ScopeBench is a new benchmark comprising 30 dead‑end agentic security tasks designed to test whether autonomous agents respect engagement boundaries under goal pressure. Each task is presented twice: once without a scope to measure raw hacking capability, and once with a natural‑language scope to assess scope adherence. The benchmark uses deterministic verification for scopeless runs and a calibrated agentic judge for scoped runs, revealing significant gaps between capability and adherence across eight tested models.
By Shane Caldwell, Max Harley, Ads Dawson, Michael Kouremetis, Vincent Abruzzo, Will Pearce
The paper introduces a synthetic ground‑truth framework for evaluating explainable AI (XAI) methods, addressing the lack of reliable evaluation procedures. By using controlled interventions to create datasets where the importance of input components is known, the framework generates ground‑truth explanations that align with the model’s actual decision process. The authors apply this approach to binary images, tabular data, and time series, and find that nine popular XAI methods exhibit significant limitations, underscoring the need for intervention‑based benchmarks.
By Miquel Mir\'o-Nicolau, Francesco Spinnato, Riccardo Guidotti
This paper introduces a new method for predicting transmembrane protein topology by employing the graph neural network SchNet. The model is trained on the same dataset used for DeepTMHMM, using 5‑fold cross‑validation, and incorporates all atom‑level embeddings rather than just sequence or alpha‑carbon features. Results indicate that GNNs hold significant promise for topological predictions without relying on pre‑trained weights.
By Sitong Chen, Xiaopeng Mao
The paper introduces a knowledge‑graph‑based evaluation framework, S3KG, to assess whether large language models truly understand context in question answering tasks. S3KG combines structural and semantic signals into a single similarity score and is paired with a diagnostic analysis that pinpoints reasoning errors at the triplet level. Across nine benchmarks, the method outperforms existing baselines, achieving up to +7.6 F1 points and an AUROC of 0.973.
By Subavarshana Arumugam, Mamta Nallaretnam, Kithuni Wickramasinghe, Chamath Gunapala, Pragatheeswaran Vipulanandan, Kamal Premaratne, Uthayasanker Thayasivam
BioEVAL is a global, multi‑institutional benchmark that evaluates large language and multimodal models on experimental reasoning tasks in bioengineering. It comprises 608 items across 11 bioengineering subfields, including 359 vetted multiple‑choice questions, 218 literature synthesis tasks, and 10 multimodal image‑interpretation problems. The benchmark was used to assess models such as ChatGPT, Gemini, and Grok, revealing accuracies up to 90% on MCQs, a 0.72 similarity score on literature synthesis, and 80% accuracy on multimodal tasks, with notable variation across subfields.
By Shun Ye, Vinny Chandran Suja, Chenlong Li, Chongming Jiang, Reza Zamani, Xiang Li, Christopher Bain, Yuqi Zhou, Walker Peterson, Huidong Wang, Chenglang Hu, Jongchan Park, Xiao Cheng, Benjamin Swedlund, Sandra Murillo, Anjali Sivanandan, Shiyu Sun, Liang Lanfeng, Mohammad Tariqul Islam, Baju C. Joy, Ishaq N. Khan, Sreedhar S. Kumar, Gabriel Mercado-V\'asquez, James V. Vizzard, Jonathan M. Matthews, Helen Huang, Xiaolu Guo, Ethan Nicklow, Guorui Chen, Ryan A. Neff, Surjendu Maity, Hyeonjin Park, Han-ho Joo, Katherine Dong, Yuyan Cai, Weihang Huang, Yichen Zou, Rui Yan, Raphael Figueroa, Artem Goncharov, Bella Rose Schremmer, Lian Elsa Linton, Keisuke Goda, Liang Gao, Ke Cheng, Leonardo Morsut, Jennifer L. Wilson, Jianping Fu, Lim Chwee Teck, Deblina Sarkar, Andreas Hierlemann, Sava\c{s} Tay, Alexander Hoffmann, Donald Richieri Griffin, Jun Chen, Shana O. Kelley, Shyni Varghese, Jinwoo Cheon, Wilbur A. Lam, James J. Moon, Wilson W. Wong, Samir Mitragotri, Dino Di Carlo
HARDEN is a constrained evolutionary search method that transforms existing evaluation cases into more challenging variants while preserving their expected outputs. It operates along domain‑specific complexity axes and enforces feasibility constraints such as task semantics, realism, and execution validity. Experiments on FinQA, PubMedQA, and ContractNLI with Qwen3.5 models show that HARDEN can reduce task‑model accuracy by an average of 22.7% and up to 49.9% compared to single‑pass baselines.
