five widely-used FSAR datasets, and experimental results fully demonstrate its effectiveness.

cs.CV↗

arXiv2026

On the paucity of lattice triangles

A rational triangle $T$ (one whose angles are rational multiples of $π$) unfolds to a translation surface ${X_T}$. The lattice triangle problem asks to classify those $T$ for which ${X_T}$ is a Veech (lattice) surface, which means that the $\operatorname{SL}_2(\mathbb R)$-orbit of ${X_T}$ is closed in its stratum (so its projection to moduli space is a Teichmüller curve). The most mysterious regime is the "hard obtuse window" (largest angle in $(π/2,2π/3]$), where it is conjectured that no lattice triangles exist. Using an arithmetic reformulation of the Mirzakhani-Wright rank obstruction, we prove a quantitative theorem that rules out all but a proportion $n^{-1+o(1)}$ of the triangles in this window with denominator $n$. The main technical result in our proof was autoformalized by AxiomProver in Lean (using mathlib).

math.DS↗

arXiv2026

DuMateBench: Evaluating Autonomous Agents in Complex Real-World Workflows

Autonomous agents are increasingly adopted to complete complex, multi-tool workflows in real-world settings. However, existing benchmarks typically separate tasks by application or capability and evaluate agents in environments that are cleaner and more stable than those encountered in practice. We introduce DuMateBench, a real-session benchmark reconstructed from anonymized and privacy-screened user sessions collected from a large-scale production agent platform. Each task preserves the relevant pre-solution interaction history, persistent configurations, and workspace state, and is then validated through human verification. The resulting benchmark comprises 200 tasks spanning 8 broad scenarios and 17 fine-grained capability categories, with most tasks requiring multiple capability coordination. We execute these tasks in isolated Docker containers injected with three forms of real-world environmental complexity: Insufficient, Unstable, and Noisy, and assess performance using a hybrid deterministic and LLM-as-Judge evaluation protocol. Experiments across five representative autonomous-agent frameworks paired with four state-of-the-art LLMs reveal substantial gaps in strict task completion. Complementary robustness, efficiency, and diagnostic analyses further show that performance under environmental perturbations is jointly shaped by the capabilities of the LLM and the surrounding agent framework. The code and data are publicly available at https://dumatebench.com/.

cs.AI↗