SearchAtlas is a framework that transforms raw search trajectories of large language model (LLM) agents into structured evidential query graphs, where edges capture how evidence is propagated from queries to the final answer. The automated parsing pipeline achieves a mean edge F1 of 86.0% against human-annotated graphs and remains consistent across repeated runs. Using SearchAtlas, the authors analyze five search agents on three benchmarks, uncovering systematic differences in search scale and evidence aggregation, and revealing process failures such as fragmented answer support, unmet question constraints, and unverified parametric knowledge that correlate strongly with incorrect answers.
By Jiacheng Sang, Mengyuan Li, Sanxing Chen, Yukun Huang, Yu Feng, Bhuwan Dhingra
The paper investigates K/V-cache interventions—transplanting a target-conditioned key/value trajectory into a source-persona generation—as a method for controlling persona in decoder-only language models. Experiments on Llama‑3.1‑8B across 13 configurations reveal that strong representation alignment (measured by V‑gap) does not guarantee behavioral persona expression, with only mid‑layer replacements achieving both alignment and lexical diversity. Position perturbations uniformly suppress persona expression, highlighting that representation similarity alone is insufficient to predict downstream behavior.
By Yu Sun, Mengyin Lu, Cong Feng, Guangming Lu, Huimin Han
The paper systematically studies how anonymizing input data affects large language models (LLMs). Five prominent LLMs were evaluated on eleven benchmarks, comparing performance on original versus pseudonymized inputs. Results show that anonymization generally degrades performance, with larger drops for more capable models and task-dependent effects; reversible anonymization preserves entity uniqueness better than irreversible redaction, and prompting about anonymization offers no benefit.
By Tobias Deu{\ss}er, Max Hahnb\"uck, Lorenz Sparrenberg, Tobias Uelwer, Christian Bauckhage, Rafet Sifa
The paper evaluates six automated jailbreak evaluators—HarmBench, JailbreakBench, JailbreakRadar, StrongReject, JADES, and JailMeter—using human-labeled data from JailbreakQR and JailMeter-Eva. It measures each evaluator’s agreement with human judgments, error types, and consistency across attack families, controlling for model-specific variation by using a shared LLM judge where needed. The study finds that JADES performs best overall, with HarmBench and StrongReject also showing strong performance.
By Yujie Mu
FreeFlow is a hierarchical transformer for optical flow estimation that eliminates traditional flow-specific inductive biases such as correlation volumes, feature warping, and iterative refinement. It relies on a single feed-forward encoder–decoder architecture that integrates window attention for local processing, shifted-window attention for cross-window communication, and global attention at reduced resolution. This design allows the model to scale naturally with capacity, achieving state-of-the-art performance on Sintel, KITTI-2015, and Spring benchmarks while remaining memory efficient at 1080p inference.
By Vladislav Bargatin, Alexander Yakovenko, Khaled Abud, Dmitriy Vatolin
MMGait is a large‑scale multi‑sensor benchmark that aligns visible, infrared, depth, LiDAR, and radar observations at the sequence level, enabling evaluation of single‑modal, cross‑modal, and multi‑modal gait recognition. The study shows that modality rankings shift with probe conditions, cross‑modal alignment remains challenging, and fusion can yield complementary gains. To address the scalability issue of training separate experts, the authors propose Omni‑Modal Gait Recognition and its implementation, OmniGait++, which unifies all recognition settings within a shared identity space using modality‑specific front ends, a shared encoder, and an anchor‑guided fusion module.
whyItMatters":"MMGait provides a common testbed for heterogeneous gait sensing and demonstrates that unified recognition across varying modality availability is feasible, offering a scalable alternative to task‑specific experts."
