arXiv AI

arXiv:2303.15619v3 Announce Type: replace-cross Abstract: The choice of which tokens to mask is an under-examined design decision in masked language modeling (MLM). Standard pretraining masks tokens uniformly at random, whereas content-aware masking schemes can improve downstream transfer. We study masking as a task-adaptive component of the fine-tuning pipeline and introduce Typhoon, a masking strategy that uses the gradient of the MLM loss with respect to one-hot token inputs to estimate online how much each token type contributes to the objective. Typhoon maintains an exponential moving average of per-token-type saliency and calibrates these scores into a masking distribution whose expected masking rate matches a target budget under a token-independence approximation. We formalize the method and evaluate it against random masking and whole-word masking on two GLUE tasks, MRPC and CoLA, across three BERT-family backbones (TinyBERT, DistilBERT, and BERT-base) and five random seeds per configuration ($90$ training runs in total). Once seed variance is accounted for, no masking strategy is reliably better than the others on these tasks: on MRPC the gap between Typhoon and the best baseline stays within $0.004$ $F_1$, across all twelve Typhoon comparisons no paired test reaches significance, and every $95\%$ confidence interval contains zero. Inspecting Typhoon's learned distribution explains why: budget-preserving min-max calibration compresses the gradient ranking into a narrow band whose Kullback-Leibler divergence from uniform random masking is $7\times 10^{-4}$ nats per token, and widening the calibration spread by an order of magnitude in KL still leaves downstream accuracy unchanged within seed noise. Typhoon's apparent advantage in single-run experiments does not survive multi-seed evaluation, showing that gradient-based task-adaptive masking is competitive with, but does not outperform, uniform random masking at this scale.

arXiv:1804.08328v2 Announce Type: replace-cross Abstract: Do visual tasks have a relationship, or are they unrelated? For instance, could having surface normals simplify estimating the depth of an image? Intuition answers these questions positively, implying existence of a structure among visual tasks. Knowing this structure has notable values; it is the concept underlying transfer learning and provides a principled way for identifying redundancies across tasks, e.g., to seamlessly reuse supervision among related tasks or solve many tasks in one system without piling up the complexity. We proposes a fully computational approach for modeling the structure of space of visual tasks. This is done via finding (first and higher-order) transfer learning dependencies across a dictionary of twenty six 2D, 2.5D, 3D, and semantic tasks in a latent space. The product is a computational taxonomic map for task transfer learning. We study the consequences of this structure, e.g. nontrivial emerged relationships, and exploit them to reduce the demand for labeled data. For example, we show that the total number of labeled datapoints needed for solving a set of 10 tasks can be reduced by roughly 2/3 (compared to training independently) while keeping the performance nearly the same. We provide a set of tools for computing and probing this taxonomical structure including a solver that users can employ to devise efficient supervision policies for their use cases.

arXiv:2610.26650v2 Announce Type: replace Abstract: A repair action proposed by a large language model (LLM) specifies which records to change and how to change their values. A failed action may still contain a useful choice of records or a useful update, depending on how the two are paired. We introduce a counterfactual evaluation framework that separates selected records (targets) from the applied change (an update). Four executions combine proposed and reference targets and updates to measure the effects of correcting or retaining each component and their interaction. We evaluate the framework on two tasks: traffic assignment, where affected demand must be identified from road counts, and database repair, where affected rows must be identified from aggregate reports. An injected fault supplies a reference action that reverses it, and the recovery rate measures an action's loss reduction relative to that reference. For sparse-target actions in the traffic experiments, retaining the proposed update on reference targets achieves 79.4% of the reference loss reduction, compared with 43.7% from tuning update strength on the proposed targets. In the database experiments, update replacement worsens 17 of 108 actions, all selecting unaffected records, with sharply different frequencies across models. The value of a proposed update depends on the targets it is paired with. Missed targets hide useful updates, while excess targets can make a milder update better than the reference correction. In a component-reuse study, testing new target-update pairings from two proposals raises mean held-out recovery rate to 0.616, compared with 0.580 when selecting between the original proposals.

