Artificial Intelligence
STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents
Guanhua Zhu, Hanwen Li, Zhi Wei et al. (+2 more)
Large
language model (LLM)
agents can self-evolve at test time by accumulating experience: retrieving relevant
memories before acting and writing new lessons back afterwards. Yet most are single-level: trajectories, distilled workflows and
reasoning items all land in one undifferentiated pool indexed by
embedding similarity, with no typed abstraction hierarchy. They never forget or deduplicate, and are prone to negative transfer, where a superficially similar but directionally opposite experience is recalled and its concrete steps are copied verbatim, actively harming the
agent. We introduce STRATA, a structured, outcome-grounded, self-maintaining
memory that organizes experience into a three-layer graph of atomic skills and code snippets, task workflows and generalized meta-frameworks, connected by typed edges. STRATA maintains itself through five mechanisms sharing one similarity geometry: a multi-dimensional similarity that scores
retrieval by operation and target; a typed reflector crystallizing procedures from successes and pitfalls from failures; community detection that induces meta-frameworks; pressure-triggered garbage collection with articulation-point protection; and contrast edges that flag confusable, direction-opposite experiences. Across four interactive
benchmarks (AppWorld, WebArena, ALFWorld, ScienceWorld), two frozen backbones and seven baselines, STRATA attains the best average score while simultaneously using the lowest latency and fewest environment steps of any
memory-augmented method, and 44% fewer lines of injected
memory than the most context-hungry baseline. We isolate two failure modes of prior
memories: single-level
memory degrades sharply as the task set grows, and single-level concrete
memories suffer catastrophic negative transfer, which contrast edges repair, lifting ScienceWorld from 78.7 to 96.7 within the same system.
Read the full abstract ▾
All authors Guanhua Zhu, Hanwen Li, Zhi Wei, Mengnan Du, Chenxi Yuan
Machine Learning
RA$^2$: Retain-Anchored Attraction Preserves Forget-Adjacent Utility in LLM Unlearning
Zinan Ling, Yang Xiao, Ruimeng Ye et al. (+2 more)
Large
Language Model (LLM) unlearning is often evaluated primarily by target suppression and clean retain utility. These metrics can miss a local failure mode: a model may perform well on standard retain
prompts, yet degrade on benign
reasoning when
prompts lie near the forgotten domain. This phenomenon is termed forget-adjacent utility loss. A fundamental question is: where should forget representations go after they leave the removed behavior? Successful unlearning depends not only on whether forget-conditioned representations depart from the harmful computation, but also on the direction in which they are redirected afterward. We propose RA$^2$ (Retain-Anchored Attraction), a latent-space unlearning method that redirects forget-
token states toward nearby retain-supported representations while preserving higher-layer behavior on retain data. On WMDP and MUSE, RA$^2$ yields higher forget-adjacent utility than strong unlearning baselines at similar broad-utility levels. These results show that the migration direction of forget representations materially affects benign behavior near the forget boundary. Our code is available at https://github.com/ZacLing/retain_anchored_attraction.
Read the full abstract ▾
All authors Zinan Ling, Yang Xiao, Ruimeng Ye, Sijia Liu, Bo Hui
Cryptography and Security
Yuke Wang, Benlei Cui, Shen Pang et al. (+7 more)
Multimodal agents increasingly rely on visual interfaces to perceive state and select actions, creating an attack surface not captured by text-only safety evaluations. We study Visual Semantic Hijacking, in which deceptive but semantically meaningful interface elements, such as forged system alerts, cause an
agent to depart from its authorized task. We propose a trajectory-guided procedure that identifies candidate injection points and inserts reproducible visual payloads into the observation stream. Using this procedure, we construct MMA-SafetyBench, which contains 1,083
adversarial trajectories across web automation, desktop GUI control, mobile navigation, document understanding, and
video processing. Across nine evaluated models, Attack Success Rate (ASR) reaches 87.96%, while Semantic Compromise Rate (SCR) reaches 98.09%. A 1–3–5 scoring scheme distinguishes target-matching outputs, semantic disruption without a target match, and successful preservation of the benign task. Controls covering all nine models and five domains show that attacks persist under neutral
prompts. A safety-priority
prompt lowers ASR without eliminating the failure mode.
Read the full abstract ▾
All authors Yuke Wang, Benlei Cui, Shen Pang, Xuemei Dong, Longtao Huang, Hui Xue, Yuwen Zhai, Junjie Li, Jingqun Tang, Haiwen Hong
Artificial Intelligence
ixi-sLM: On-Device sLM for AI Call Agent Service
Han-Sang Lee, Kyuho Lee, Juneyoung Park et al. (+10 more)
Running an AI call
agent on-device makes it private, network-independent, and free of per-query cloud cost, but shipping one as a real-world service imposes constraints a cloud deployment never faces: ~1GiB
memory budget, in-call latency and thermal limits, static-integer NPU execution, Korean STT input far from sLM pretraining, and cloud-level quality expectations. We bring ixi-sLM, a 1.2B Korean sLM, to mobile NPUs through ON-DEVICE MODELING: domain-adaptive modeling trims the vocabulary by call frequency to shrink the
embedding and LM head that take nearly half the model under INT4
transformer weights and refits the tokenizer to call
speech (18% fewer
tokens on call traffic), while task-aware quantization calibrates on real calls and assigns per-module mixed precision by activation sensitivity. A single domain-adapted model runs ixi-O's two production services, AI Call Summary and AI Call Answering, at 94% of a state-of-the-art cloud LLM's task-weighted quality; our on-device modeling carries that quality onto the NPU with only a marginal drop in overall score, in an 803MiB model (28% of FP16) at 23% of the CPU's per-call energy. The same recipe deploys across Qualcomm and Apple NPUs and serves several million ixi-O subscribers in production.