By Saihui Hou, Chenye Wang, Qingyuan Cai, Aoqi Li, Yongzhen Huang
TextAlign is a post‑training preference‑alignment framework that improves text rendering in large text‑to‑image generative models without changing the generator architecture. It uses a hierarchical vision‑language model to reward global, word, and glyph‑level accuracy, converting binary defect judgments into a scalar preference signal that can be optimized with GRPO or DPO. Experiments on FLUX.1‑dev and Z‑Image‑Turbo demonstrate higher OCR‑based text accuracy while preserving overall generation quality, outperforming several foundation and text‑rendering baselines.
By Mingxuan Cui, Jingpu Yang, Fengxian Ji, Qian Jiang, Zhecheng Shi, Jiaming Wang, Zirui Song, Zhuohan Xie, Fajri Koto, Xiuying Chen
The paper introduces Word-level Probability MIA (WPMIA), a black-box membership inference attack that estimates word-level generation probabilities via Monte Carlo sampling and local kernel smoothing, then aggregates them into a sequence-level likelihood estimator. By conditioning on different prefixes, WPMIA amplifies distributional differences between member and non-member texts, outperforming existing black-box baselines on open-source LLMs and achieving an average TPR@5%FPR of 42.0 on proprietary models such as GPT‑5‑Chat, Gemini‑2.5‑Flash, and Claude‑4.5‑Haiku.
By Shengjie Niu, Yeheng Ge, Jian Huang
PEARL is a task-aware framework that evaluates differentially private synthetic educational data by checking validity, privacy protection, predictive usefulness, and suitability for the intended educational task. In a study of 96 settings, only 12 datasets passed all PEARL checks, with many failures due to missing outcome groups or distorted learning activity order. Even datasets that met privacy and predictive-usefulness criteria sometimes exhibited fairness issues and failed to support knowledge-tracing models, indicating that privacy alone does not guarantee practical usefulness.
By Xianghui Meng, Yujing Zhang, Jionghao Lin
The paper introduces a posterior reweighting framework to explain and counter in-context learning jailbreaks in multimodal large language models. It models the model as switching between safe and harmful behavioral modes, interpreting prompt demonstrations as evidence that shifts the posterior. Using this view, the authors derive scaling laws for jailbreak effectiveness and propose a defense that injects benign counter‑evidence to suppress harmful drift while maintaining utility.
By Xu Zhang, Dev Mistry, Xiang Xu, Ren Wang
The paper surveys 25 studies that use explainable AI to compromise machine learning models, covering attacks such as model extraction, membership inference, and model inversion. It distinguishes between how explanations are obtained—through target releases, attacker-derived methods, secondary disclosure, privileged access, or global artifacts—and shows that explanations can lower extraction costs and reveal membership signals via statistics, recourse distance, and robustness. The authors compare threat models, signals, and defenses, concluding that no single explanation type is always unsafe and that protection must be tailored to the specific acquisition path and target asset.
By Abdullah Caglar Oksuz, Anisa Halimi, Erman Ayday
The paper introduces Cros, a risk‑constrained stopping layer for sequential clinical diagnosis agents that determines when to stop testing and make a diagnosis. Cros combines state‑wise error ranking, policy design on disjoint development splits, and exact tests of selective diagnostic error to provide finite‑sample guarantees. On a MIMIC‑derived abdominal‑pain benchmark, Cros achieves higher state‑error AUROC and lower selective error rates compared to baseline stopping methods, though its performance varies across development resplits.
By Yuexin Wu, Vasile Rus
LexAgentHallu is a new benchmark that profiles hallucinations in legal agents across multi-step interactions. It contains 3,414 instances spanning 17 legal categories and 6 task types, each annotated with a dual-layer taxonomy of 7 high-level and 27 fine-grained hallucination categories. The benchmark introduces fine-grained metrics to quantify and localize failures along an agent’s execution path, revealing patterns such as the Right-Answer-Wrong-Reason effect and clustered hallucination subclasses.