arXiv:2610.26636v2 Announce Type: replace Abstract: Code generation models are increasingly used to produce interactive artifacts, such as websites and user interfaces. Many properties that determine whether these interfaces are useful only become apparent through realistic interaction, making them difficult to incorporate into model training. In this paper, we introduce RELIEF (REvision-based Learning from Interactive Experience and Feedback), a method for improving code generation models using feedback on their generated interfaces. RELIEF uses a coding agent to revise these interfaces, then incorporates the improvements into their original generation traces to produce supervision for fine-tuning. To evaluate RELIEF, we collected a dataset of 200 synthetically generated websites paired with 355 interaction recordings and accompanying feedback from 24 users. We used critiques derived from code inspection, automated interaction, explicit human feedback, and human interaction recordings to fine-tune variants of two code generation models. Our results show that feedback effectiveness varies with base model capabilities, with critiques inferred from human interaction recordings leading to the largest usability improvements overall under our tested models.

arXiv:2610.26342v2 Announce Type: replace Abstract: Recursive self-improvement (RSI) requires agents to diagnose system bottlenecks and coordinate improvements across training and inference. We refer to the agents' capability to jointly optimize training data, the trainer, and the inference harness as cross-component co-design. We introduce VibeTraining-Bench, a benchmark for evaluating this capability across tasks spanning terminal operation, software engineering, multi-turn tool use, and mathematical reasoning. Within each task, research agents start from an identical initial agent system under a fixed resource budget. Guided by validation feedback, they iteratively refine data, trainers, and inference harnesses, then select and submit a final agent system for evaluation on the test set. To prevent reward hacking, we enforce three safeguards: test data isolation, separate research and grading containers, and network restrictions for research agents. Our evaluation results show that frontier research agents substantially improve target agent systems. For example, on SWE-bench Verified, GPT-6-Astra raises the target agent system's performance from 32.9% to 53.7%, comparable to the human-developed instruct model's 55.4%. However, there is still room to improve their ability to co-design and iteratively refine data, trainer, and harness to achieve cumulative performance gains. We hope that VibeTraining-Bench will serve as a unified platform for advancing RSI, deepening our understanding of cross-component optimization for agent systems. Code and data are available at https://github.com/D2I-ai/VibeTraining-Bench.

arXiv:2610.25496v2 Announce Type: replace Abstract: In many human-AI systems, candidate actions are not executed directly but are first evaluated by another decision-making entity under a potentially different objective or decision model. Standard decision-theoretic frameworks do not require this interaction to be represented explicitly, as they define transition dynamics directly over state-action pairs, allowing evaluation to be absorbed within the transition model. We introduce a semantics-aware decision framework in which a candidate action is proposed, evaluated under an interaction context, and then resolved into an executed outcome. We formalize this through a joint action language and an interpretation function that maps each joint action and context to a semantic outcome. Transition and reward dynamics are induced through these semantic outcomes, thereby introducing an explicit representational layer for interaction semantics. Semantic outcomes provide explicit units for organizing interaction principles, which can be instantiated through application-specific evaluation mechanisms and reused across distinct interaction instances. This provides a structured interface for specifying context-dependent evaluation separately from the resulting execution dynamics, without requiring changes to the underlying planning algorithm. We instantiate the framework in shared human-AI autonomy, capturing interaction patterns such as advisory and authorization, and illustrating how interaction context influences decision making. The resulting formulation remains compatible with standard Markov decision process (MDP) planning.

arXiv:2610.24611v2 Announce Type: replace Abstract: Evaluating software agents on a full benchmark is expensive. A small task subset can preserve agent rankings while yielding inaccurate scores. Historical responses provide useful structure, but tasks with identical source responses can have different outcomes for a new target agent. We introduce Hierarchical Score Posterior Recovery (HSPR), which combines source-fitted response factors with uncertain group success rates. Observed target outcomes update the posteriors over the target latent state and group rates; posterior-guided acquisition selects subsequent tasks. Its score-based criterion equals the expected one-step reduction in full-score posterior variance. Observed successes and posterior-predictive counts for unqueried tasks yield a full-score posterior. The final estimate uses its median with source-selected calibration. At a 10% task budget on four task banks from the SWE-bench family, HSPR achieves the lowest mean absolute error (MAE) on three banks and the lowest maximum absolute error (MaxErr) on all four among the methods compared. Removing group uncertainty or replacing adaptive queries increases macro MAE in the respective controls. Macro MAE decreases across 5-20% budgets. Code and project resources are publicly available.