Read the full abstract ▾
All authors Han-Sang Lee, Kyuho Lee, Juneyoung Park, Seongbae Lee, Jin Kyu Park, Minjae Kim, Yuri Hong, Jinwoo Lee, Youngwook Kwon, Seongwan Kim, Seungheon Hyeon, Jaeho Lee, Kiduk Kwon
Artificial Intelligence
EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs
Guanhua Zhu, Bo Shen, Lei Zhang et al. (+4 more)
Different tasks often require interactive
agents to perform similar operations at intermediate steps. Retrieving an entire past workflow can introduce irrelevant steps, whereas retrieving isolated operations can discard the dependencies needed to reuse them. We introduce EvoWeave, a framework for experience
memory that represents abstract states as nodes and reusable operations as directed edges with input and output contracts, provenance, and capability references. A planner decomposes the current task into a dependency graph; operation
retrieval and bounded directed search then assemble relevant experience fragments. An execution
agent can inspect these fragments and their referenced capabilities, while reflection separates concrete execution evidence from reusable abstractions. We study the framework across four
benchmarks, five evaluation settings, and three
language model backbones. Across backbones, AppWorld Normal and Challenge task goal completion average 51.91 and 27.99, respectively, exceeding the baseline without
memory by 6.99 and 4.99 percentage points. An AppWorld ablation associates directed connections with an additional improvement of 2.08 points over independent operation cards. The results motivate composition of reusable operations while exposing tradeoffs in
retrieval overhead and conditions for valid reuse.
Read the full abstract ▾
All authors Guanhua Zhu, Bo Shen, Lei Zhang, Fadi Deek, Zhi Wei, Mengnan Du, Chenxi Yuan
Computation and Language
Learning What to Keep: Transfer-Aware Personal Skill Learning from User Feedback
Li Wang, Chen Zhang, Yirong Yang et al. (+4 more)
User feedback enables LLMs to refine reusable personalized skills, but often mixes persistent preferences with request-specific corrections. Indiscriminately retaining both risks negative transfer, motivating selective retention at the level of individual requirements. We frame this challenge as selective transfer over atomic rules and propose SkillGrain, a framework for learning reusable personalized skills from user feedback. SkillGrain decomposes feedback into independently actionable rules, predicts their transferability, and retains only those suitable for future requests. To learn these decisions, we combine rubric-guided skill rewards with fine-grained credit assignment that separately evaluates rule content and retention decisions. We further jointly optimize skill induction and skill-conditioned generation through reinforcement learning, ensuring that learned skills support downstream personalization. We construct a
benchmark spanning five personalized writing tasks to evaluate cross-request generalization. Experiments across three base models in both in-domain and out-of-domain settings show that SkillGrain achieves the best overall average performance among competing methods.
Read the full abstract ▾
All authors Li Wang, Chen Zhang, Yirong Yang, Xingyu Chen, Jinkun Zhao, Tianhao Peng, Wenjun Wu
Machine Learning
Robust Explanatory Analysis for Structured Optimization Networks (REASON) in Learning
Vahid Jebraeeli, Hamid Krim and Keisuke Fujii
We seek a hierarchical resolution of residual error under a composite learning objective as a principled mechanism for identifying and exploiting additional information, particularly under distribution shift and perturbation. Inspired by the System 1/System 2 paradigm, we formulate this principle as a two-loop mechanism serving a common inference goal. A fast, System 1-like loop resolves the residual addressable by the available evidence; the remaining epistemic residual triggers a medium speed, System 2-like loop that actively acquires additional evidence and measures its informational value through residual reduction. Over nested solution spaces, the medium-loop optimum is provably no worse than the fast-loop optimum. Experiments show that the resulting two-loop model consistently improves upon the fast loop at approximately half the computation of the medium loop, including under unseen corruptions, and, when provided a cleaner observation, outperforms staged classifiers and learning-to-defer methods at equal computational cost.
Read the full abstract ▾
Artificial Intelligence
DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents
Tianle Chen, Pushkar Singh, Ketan Todi et al. (+8 more)
Language-model
agents that run long tasks often redo work they have already done: they re-issue a tool call they made earlier and get back exactly the same observation, a step that adds no new information. We identify this failure as a loop. Loops appear natively in Qwen, Llama, and Gemini
agents and become far more common once the
agent's history is compressed to fit its context window. To reduce these loops, we propose DeLoop, a self-supervised training method. DeLoop builds on a simple observation: the
agent's own rollout can provide the supervision needed to learn from its loops. We use a looping action as a negative example and the first later non-looping action in the same rollout as a positive example. DeLoop learns from these pairs through direct preference optimization and does not require task rewards, gold answers, or LLM judges. On CodeGym under compression, DeLoop reduces the share of Qwen3-8B tasks containing loops from 57.3% to 6.1%. It also uses about one third as many steps per task, with similar task success. On Qwen3-32B, DeLoop reduces loops and steps while raising task success from 37.7% to 44.8%.