By Aditya Kumaran, Rahul Singhal, Karime Maamari, Amine Mhedhbi, Pradyumna Tambwekar
T‑RoPE introduces a time‑aware Rotary Position Embedding for sequential recommendation, replacing index‑only rotations with timestamp‑based angles, learnable temporal coefficients, multiscale frequency banks, shifted query alignment, and non‑stationary key rotation. The authors prove that standard RoPE is time‑translation invariant and cannot capture seasonal contexts, while T‑RoPE breaks this invariance while preserving the RoPE interface. Across five public benchmarks and an industrial‑scale e‑commerce dataset, T‑RoPE outperforms baselines by up to 130 % in HR@10 and delivers significant online lift in conversion and order count.
By Yang Liu, Noel Loo, Ali Khanafer, Shuying Sun, Akshay Soni, Zhong Wu, Linjun Yang
LAVOIR is a single‑pass decision encoder that not only predicts answers to typed questions but also identifies which missing pieces of information (slots) would most improve its confidence. By placing candidate slots next to answer options, one forward pass yields both the decision distribution and the expected value of asking each slot, without requiring human labels. In controlled experiments, LAVOIR’s question policy matches a greedy oracle and improves accuracy by up to 14.1 points over never asking, while on real conversations it raises accuracy by 8.3 points with minimal questioning.
By Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozay
CRC‑Router is a risk‑constrained, uncertainty‑aware routing module designed for medical AI systems, particularly in imaging. It fuses multiple uncertainty signals with predictive scores to estimate a per‑finding wrong‑accept risk, then uses Conformal Risk Control to set acceptance thresholds that meet a user‑specified risk target. Applied to chest X‑ray triage on the NIH ChestX‑ray14 dataset, CRC‑Router outperforms baseline methods in risk–coverage trade‑off, both as a standalone layer and when integrated with the MedRAX agent, demonstrating its effectiveness and model‑agnostic compatibility.
By Xueyang Li, Mingze Jiang, Gelei Xu, Jun Xia, Ching-Hao Chiu, Mengzhao Jia, Danny Z. Chen, Yiyu Shi
ORCA is a new benchmark for evaluating large language models on Data Science Code Translation (DSCT), comprising two settings: ORCA-MAIN with 1,600 grounding-level tasks across data querying, manipulation, and deep learning, and ORCA-PROJECT with 200 full-project translation tasks across seven data‑science task types. Each task includes reference translations and test cases to verify functional equivalence, and a multi‑stage quality verification process ensures task correctness. Experiments show that even state‑of‑the‑art LLMs perform poorly on DSCT, with Claude‑Opus‑4.6 achieving only 56.92% success on ORCA‑MAIN and 33.67% on ORCA‑PROJECT, while an intent‑augmented approach improves success rates by 4.80% and 5.33% respectively.
By Xiaolong Li, Jinyang Li, Bowen Qin, Ge Qu, Nan Huo, Xiaohan Xu, Shipei Lin, Reynold Cheng
Backbone-Adaptive Evidence Routing (BAER) is a method that dynamically selects an evidence gathering protocol—evidence stacking, reliability-based expert routing, or candidate-blind reference verification—based on the judge backbone and benchmark during development, then locks that choice for testing. BAER maintains candidate symmetry, ensuring that swapping responses reverses preference but not its strength. In experiments across four benchmarks and two 8B judge backbones, BAER outperforms all compared methods, achieving the highest test accuracy in all eight conditions and improving over the strongest baseline by 0.87–7.32 points.
By Zeyan Li, Jing Peng, Jianfeng Xu
The paper investigates whether the reasoning process in large language models mitigates or exacerbates bias. Using a within-model ablation on three high-stakes datasets (Adult, COMPAS, Credit) across three 32‑B models, the authors find that reasoning resolves some counterfactual fairness flips but creates roughly five times as many new flips at high confidence. They introduce two dynamic tools—Counterfactual Depth Probability Gap and Bias Transition Matrix—to trace how bias propagates and amplifies during reasoning depth and to explain the asymmetric dual effect.
By Deng Pan, Joe Germino, Yihong Ma, Elizabeth Daly, Nuno Moniz, Ting Hua, Nitesh Chawla
The paper reviews the fragmented literature on real‑time voice agents, noting that architecture, turn‑taking, and agentic evaluation communities rarely cite each other. It presents three evidence‑based claims: (1) architecture choice is a deployment constraint rather than a definitive solution, (2) evaluation has shifted from component quality to grounded outcomes, and (3) the dyadic assumption is breaking down as multiparty turn‑taking and reasoning become essential. The authors propose the TRG reporting standard to characterize agents by timing, recovery, and state‑verified outcomes, with an optional fourth axis for multiparty contexts.