arXiv:2610.24520v2 Announce Type: replace Abstract: In long-term interactions with large language model (LLM) agents, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversations. However, individual observations may provide only partial evidence about a habit or when it applies. Retaining past interactions alone does not ensure that agents can infer habits and apply them appropriately. We introduce HABIT-Bench to evaluate how well agent memory systems infer habits from weak evidence and apply them to later requests. We construct synthetic user histories from task-oriented dialogue data and controlled habit graphs. The histories span hundreds of sessions across food, finance, software, and travel. Four-choice tasks assess agents' ability to infer user habits and apply them appropriately. They include temporary exceptions, changes in habits, and suggestions that users have not accepted. Sampled tasks undergo independent human review. We evaluate eight memory methods with Qwen3-235B-A22B and GLM-5.3-Flash. The best-performing memory method achieves 33.61\% macro accuracy across the four domains. Performance remains limited even when annotated evidence is supplied directly. These results expose a persistent gap between access to past interactions and the ability to infer and appropriately use user habits.

arXiv:2610.24264v2 Announce Type: replace Abstract: As LLM-based agents are deployed with increasing autonomy in real-world environments, a critical failure mode emerges: agentic misalignment, where agents choose harmful actions over failure. However, existing safety benchmarks primarily evaluate whether agents comply with explicitly harmful instructions, leaving the conditions that cause agents to choose harmful actions poorly understood. To address this limitation, we propose AM-Bench, a unified taxonomy and 35 scenario evaluation benchmark spanning seven industries for agentic misalignment grounded in the fraud triangle, a well-established criminological framework explaining goal-directed misconduct. We adapt this framework to agentic environments and construct our evaluation suite and detection systems along three dimensions: perceived threat, perceived opportunity, and reasoning. We operationalize each dimension into structured subcategories rooted in prior theory: five threat categories derived from the instrumental convergence thesis, seven opportunity categories drawn from the Emergent Strategic Reasoning Risks (ESRR) taxonomy, and a novel four-question binary reasoning framework synthesized from prior literature. Our evaluation of 10 frontier models reveals a 62-percentage-point gap in misalignment across scenarios, with Gemini 3 Flash and Gemini 3.1 Pro exhibiting the highest levels of agentic misalignment while Claude Sonnet 4.6 displays near-zero rates across all conditions. Our reasoning framework further reveals that models may exhibit harmful behaviours without explicit situational awareness or explanation, complicating both mitigation and detection. AM-Bench scenarios and code are available at https://github.com/AI4Collaboration/Agentic-Misalignment-Bench

arXiv:2610.24161v2 Announce Type: replace Abstract: TorchTPU is an open-source collaboration between Google and Meta that introduces a native and principled PyTorch backend for Google TPUs. It enables seamless out-of-the-box model execution and TPU migration while delivering industry-competitive inference and training performance through eager execution optimizations, torch.compile integration, and native hardware acceleration.

arXiv:2610.23722v2 Announce Type: replace Abstract: Understanding and simulating how subjective cognition and mental states emerge, evolve, and shape behavior is essential for advancing artificial intelligence toward deeper social cognition and general intelligence. Existing large language models (LLMs) and Theory of Mind (ToM) methods primarily infer mental states from linguistic information, lacking unified mechanisms for subjective world construction, cognitive-psychological interactions, dynamic evolution, and historical dependencies. This paper introduces the Mental World Model (MWM) and its implementation, the Multi-Branch Mental Evolution Architecture (MMEA), which explicitly constructs subjective worlds, models interactions between cognition and mental states, and simulates continuous mental evolution shaped by historical experiences. On DynToM and RecToM, MMEA achieves 96.48% and 94.00% accuracy, respectively, outperforming leading LLMs and specialized ToM methods. Additional experiments demonstrate improvements of 28.7, 39.3, and 16.6 percentage points in mental state and behavior prediction, long-sequence prediction, and implicit mental state inference, respectively. The results demonstrate that MMEA provides an interpretable framework for continuous mental modeling, enabling AI to move beyond surface-level semantic representations to infer individuals' underlying mental states from linguistic information and dynamically simulate their internal cognitive processes and evolution over time. MWM also provides a promising foundation for causal reasoning, personalized AI, sustained human-AI interaction, and future advances toward artificial general intelligence (AGI).