Read the full abstract ▾
All authors Tianle Chen, Pushkar Singh, Ketan Todi, Biswajit Paria, Qijia He, Wenqi Wang, Yuwen Tan, Jike Zhong, Sugato Basu, Deepti Ghadiyaram, Arjun Akula
Machine Learning
One-Step Prediction Fidelity Is a Poor Proxy for Planning Performance in JEPA World Models
Joshua Cheung and Zhang He
Latent world models are routinely selected, checkpointed and ablated on one-step prediction loss, on the assumption that a model which predicts the next latent better is a better model. In model-based RL this assumption is known to be unreliable; we ask whether it also fails for latent world models, where the training target is the model's own
embedding and the loss is not expressed in interpretable units. We introduce a frozen-decoder protocol reporting one-step error in physical units, decomposed into the part already present in the encoder's representation and the part the predictor adds. Applying it to a recent end-to-end JEPA world model across three continuous-control environments, we move the metric by fine-tuning and measure goal-conditioned planning on the same checkpoints, with three training seeds and paired planning seeds throughout. Within one environment, a per-epoch ladder shows the relationship saturating after a single epoch: the first epoch improves fidelity by 32% and planning by 4.4 pp, but across the fifteen checkpoints that follow – a further 15% fidelity improvement, monotone in every training seed – the correlation between one-step error and planning success is r = -0.010. This is not a task-ceiling artifact: stratifying by a model-independent measure of training-data support, planning in the lowest-support stratum starts at 80%, leaving ample headroom, and is nonetheless flat from epoch 1 onward. Across environments the exchange rate is equally unstable: a 93% fidelity reduction on TwoRoom yields -0.7 pp of planning, while a 40% reduction on PushT yields +4.9 pp [+1.1, +8.9]. All results replicate under a decoder-free measure, latent drift. We also show that a matched continued-training control absorbs the entire fidelity gain we had attributed to an auxiliary objective.
Read the full abstract ▾
Systems and Control
A Critical Review of Grid Integration Challenges and Flexibility Opportunities for Large-Scale Data Centers
Bibek Dhakal, Manohar Chamana, Biswambhar Gayen et al. (+8 more)
The rapid proliferation of cloud computing and artificial intelligence workloads is fundamentally reshaping the magnitude, spatial concentration, and electrical characteristics of data center power demand. Data centers differ from conventional commercial loads because they combine high rack densities, converter-dominated interfaces, thermally coupled cooling, uninterruptible power supplies, programmable workloads, and increasingly complex behind-the-meter generation and storage within a single facility. This review examines their grid integration from utility and sustainable-energy perspectives across resource planning, large-load interconnection, distribution feeders,
market operation, real-time dynamics, power quality, and asset management. Particular
attention is given to distribution-level issues that receive limited treatment in broader energy reviews: bidirectional operation on renewable-rich feeders, harmonic propagation and resonance, ride-through and reconnection,
transformer derating, cable ampacity, neutral-conductor heating, protection, and measurement. Static, composite, physics-based dynamic, calibrated, impedance, electromagnetic transient, and hybrid load models are compared according to study purpose and data requirements, and utility coordination platforms, uninterruptible power supply and battery control, workload flexibility, renewable co-location, cogeneration, fuel
cells, thermal storage, and clean firm supply are assessed. The central finding is that requested peak demand and annual energy matching alone cannot establish reliable or sustainable integration. Credible assessment must consider hourly and locational supply, network deliverability, marginal emissions, lifecycle and water impacts, backup obligations, rebound, asset health, and verified operating capability.
Read the full abstract ▾
All authors Bibek Dhakal, Manohar Chamana, Biswambhar Gayen, Jatendar Kumar, S M Shams Rahman, Aaqib Peerzada, Ki Yeob Lee, Olatunji M. Adeyanju, Bhaskar Mitra, Soumya Kundu, Stephen Bayne
Computation and Language
Evaluating Compressed Memory: Dependence, Utility, and Content Benefit
Minsu Oh, Seunghui Jwa, Chanjun Park et al. (+1 more)
Does a
language model answer better because its compressed
memory contains useful document information? Replacing
memory vectors with zeros can lower a trained model's score without showing that it outperforms a model trained without
memory. We combine zeroing with separately trained models that cannot access stored text or receive the same constant input for every document. Across nine systems, some models rely on
memory yet perform worse than the trained alternative. For example, in experiments that predict states in generated stories, zeroing
memory worsens a model's prediction loss, yet a model trained without
memory achieves lower loss than the original model with intact
memory. On TriviaQA, giving xRAG
memory from another question's passage raises F1, which measures word overlap with accepted answers, from 19.38 without a document to 24.78. Its effect on how often outputs contain an accepted answer is statistically uncertain. The question's retrieved passage scores higher still (29.88 F1). Thus a gain over no document alone does not establish that the gain requires the question's passage. Evaluations must therefore specify the comparison model, absolute performance, and generation and scoring conditions.
Read the full abstract ▾
All authors Minsu Oh, Seunghui Jwa, Chanjun Park, Yeo-Chan Yoon
Machine Learning
Policy-Verified Self-Training: Learning Agent Compliance from Verifiable Policy Rewards
Min-Young Yu
Runtime policy verifiers can prevent
agents from executing disallowed actions, but their decisions are typically used only for enforcement and discarded afterward. We ask whether these decisions can instead become verifiable policy rewards for training the
agents they guard. We introduce Policy-Verified Self-Training (PVST), a post-training paradigm in which an
agent generates its own interaction trajectories and an executable policy verifier, here a deterministic gate compiled from a written policy, evaluates whether each action complies. Compliant trajectories are reinforced, and violations become negatives for preference optimization; beyond the gate and the
benchmark's task-success signal, PVST needs no stronger teacher model, no human preference labels and no attack-specific supervision, although rollouts are collected in the
benchmark's attacked environment. On held-out AgentDojo tasks under
prompt injection, policy-verified preference training of Qwen3-8B reduces violating trajectories from 37.2% to 10.3 $\pm$ 0.9% (three seeds) and successful attacks from 12.4% to 0.5% at equal or higher pooled task utility, although utility falls on travel, compared with 22.0% when negatives are selected by an attack oracle and 31.9 $\pm$ 0.9% with negatives that contain no violation. The learned behavior holds under an attack template never used in training (37.7% to 11.0%) and on a suite excluded from training (34.8% to 16.7%), and the retained verifier's intervention rate falls from 39.2% to 8.8% (seed 0). Our analysis identifies two conditions under which the verdict is most informative: success must not already imply compliance, and guarded actions must appear in both permitted and forbidden contexts.