By Yujin Zhou, Mingxuan Zheng, Chuxue Cao, Huang Yidan, Jiale Chen, Yike Guo, Sirui Han
The paper introduces the concept of proof‑carrying cognition, aiming to close the verification gap in language‑model reasoning by using reality‑settled rewards. It presents a theoretical framework linking verifier‑gold correlation to compute‑capability trade‑offs, demonstrates that unsound verifiers degrade under best‑of‑N selection while sound verifiers improve, and proposes a new benchmark metric, Soundness‑under‑Pressure, for evaluating reality‑settled reasoning systems.
By Eshwar Reddy M, Sourav Karmakar
The Era by Eon Benchmark is a new dataset for evaluating large language model agents that interact with enterprise tools. It constructs a complete fictional company with product simulators, internal databases, and benchmark questions, all generated from a shared entity graph to ensure consistency. Exact answer keys are computed from the generated records, allowing precise grading and validation of realism and adversarial robustness across 23 simulated companies.
By Benjamin Gruenbaum, Doron Porat, Assaf Natanzon, Roy Zavida, Chen Dinachi, Or Itzahary
The paper introduces Time‑Frequency Geometric Cross‑Attention (TFGCA), a module that enhances vision‑language‑action models by decomposing action chunks into time‑frequency tokens using a learnable wavelet transform. TFGCA fuses dot‑product similarity with wedge‑product magnitude to better capture both frequency‑based smooth trends and cross‑phase orthogonal motion structures. When added to a pretrained VLA model, it yields significant performance gains across in‑distribution and out‑of‑distribution benchmarks, including a 28.5‑point improvement under RoboTwin domain randomization and an 11.67‑point increase on real‑robot AgiBot A2 tasks.
By Shengye Dong, Haochen Niu, Hao Liu, Peiwen Lin, Chuang Wang, Shanmin Pang
The paper introduces a reference‑based bias detection method that audits hidden‑state representations of language models by encoding sentences as similarities to a fixed set of anchor sentences. This relative representation allows comparison across model variants, such as before and after fine‑tuning, and yields a metric called Representational Bias Shift (ΔB). ΔB correlates strongly with output‑level bias changes, can detect bias‑increasing checkpoints with high ROC AUC, and is computationally efficient, requiring only a few minutes and far less compute than traditional benchmarks.
By Marek Jeli\'nski, Jan Dubi\'nski, Maciej Chrabaszcz, Sebastian Cygert
The paper introduces kernel‑managed shared memory, a system‑level abstraction that lets specialized agents write structured, tagged memories while the agent‑system kernel controls retrieval, privacy, and prompt injection. Implemented on AIOS, this design outperforms unmanaged external memory, standard retrieval‑augmented injection, and full context concatenation across GPT‑4o, Llama‑3.1:8B, and Qwen‑2.5:7B, improving personalization scores by 2.4‑4.0 points on a 5‑point scale and reducing latency and token usage by 15‑61%. The results show that centralizing memory management in the kernel delivers most personalization benefits at a fraction of the cost.
By Ryan Lum, Yongfeng Zhang
The paper investigates how incorporating physical-state (geometric) inputs influences a 0.8B hybrid language model with 6.2M trainable parameters for manipulation tasks. Training with geometry-conditioned recurrent decay gates achieves a 28.9% success rate, slightly lower than the 36.7% success when geometry increments are shuffled, and higher than the 24.4% success without explicit geometry. Robustness tests show that a state-only relative-coordinate policy retains most success under frame relabeling, whereas visual policies degrade significantly after small object displacements, indicating no clear advantage from training-time geometric alignment under the tested conditions.
By Hao Li, Haofei Sun, Lin He
The study investigates how speech‑to‑speech (S2S) models handle gender, distinguishing between the acoustic voice and the content’s gender cues. Experiments across five models in English, Spanish, and Mandarin show that while the rendered voice remains unbiased, the models consistently attribute speaker gender based on textual content rather than voice. When content and voice disagree, misgendering rates soar to 90%, whereas agreement yields only 2% misgendering.
By Xiaoqun Liu, Tanu Mitra, Harshit Rajgarhia, Abhishek Mukherji