arXiv:2610.20022v2 Announce Type: replace Abstract: Memory systems for persistent personal agents must move beyond isolated memories, maintaining both personal memory about whom they serve and procedural memory about how to act. Crucially, in harness-based personal assistance, procedures or skills are inherently user-conditioned: personal context determines which procedures apply, how they are instantiated, and when they should change. Existing benchmarks typically emphasize either personal or procedural memory, leaving their interaction underexplored. We introduce PraxisMemBench, a harness-based benchmark for persistent personal assistance across daily-life and work trajectories, where personal and procedural memory evolve together and later tasks depend on both. Experiments show that maintaining two memory types separately is insufficient, with completion dropping from single-type to interdependent dependencies (84.6% to 77.9%). More broadly, across existing methods, relation-related failures increase from early to late sessions as interaction history grows (59.8% to 72.9%), revealing a continual-maintenance challenge: writing correct memories is not enough if their evolving content and organization cannot be preserved for later retrieval. To address this, we propose PraxisMem, which uses privileged joint-state supervision to internalize the ability to maintain memory content and organization into the memory manager, with location-aware GRPO and dynamic residual-routed OPSD. PraxisMem reaches 40.78% overall Session Success Rate, compared with 20.60% for Claude Code's native memory. Our code is publicly available at https://github.com/QMMMS/PraxisMemBench.

arXiv:2610.19432v2 Announce Type: replace Abstract: Comparing financial disclosures across languages requires identifying shared accounting concepts and recovering all numerical facts that express them. Existing benchmarks evaluate aspects of these capabilities, but provide limited insight into whether both succeed on the same multilingual evidence. In particular, it remains unclear whether language models can recognize these concepts across languages and exhaustively recover their corresponding numerical evidence from the same contexts. We introduce MULTIFINGROUND, a benchmark designed to investigate this recognition-grounding gap in multilingual financial reports. Constructed from FY2025 disclosures of 90 issuers in English, French, and Greek, MULTIFINGROUND aligns independently authored reports through a shared accounting taxonomy and provides traceable fact-level annotations. The benchmark evaluates financial fact identification, concept recognition, exhaustive fact selection, and concept discrimination without candidate labels. Across 15 evaluated models, eight API and open-weight reference models achieve 97.3-99.4% concept recognition accuracy, yet miss at least one target fact in 26.6-41.9% of correctly recognized groups. Further analyses reveal substantial variation in concept discrimination and limited gains from the evaluated finance-specialized models. Our findings demonstrate that strong concept recognition can coexist with incomplete evidence recovery, highlighting the importance of evaluating semantic recognition and evidence completeness together.

arXiv:2610.18603v2 Announce Type: replace Abstract: Benchmarks report what an Artificial Intelligence (AI) agent can do on a task suite, but not whether one released version has been engineered with the controls that make it reliable, safe and observable, or how strongly that is shown. This paper proposes AgentStripes, an evidence-graded maturity model for AI agents. It defines levels (0 to 7 stripes), process areas (12 areas holding 40 capabilities), three gates (basic, safety, proof) that hold an agent at a level however high its score, and evidence grades E0 to E5. Each applicable capability receives a rating from 0 to 3 and a separate evidence level; these are combined into a credit, aggregated into a severity-weighted score and converted into stripes that the gates limit. Instruments gather evidence, language-model judges interpret it under explicit criteria, and the arithmetic is deterministic. The model is compared with standards at topic level without a claim of conformance, and an unexecuted validation protocol is specified. An empirical evaluation (exploratory, non-confirmatory) of 30 public Python agent repositories, with static evidence only, is a demonstration and not a validation: static evidence caps every final stripe at 2, 26 repositories end at one stripe and 4 at two, and this distinction is unreliable, since re-judging changed the final stripe of 4 of 22 repositories (engine, dossier and panel changed), whereas none of 8 panel-only repeats, all at stripe 1, changed. The parameters are expert conventions without outcome based calibration, and the stripes have not been shown to predict reliability or safety.