Read the full abstract ▾
Computer Vision
The Impact of Detector Architecture, Model Scale, and Scene Density on Multi-Pedestrian Tracking
Pedro Manoel Herminio Alves, Dalton Cézane Gomes Valadares and Cláudio de Souza Baptista
Pedestrian tracking in
video is important for applications such as surveillance and scene analysis, but it becomes more challenging in crowded scenarios with frequent occlusions and rapid motion. This paper analyzes how detector architecture, model scale, scene density, and inference regime affect the tracking of multiple pedestrians. We compare the YOLO11, RT-DETR, and YOLO26 detector families with the ByteTrack and BoT-SORT trackers on sequences from the MOT17, MOT20, and SOMPT22
benchmarks. For YOLO26, we also compare traditional overlap filtering via Non-Maximum Suppression (NMS) with end-to-end inference. The RT-DETR family achieved the highest mean scores on the HOTA tracking quality metric across all three
benchmarks with both trackers, with a more pronounced advantage on MOT20, albeit at a higher computational cost. On MOT17 and SOMPT22, YOLO26 variants, especially YOLO26-L with NMS, offered favorable combinations of quality and speed compared with RT-DETR. In the YOLO26 inference comparison, the end-to-end mode did not yield consistent tracking gains over the NMS-based regime. Taken together, these results provide a practical guide for selecting detectors and trackers for this task based on application priorities.
Read the full abstract ▾
Software Engineering
OS-Bench: Evaluating SWE Agents on Cross-OS Code Migration
Qirui Jin, Kenan Li, Liao Zhu et al. (+8 more)
Language-model (LM)
agents are increasingly evaluated on repository-level software engineering (SWE)
benchmarks derived from GitHub issues and pull requests. Although these
benchmarks cover a broad range of development tasks, they provide limited insight into a common yet technically distinct challenge: migrating an existing implementation from one operating system (OS) to another while preserving behavior on both. We introduce OS-Bench, an executable
benchmark for cross-OS code migration between Linux and Windows. Motivated by case studies showing that cross-platform features are frequently implemented on one OS before support is added for another, OS-Bench identifies migration opportunities by comparing execution coverage collected from the same repository on both systems. Platform-specific regions are then verified, organized into five migration scenarios, and transformed into repository-level tasks with target-platform tests and source-platform regression tests. The resulting
benchmark contains 140 validated tasks from 111 repositories across C, C++, C#, Go, and Rust, comprising 91 Windows-to-Linux and 49 Linux-to-Windows migrations. We evaluate six models with SWE-
agent, Claude Code, OpenHands, and Win-
agent. The best configurations achieve only 26.4% success on Linux-target tasks and 20.4% on Windows-target tasks while preserving correctness across both systems. These results indicate that cross-OS migration is a challenging and underexplored capability, and OS-Bench is a testbed for improving platform-aware SWE
agents.
Read the full abstract ▾
All authors Qirui Jin, Kenan Li, Liao Zhu, Lingching Tung, Yikai Zhang, Yijia Wu, Zijian Jin, Pengfei Gao, Yu Kang, Qingwei Lin, Wenke Lee
Computation and Language
Beyond Corrected Actions: Training LLM Agents on Textual Critiques
Thomas Zeng, Zelei Cheng, Mingyang Zhou et al. (+3 more)
Large
language models (LLMs) are increasingly used for multi-turn agentic tasks that require interactions with users and environments over extended sequences of actions. A common approach to improving LLM
agents is to use more capable expert models to provide training supervision, either by behavior cloning the expert's trajectories or by using the expert to correct the actions in trajectories generated by the weaker student model. However, expert LLMs can provide richer supervision than actions alone: they can directly generate textual critique that explain why specific actions are incorrect. We study whether training on such feedback benefits LLM
agents and propose Critique-Augmented Agentic Fine-Tuning (CAFT), a framework that elicits textual expert feedback on student-generated trajectories. For each incorrect action, the expert yields both a corrected action and a textual critique explaining the student's mistake–which we then use as the training target for fine-tuning. Across four student models from the Nemotron 3 and Gemma 4 families on the agentic domains $\tau^2$-Bench Airline+Retail and WebShop, adding critiques to action-level supervision improves performance by 8% on average over action-only
distillation. Additionally, CAFT checkpoints can also provide a better warm start for reinforcement learning than expert-trajectory
distillation. Our results demonstrate that textual critiques can strengthen supervision for multi-turn LLM
agents.