By Shivam Negi, Arpit Rawat, Rashi Jain
The paper introduces the Correlated Promotion Benchmark (CPB) to evaluate how agents decide whether to admit claims into shared memory, addressing the risk of repeating false claims. CPB offers two modes: CPB-Static, a frozen test set with fixed gold actions, and CPB-Live, which runs multi‑agent teams and tracks source lineage. Experiments across eight admission policies and four agent families show that deduplication reduces false claims but also discards true ones, while gating on declared source type most effectively limits false adoption.
By Xiaoyang Li, Yiqi Wang, Chencheng Zhu, KE XU, Wencheng Yang, Zequn Sun, Pingan Song, Yiqun Duan, Taotao Cai
The paper introduces PTC-Decoder, a training‑free, plug‑and‑play decoder designed to improve the reliability of small language models (SLMs) on offline, resource‑constrained edge devices such as remote sensing satellites. PTC-Decoder combines a Plan‑to‑Act paradigm that forces the model to invoke a planning tool at the first inference step with a deterministic finite automaton (TC‑Decoder) that imposes token‑level hard constraints on tool names while preserving the model’s reasoning capability. Evaluated on 200 real satellite tasks across seven SLMs, the approach achieves a statistically significant mean overall score gain of +1.21, with the TC‑Decoder component identified as the primary contributor to performance gains.
By Minghui Yu, Ke Mu, Gang Wu
MoMHa is a system that optimizes large language model harnesses across three objectives—accuracy, behavioural safety, and token cost—using a single‑phase joint‑reward proposer. It outperforms alternative strategies on seventeen domains, including synthetic suites and real‑world benchmarks, achieving higher joint scores and better safety while reducing token usage. The approach demonstrates that multi‑objective harness design can transfer effectively to unseen models and tasks.
By Subhojyoti Mukherjee, Md Mehrab Tanjim
arXiv:2609. 30971v1 Announce Type: new Abstract: Embodied agents offer a promising route to automating scientific experimentation, yet their progress is constrained by the lack of reliable and systematic evaluation environments.
By Maokai Qin, Chuan Qin, Qi Zhang, Dianyu Liu, Zirui Liu, Hongting Niu, Yuanchun Zhou, Hengshu Zhu
The paper introduces Governed Deduction (GD), a framework that distinguishes between premise relevance and premise authorization in reasoning systems. Using an RBAC-augmented Spider benchmark, the authors create 4,461 authorization pairs where the same premise is either allowed or denied for a transition, revealing that simple linear models fail to capture policy constraints. The study demonstrates that matched controls and leakage audits are essential for evaluating policy-sensitive reasoning.
By Wesley Shu, Hsi-Ching Lin
The paper introduces LIFT, a lightweight vector‑intervention technique that transfers reasoning capability from a base large language model (LLM) to a vision‑language model (VLM) without retraining the VLM backbone. LIFT defines Reasoning Vectors as differences in hidden states between a reasoning path with an explicit trace and a solver path without it, and injects these vectors into the VLM’s language‑side activations. Experiments on two VLMs across six reasoning benchmarks show that vectors derived from the base LLM consistently outperform those derived from the aligned VLM, indicating that the base LLM is a more effective source for recovering degraded reasoning.
"whyItMatters":"The study demonstrates that a simple, frozen‑backbone intervention can partially restore reasoning abilities in multimodal models, highlighting the value of leveraging the original language model’s reasoning power."
By Ziyi Wang, Li Li, Aolin Zhou, Yankun Shen, Chonghan Liu, Shuxia Lin, Xu Yang
The paper introduces Stackelberg Program Optimization (SPO), a framework that uses large language models to discover adaptive destroy‑repair operators for large neighborhood search. SPO conditions operator decisions on a compact state representation, enabling state‑dependent behavior, and frames the discovery process as a Stackelberg game where destroy operators act as leaders and repair operators as conditional followers. Experiments on the traveling salesperson and capacitated vehicle routing problems show that SPO outperforms strong baselines, generalizes to larger instances, and exhibits coupled improvement in operator behavior during discovery.
By Xinyi Ke, Kai Li, Junliang Xing, Yifan Zhang, Jian Cheng