arXiv:2610.18447v3 Announce Type: replace Abstract: World models have emerged as a promising paradigm for embodied AI, yet existing systems often remain fragmented in the modalities they model and the functions they support. In particular, visual dynamics, geometric structure, and executable action are commonly coupled through separate interfaces or downstream modules rather than being jointly represented and evolved within a shared dynamics model. In this work, we present Thinker-WM, a unified multimodal world model that jointly models and generates RGB, metric depth, and robot action through a shared intermediate Unified Dynamics Transformer with modality-specific encoders and decoders. The model performs multimodal unification at the representation level: appearance, geometry, and action are transformed into a common dynamic representation and jointly evolved by shared parameters, while their distinct output structures are preserved. The same parameterization supports multiple functions, including joint video-action generation, world prediction, policy inference, action-conditioned generation, video-to-action generation, and inverse dynamics. To examine whether the shared representation captures action-related information, we introduce a lightweight motion probe that shows the frozen visual representation contains recoverable motion information. Thinker-WM is trained on a heterogeneous corpus comprising approximately 20K hours of egocentric video, 10K hours of real-robot data, and 1K hours of simulation data, including approximately 4K hours of real-robot data with aligned depth supervision. Extensive experiments demonstrate strong performance across embodied world-model evaluations. Thinker-WM ranks first on EWMBench, WorldOlympiad and WorldArena, while controlled studies further demonstrate the benefits of action conditioning, geometric interaction, and multi-function training within the shared representation.

arXiv:2610.17710v2 Announce Type: replace Abstract: Training graphical user interface (GUI) agents with reinforcement learning suffers from sparse outcome rewards, leading to inefficient learning. Process rewards offer a promising direction for credit assignment to intermediate actions; however, accurately evaluating an action requires identifying evidence across long-horizon GUI trajectories, as its value may depend on information beyond its immediate state and observation. In this paper, we propose an adaptive-group process reward framework that dynamically partitions a GUI trajectory into semantically coherent subgoal groups, enabling process reward estimation at both the action and group levels. At the action level, a critic evaluates each action based on the visual changes and subsequent actions within the group for action correctness and contribution to the subgoal. At the group level, the framework aggregates action-level evaluations to assess subgoal completion and contribution to the overall task, and further refines these assessments by leveraging the context of all groups in the trajectory. The process reward of each action is obtained by combining its action-level evaluation with the refined group-level evaluation. We further augment GRPO's outcome advantage with the proposed process rewards to provide finer-grained guidance for GUI agent training. Experiments on process reward benchmarks and downstream RL training demonstrate the effectiveness of the proposed framework.

arXiv:2610.14958v2 Announce Type: replace Abstract: Activation steering has recently been used as a memory for skills: past experiences of an agent are stored as steering vectors and retrieved later. Facts and preferences about the user are assumed to need text instead. We test that assumption with a training-free activation memory beside a frozen LLM. The change an experience causes in late-layer states is stored as a shift, filed under a whitened key for the conversation's gist, and added back in a later session only when the gist matches. We argue that such a shift is a fixed average of an effect that attention, in context, recomputes for every question, so only question-independent content can survive. In a pre-registered study on Qwen3.5-2B and 9B, with judges from four model families, the memory is selective (about 98% of unrelated replies are unchanged), but facts are stored as the word to say, in the last layers only, and are almost never used (at most 15% clean use, against 96-98% in context). Leanings transfer only in a narrow strength window. Most strikingly, the reference subtracted when writing a preference decides whether its concept or its direction survives: unless the opposite is subtracted, a stored dislike is recalled as a like.