Read the full abstract ▾
All authors Thomas Zeng, Zelei Cheng, Mingyang Zhou, Sambit Sahu, William Campbell, Genta Indra Winata
Artificial Intelligence
Minghao Guo, Qingyue Jiao, Xiujin Liu et al. (+77 more)
Multimodal models and
agents have emerged to enable AI systems to learn from and interact with the real world more like humans by integrating information across multiple modalities. However, enabling
multimodal agents to continually acquire useful information from complex, heterogeneous
multimodal data remains a significant challenge as they move toward long-horizon tasks and lifelong settings. Addressing this challenge requires effective
memory mechanisms that can maintain continuity across sessions, adapt to changing users and environments, and refine
memory-management policies through accumulated experience. We introduce a two-axis analytical framework for
multimodal agent memory structured around two key questions: how
multimodal information is formed and represented (Axis I, Formation and Representation), and how its management adapts with accumulated experience (Axis II,
Memory Management). Axis I, Formation and Representation, characterizes what information is retained and how it is encoded (e.g., modality-specific
embeddings, textual and structured records, latent and parametric representations) to support subsequent
reasoning and action. Axis II,
Memory Management, distinguishes three tiers: T1 Static
Memory reuses fixed
memory; T2 Dynamic
Memory updates
memory through a stable policy; and T3 Self-Evolving
Memory updates
memory through an evolving policy. Based on this framework, we synthesize recurring design patterns and derive principles for balancing source fidelity with abstraction, maintaining temporal validity and provenance, grounding
memory use in recoverable evidence, and evolving
memory policies over time. Furthermore, we discuss the challenges and potential future directions for the
memory mechanism of
multimodal agents.
Read the full abstract ▾
All authors Minghao Guo, Qingyue Jiao, Xiujin Liu, Jianhong Pang, Ruimeng Yang, Naihao Xu, Liwei Che, Haochen Xue, Ziqi Wang, Xinying Cai, Jiaxuan Luo, Jiaxin Deng, Xuesong Ye, Jintao Qu, Wenyao Gao, Jingwen Gu, Chenyang Zhao, Shichun Liu, Guibin Zhang, Zefang Liu, Siyin Wang, Wujiang Xu, Danrui Li, Zeru Shi, Woongyeong Yeo, Shuang Liu, Wei-Chieh Huang, Weizhi Zhang, Tianyu Yang, Kangsan Kim, Yebowen Hu, Bowen Wen, Bowen Jiang, Haohao Su, Ming Yin, Sheldon Yu, Mian Zhang, Yuyan Chen, Hanrong Zhang, Kai Mei, Zihao Zhang, Mingyu Jin, Xi Zhu, Hengyi Wang, Tianxin Wei, Linyong Nan, Zhenwen Liang, Zhendong Chu, Ziqi Huang, Yongye Su, Sixun Dong, Shujian Liu, Kaiqu Liang, Jiaojiao Han, Zhaoye Fei, Kun Wang, Shihan Dou, Zhichun Guo, Xingjian Diao, Zeqi Gu, Heng Wang, Tao Feng, Cheng Qian, Hamid Palangi, Silei Xu, Xi Yun, Yi Zhu, Shilong Liu, Qingsong Wen, Sung Ju Hwang, Zuxuan Wu, Yu-Gang Jiang, Yiyu Shi, Dimitris N. Metaxas, Jiaxuan You, Julian McAuley, Mengdi Wang, Jiawei Han, Philip S. Yu, Ziyi Ye
Artificial Intelligence
DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment
Tianzhuo Yang, Shaoheng Yan, Lin Lu et al. (+5 more)
Safety
alignment teaches
language models to refuse harmful requests, yet the internal accessibility of hazardous concepts can retain traces of earlier training. Can checkpoints with equally safe behavior differ in what remains accessible behind their refusals? In this work, we investigate this question through paired training histories and a direct audit of intermediate representations. We introduce DANGERLENS, a probe that maps hidden states into vocabulary space and uses blinded semantic judgments to distinguish harmful goals and operational means from hazard recognition and safety control. Experiments on three Qwen1.5 scales show that checkpoints with prior harmful exposure generally retain greater hazardous concept accessibility after the same safety
alignment, with the strongest separation in intermediate and later layers. The same harmful update also produces different behavioral trajectories from closely matched starting points. Human relevance validation and readout ablations reinforce the measured accessibility contrast. Our findings show that training history remains visible inside behaviorally matched models and motivate internal accessibility audits alongside output safety evaluations.
Read the full abstract ▾
All authors Tianzhuo Yang, Shaoheng Yan, Lin Lu, Jiawei Chen, Kaile Wang, Fanyu Meng, Minglu Liu, Yaodong Yang
Machine Learning
Dynamic Graph World Modelling: Learning Future Dynamics from Temporal Context
Bo Li, Yue Yang, Yujin Huang et al. (+6 more)
Dynamic graph learning models how entities and their relations evolve from historical observations to predict future graph states. Existing methods do not fully address dynamic graph world modelling, which requires learning transitions from temporal context to future latent graph states. This is difficult for three reasons: nodes and relations rely on different historical evidence, structural
evolution is path-dependent, and future snapshots provide sparse and realization-specific targets. We propose Dynamic Graph World Modelling (DyG-WM), the first framework to jointly predict future node and relation representations from temporal context by answering three questions: what has evolved, how it evolved, and what will evolve next. To capture what has evolved, a historical node-relation context encoder summarizes node-neighborhood trajectories and pair-specific histories. To characterize how it evolved, a node-relation path-signature encoder captures ordered structural changes. To predict what will evolve next, a dual-level joint-
embedding objective predicts target-snapshot node and relation representations without reconstructing the raw graph. Experiments on 12 link-prediction and two node-classification
datasets demonstrate the effectiveness of DyG-WM across both tasks.
Read the full abstract ▾
All authors Bo Li, Yue Yang, Yujin Huang, Siao Liu, David Lemphers, Sara E Rodriguez Gomez, Luke Borgnolo, Can Wang, Xin Zheng
Computation and Language
Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems
Taro Yano, Yoichi Ishibashi and Masafumi Oyamada
Best-of-N sampling is a simple inference-time scaling strategy for large
language models, yet its behavior under incomplete reward models remains poorly understood. We study Best-of-N as a competition between the extreme reward tails of correct and incorrect responses. This view shows that success depends not only on the probability of sampling a correct response, but also on whether correct responses dominate incorrect ones in the upper tail. Using extreme value theory, we characterize three asymptotic regimes: correct-tail-dominant, incorrect-tail-dominant, and critical explaining both monotonic gains and reward-hacking-induced degradation. We further derive finite-sample scaling laws for representative reward distributions, showing that, under Gaussian rewards, incorrect responses with sufficiently large variance can dominate selection even when their mean reward is lower. Unlike oracle selection, convergence with an incomplete selector can be
polynomial rather than exponential. Finally, we propose a lightweight tail-tension criterion to estimate when additional sampling may become harmful. Experiments on synthetic data and real reward model scores are consistent with the theoretically predicted regimes, scaling laws, and diagnostic behavior, suggesting that tail competition provides a unified lens on Best-of-N sampling for verifiable problems.
Read the full abstract ▾
Computer Vision
VLX-Go: Robust Embodied Tracking through Multi-Stage Learning and On-Policy Refinement
Kyusong Lee, Heting Ying and Tiancheng Zhao
Embodied visual tracking requires a
robot to identify a designated target from visual observations and control its own motion to follow that target through an environment. Following an isolated person is a comparatively simple demonstration of embodied visual tracking, but following the intended person among distractors remains challenging. Existing methods exhibit a substantial drop in success from single-target to distractor tracking, motivating a focus on the shared spaces that person-following
robots must navigate. We present \method{}, a
vision–
language waypoint policy built around the hypothesis that visual and
language representations should be aligned before tracking-specific motion is learned. The policy combines complementary frozen visual encoders, multiscale temporal context, a
transformer backbone, and a lightweight head for
robot-centric waypoint prediction. Its training progresses through
vision–
language alignment,
multimodal reasoning mid-training, and tracking specialization on 19,532 oracle episodes. Compared with 41.54% distractor-tracking success for our earlier OmTrackVLA model, the complete multi-stage recipe achieves 55.80%. We then explore on-policy refinement using filtered oracle labels for states visited by the policy. This preliminary study adds 2.42 percentage points of distractor-tracking success, reaching 58.22%, while reducing collision rate by 1.99 pts. The final policy records the highest distractor- and ambiguous-tracking success rates among the compared methods. These results support staged
multimodal learning as a promising direction for tracking among distractors and motivate further study of on-policy refinement for person following in shared spaces.
Read the full abstract ▾
Artificial Intelligence
Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering
Weilun Xu and Zipeng Sun
Activation steering adds a vector to a model's hidden states to control a named behaviour. A gain on that behaviour, however, does not show that the vector controls it directly: it may act through a component shared with neighbouring behaviours, which unsteered and random baselines cannot separate. We introduce the competing-target test: a vector is compared with the vector built the same way for a competing target, and an equivalence test states when their difference is below a useful margin. We apply it in open-weight models to vectors for Kantian deontology, which judges an act by the rule it follows, and act utilitarianism, which judges it by the welfare it produces. On held-out situations, prompting a framework raises its score and lowers the other's, whereas vectors for either framework move both scores alike, and none reaches the margin. The two vectors share most of their direction, a request to apply some framework; the same construction separates an optimistic from a pessimistic tone, whose vectors share much less. What a vector shares with its competitors therefore sets what it controls: steering a perspective can deliver a mixture, and measuring a disposition can measure its neighbours, unless tested against the competing target.
Read the full abstract ▾
Machine Learning
Katy Blumer, Da-Cheng Juan, Jon Kleinberg et al. (+3 more)
In this paper we study the behavior of
transformer networks generating text over long periods of time. We show that even a three-
token, single-layer
attention-only
transformer can fail to converge. We then study a rich class of
transformers with arbitrary finite depth, layer normalization, MLPs, RoPE encodings, and an arbitrary mix of local and global layers. Every
transformer in this class has a softmax temperature threshold above which its empirical cache distribution converges to a unique limit, independent of initial conditions. We show further that, while each
transformer has a temperature threshold, the supremum of this set is infinite, so there is no global minimal temperature that forces all
transformers to converge.
Read the full abstract ▾
All authors Katy Blumer, Da-Cheng Juan, Jon Kleinberg, Ravi Kumar, Cyrus Rashtchian, Andrew Tomkins
Numerical Analysis
Discrete Edge Multiscale Space based Hybrid Schwarz Preconditioner for High-Frequency Helmholtz Problems
Shubin Fu, Shihua Gong, Guanglian Li et al. (+1 more)
In this work, we develop a novel hybrid Schwarz method, termed as edge multiscale space based hybrid Schwarz (EMs-HS), for solving the Helmholtz problem with large wavenumbers. The problem is discretized using $H^1$-conforming nodal finite element methods on meshes of size $h$ decreasing faster than $k^{-3/2}$ such that the discretization error can be bounded by the data uniformly with respect to $k$. EMs-HS consists of a one-level Schwarz preconditioner (RAS-imp) and a coarse solver in a multiplicative way. The RAS-imp preconditioner solves local problems on overlapping subdomains with impedance boundary conditions in parallel, and combines the local solutions using partition of unity. On the PDE level, this coarse space corresponds to an edge multiscale space proposed in {[S. Fu, G. Li, R. Craster, and S. Guenneau, Multiscale Model. Simul., 19(4):1684-1709, 2021]}. The approximation properties of this coarse space to the discrete solutions from the primal and adjoint problems are derived by a novel conforming Galerkin finite element method based upon a nonstandard variational formulation. This work represents the first theoretical analysis of an edge multiscale space at the discrete level. We prove that the EMs-HS preconditioner leads to a convergent fixed-point iteration uniformly for large wavenumbers, by rigorously analyzing the approximation properties of the coarse space to the global Helmholtz harmonic extension part and to the solution of the adjoint problem. We provide extensive numerical results on the sharpness of the theoretical findings and also demonstrate the method on challenging heterogeneous models.