arXiv:2610.14539v3 Announce Type: replace Abstract: Real-world business requests often leave key requirements unstated because frontline practitioners struggle to articulate rules they know from experience. Forward-deployed engineers (FDEs) must elicit these requirements through dialogue and iteratively optimize solutions until they meet customer acceptance criteria. Existing optimization benchmarks typically start from explicit objectives, while clarification benchmarks often introduce information gaps into existing tasks. We introduce FDE-Bench, the first benchmark built from real customer projects to evaluate agents on end-to-end delivery tasks that require interactive requirement clarification and iterative optimization. Its 49 cases span combinatorial optimization and machine learning. Clarification targets are distilled from discussions between frontline practitioners and FDEs. Agents elicit missing requirements from a simulated customer, and case-specific evaluators assess final artifacts against acceptance criteria using customer-accepted FDE solutions as quality references. Across 18 agent configurations, the highest mean delivery score is only 0.5510, compared with an FDE reference of 1.0. The highest delivery success rate (DSR), defined as the fraction of planned runs reaching this reference, is only 6.12%. These results show that reliable FDE-level end-to-end delivery remains out of reach for the evaluated agents on underspecified real-world business tasks. The benchmark project page is available at https://baidubce.github.io/FM-Agent/FDE-Bench/.

arXiv:2610.11723v4 Announce Type: replace Abstract: Multiplayer world models must ensure that independently controlled views remain consistent with a single shared and persistent world. We introduce MultiWorldBench, a diagnostic Minecraft benchmark comprising 495 case configurations across seven task suites and ten capabilities, including independent control, cross-view motion, shared-state synchronization, persistence, structural reasoning, concurrent interaction, and delayed revisit. Using Engine GT as a reference, we evaluate two multiplayer world models, Solaris (MP) and Gamma-World, together with a single-player generation variant, Solaris (SP). Among the generated systems, Gamma-World achieves the highest ten-capability average at 21.39, followed by Solaris (MP) at 20.88 and Solaris (SP) at 10.06, while Engine GT reaches 91.69. Gamma-World performs better on several control, shared-state, and revisit capabilities, whereas Solaris (MP) leads in cross-view motion and race-condition consistency. Nevertheless, all generated systems score at most 8.00 on state persistence and 1.33 on structural consistency, and none succeeds in spatial reasoning or building-identity preservation. Human pairwise evaluation yields the same overall system ranking and shows strong alignment with the automatic evaluation, with a mean dimension-level Spearman correlation of 0.95. These results show that current multiplayer world models still exhibit a substantial gap relative to Engine GT in maintaining a coherent shared world.

arXiv:2610.04699v2 Announce Type: replace Abstract: A verifier for an agent faces rules of two kinds: the ones a fixed check can settle and the ones that require a judge. A team that derives its own checks fixes that split up front. Where the requirements come from outside, as in finance, healthcare and law, the agent enforces rules it did not write, so the split falls to runtime, recurring for every predicate of every rule on every action at a rate no reviewer can audit. Every escalation scheme assumes a model can make that decision itself, that it is self-decidable. Across six corpora, including the EU AI Act, FINRA guidance and a deployed credit agent, we collect roughly 22,000 labels from four models built by three labs. They agree almost perfectly where the answer is obvious and collapse on regulatory text; their errors run in opposite directions, so no model can be trusted as the conservative choice; and on the deployed agent's own rule-set they err together, over-claiming that a fixed check will do, the direction that never gets escalated. We introduce CoVer (corroborate-then-verify), which treats unanimity as a nomination, admitting a predicate only when the check synthesized for it survives intervention, reading fields the agent cannot write and holding under deterministic rewording. That gate rejects most of what corroboration wrongly admits, at a cost in coverage we report rather than tune away. The obvious alternative, agreement with a reference judge, certifies nothing: it climbs from 30% to 77% across calibration bands while the genuinely decidable share does not move, because a judge drawn from the population under indictment ratifies the blind spot it shares. Self-decidability is not a capability to elicit from a model but a boundary the verifier must construct.