Read the full abstract ▾
All authors Shubin Fu, Shihua Gong, Guanglian Li, Yueqi Wang
Machine Learning
BFHD: Bidirectional Feature Harmonization Decomposition for Heterogeneous Clinical Assessments
Yuanhao Zhuo, Zixi Qin, Ling Qin et al. (+1 more)
Clinical assessments are often collected using heterogeneous assessment systems across centers and time, leading to records that mix different but related sets of measurements. This motivates harmonization beyond total-score linking. We formulate clinical harmonization at the measurement level as a bidirectional recoverability problem: given paired observations from two assessment systems, the goal is to identify which measurements can be reliably translated in both directions within an application-defined tolerance, while separating non-translatable components. We propose Bidirectional Feature Harmonization Decomposition (BFHD), a feasibility-driven framework that enforces bidirectionally coupled translation and uses feature-wise output gating to produce an explicit decomposition in the original measurement space. Experiments on synthetic data and real clinical assessment pairs show that BFHD achieves broader feasible harmonization coverage and improved subset stability compared to baselines.
Read the full abstract ▾
All authors Yuanhao Zhuo, Zixi Qin, Ling Qin, Wanqing Li
Machine Learning
Where to Start, How to Move: Analyzing Curricula under GRPO for Long-Horizon LLM Agents
Minchan Kim, Sungho Bae, Youngjin Kim et al. (+1 more)
Group Relative Policy Optimization (GRPO) provides no reward-relative learning signal when all rollouts receive identical rewards. In long-horizon tasks that require multiple interactions, rewards are given only upon successful task completion, making all-fail groups common early in training and limiting opportunities to practice later steps. Guide Assisted GRPO replays a successful reference prefix before the learner continues. We study how initial positions, handoff units, and movement rules affect unassisted task completion. Across three model sizes on BFCL, starting near completion is more effective under a fixed turn-level threshold rule. Finer handoffs alone do not consistently help, but outcome-informed tool-call selection outperforms GRPO on trained and untrained task categories. Guided starts also improve AppWorld success on both evaluation splits. These findings highlight where learning starts and how starting positions move as key curriculum choices.
Read the full abstract ▾
All authors Minchan Kim, Sungho Bae, Youngjin Kim, Hyunwoo Park
Machine Learning
Integrated Novelty Detection Systems: From Sensor Networks to Multi-Novelty Computer Vision
Irené Tematelewo
Novelty detection is crucial in various technological and scientific domains. Its importance spans from ensuring the reliability of sensor networks to enhancing the adaptability of computer
vision systems. As AI systems become increasingly sophisticated and complex, effective anomaly detection is essential for maintaining reliability and accuracy. This work addresses novelty detection in weather sensor networks and computer
vision. In weather sensor networks, particularly for monitoring rainfall, detecting blocked rain gauges can significantly impact data reliability, food security, and disaster management. This work evaluates four existing quality control methods for detecting blocked rain gauges and demonstrates their limitations, particularly in detecting clusters of faulty sensors. By comparing methods such as the Conditional Mixture Model (CMM), PSWQ (Fz-filter), Kernel Regression, and Gaussian Method, we identify their shortcomings and propose three novel quality control techniques that integrate probabilistic fault detection and consistency-based diagnosis framework to improve timely detection of blocked sensors, while reducing false alarms. In the realm of computer
vision, current systems often fail to handle multiple types of novelty due to their focus on single-novelty scenarios. We highlight the need for integrated systems capable of handling multiple types of novelty. To support this, we introduce the Novel Snapshot Serengeti (NOVEL-SS)
dataset, derived from the Snapshot Serengeti camera trap data, which includes various forms of novelty. We also present the DCA (Detect, Characterize, and Accommodate) baseline system, designed to handle multiple novelty by detecting, characterizing, and accommodating them. This work aims to advance the robustness and adaptability of novelty detection systems across both weather sensor networks and computer
vision applications.
Read the full abstract ▾
Artificial Intelligence
When Is Smaller Greener? Right-Sizing Language Models for Agentic AI
Sajid Siraj and Maria Barbati
AI
agents use large
language models (LLMs) to plan and act over multiple steps, yet models are usually compared per call rather than per completed task. This paper asks: which model should an
agent use to complete tasks reliably while minimising energy, cost and time? We show that per-call comparisons can be misleading for agentic AI. Because a task succeeds only if all required steps succeed, small differences in per-step reliability can produce large differences in overall task success. Failed attempts also consume resources. We therefore evaluate models using energy, cost and time per successfully completed task, treating the completed task as the functional unit. On this basis, we develop a decision model for choosing LLMs in agentic workflows. The model applies to tasks with binary or graded outcomes and to settings with or without retries. We show that each resource has a crossover task length at which a smaller, less reliable model ceases to be the greener, cheaper or faster option. These crossovers divide tasks into three regimes: for short tasks smaller models are preferred, for long tasks larger models are preferred, and only between these extremes do preferences among objectives matter. We embed these results in a two-stage framework that separates efficiency from preference. A Pareto screen first removes models that are never competitive; a weighted decision is required only when a genuine trade-off remains. A pilot study of 15 open-weight models from five developers, ranging from 0.6B to 27B parameters, on two multi-step
agent tasks with measured GPU energy supports the predicted structure. The model with the lowest energy per completed task changes with task length and task type, and observed crossover points generally match those predicted from per-step measurements. These findings provide a practical framework for right-sizing
language models in agentic AI.
Read the full abstract ▾
Computation and Language
Target-Token Asymmetry Can Masquerade As Depth Specialization
Danlin Li, Miao Li, Xu Han et al. (+1 more)
Layer-ablation studies often perturb early or late layers of a
language model, compare the damage to prediction on two corpora, and read a larger early-layer effect on one corpus as early-layer specialization for its task. Such a comparison averages over the
tokens that each corpus asks the model to predict, and these target
tokens differ: math solutions, for example, ask for far more whitespace and numbers than encyclopedia prose. We show that this difference can create the appearance of depth specialization. We split a cross-corpus early–late damage interaction by target-
token type and recompute it under a target mixture shared by both corpora. Between GSM8K and Wikitext, a positive interaction replicates across five model instances. In the three fully decomposed models, only whitespace has a positive interaction. Removing whitespace reverses the pooled effect in all three, and matching the full target mixture reverses it in two. The diagnosis holds under position matching,
adversarial reassignment of
token types, and a continuous readout. The risk can also be anticipated from the tokenizer alone. A screen computed before any model runs places every held-out pair it scored in the region its label predicts, and on the pairs it rates at high or medium risk, removing whitespace reverses every positive pooled estimate. A smaller content-word effect one
token ahead follows the share of content targets that the unablated model already predicts. A cross-corpus depth effect therefore supports an interpretation of task specialization only after the target-
token mixture is reported, the effect is decomposed by
token type, and a mixture-matched summary keeps its sign.
Read the full abstract ▾
All authors Danlin Li, Miao Li, Xu Han, Jiecheng Lu
Machine Learning
Counterfactual Gate-Support Decomposition: Separating Selection from Redistribution in Training-Data Filtering
Hyojung Han, Jongmin Kim and Seung-Hun Jeon
A training-data quality gate changes both which examples survive within regions of the input space and how training mass is distributed across those regions, so a gate-versus-random comparison does not isolate these mechanisms. We introduce Counterfactual Gate-Support Decomposition (CGSD), a four-arm, budget-matched design that crosses gate eligibility with support distribution and decomposes the observed gate effect into within-support selection, between-support redistribution, and their interaction. The practical constraint is finite-pool realizability: we characterize mass and
token-assembly conditions under which the required arms can be constructed, together with a support-extinction case in which the ordinary design is undefined. On Korean-English and Korean-Japanese
retrieval, the mean within-support selection component is the only one that exceeds its calibrated null band (+29.7 and +5.6 holdout items; 30/30 positive replicates), decisively in Korean-English and marginally in Korean-Japanese; redistribution and interaction remain unresolved at the available power. The same pattern holds for a second, unrelated gate family – a model-confidence threshold rather than a deterministic verifier – judged against its own null (+27.3 items, 9/9 positive), so the result is not a property of one gate's mechanism. On a mathematics gate, zero-eligibility supports leave the ordinary design undefined at every budget, while the extended design is constructible only inside narrow budget bands that an exhaustive scan locates and that the natural operating point misses – the assembly condition separating a property of the pool from a limitation of the packer. CGSD therefore returns a decomposition, a constructible-but-underpowered design, or an explicit non-constructibility result, and says which; it prevents support redistribution from being silently interpreted as sample-level gate quality.
Read the full abstract ▾
Multiagent Systems
Why Do Agent Teams Fail Across Boundaries? A Failure Taxonomy for Agent-to-Agent Systems with Private States
Sarath Shekkizhar, Prathyusha Jwalapuram, Zixuan Ke et al. (+12 more)
Agents are increasingly interacting with other
agents that act for different principals (people or organizations on whose behalf they work), such as a buyer's personal
agent negotiating with an enterprise's seller
agent. In such
agent-to-
agent (A2A) scenarios, each
agent has its own environment and holds information and authority that the other cannot see. Moreover, the interests of each
agent may be aligned (cooperative), partly aligned (mixed-motive), or opposed (competitive). Existing multi-
agent system failure studies assume one orchestrator, shared workspace, or single principal. We ask how A2A systems fail when these assumptions are removed. We study more than 2,500 recorded runs of A2A systems built from different
agent harnesses and models, spanning 488 scenarios across eight families (legal, CRM, security, commerce, etc.). From these traces, we derive $A^2$-Failures, a taxonomy of 20 A2A failure modes in four categories: (i) discovery and delegation, (ii) grounding and record consistency, (iii) verification, and (iv) conduct and fidelity. Three findings hold across every harness, model, and scenario studied: (1) discovery and interaction are driven by the phrasing of the task rather than the availability of peer
agents; (2) no
agent verifies a peer's factual claim on its own initiative, so an A2A system inherits the reliability of its least reliable participant unless the protocol or mechanism performs the check; and (3) the aggregate outcome score credits the
agent completing the task while staying blind to the coordination, verification, and conduct failures beneath it, so A2A systems must be evaluated per
agent, against each
agent's own principal. We release the run logs behind the taxonomy at https://huggingface.co/
datasets/a2a-salesforce/a2-failures.
Read the full abstract ▾
All authors Sarath Shekkizhar, Prathyusha Jwalapuram, Zixuan Ke, Vaidehi Patil, Jiayu Wang, Srijan Bansal, Trong Huy Nguyen, Dingfeng Jiang, Dongqi Liu, Ruochen Zhao, Wei Lu, Giuseppe Carenini, Mohit Bansal, Silvio Savarese, Shafiq Joty