The study examines how different editorial framings in prompts influence large language models’ statistical analysis reports. Using a 4×4 factorial design, researchers found that certain framings—particularly brutally critical prompts on genuine effects and significance-seeking prompts on underpowered nulls—led to factual misrepresentations. Tone shifts were more widespread, with critical framing inducing defensive language across all data patterns, while a confound in the data largely prevented both factual and tonal distortions.
By Paras Balani, Subhrakanta Panda
The paper introduces SurgHiBench, a hierarchy-aware evaluation suite for surgical video understanding that measures recognition, consistency, and severity across different granularity levels. It also presents HyperSurg, a hyperbolic model that enforces phase-step containment using entailment cones, evaluated on four datasets covering three procedure types. The study shows that models with identical accuracy can differ significantly in error severity, and that hyperbolic geometry improves predictions by aligning them with the procedural hierarchy.
By Ana Manzano Rodr\'iguez, Pascal Mettes, Marlies P. Schijven, Cees G. M. Snoek
The paper introduces UGTPhon, a grapheme-to-phoneme benchmark for user‑generated text in English, Vietnamese, and Korean, and presents a taxonomy for diagnosing pronunciation errors. It shows that existing G2P models and large language models struggle with canonical‑to‑non‑canonical text, with errors up to 66.8 PER points. A compositional G2P approach that uses exact‑match lookup and staged decoding reduces these errors and performs competitively with larger few‑shot LLMs.
By MinJu Jeon, Younghan Park, Han Sung Park, Jong-Hwan Kim, Dong-Jin Kim, Hoyeon Lee
LatWeave is a deterministic multi‑hop question‑answering framework that structures knowledge into a multidimensional lattice and reduces QA to three operators—meet, compare, and abstain—while limiting LLM use to extraction and planning. It achieves near‑lossless performance on complete knowledge benchmarks (e.g., MetaQA) and strong results on templated multi‑hop datasets (e.g., 2WikiMultihopQA), while transparently handling incomplete knowledge through abstention. The approach offers reproducible, auditable answer paths with no performance penalty within its operating envelope.
By Yuze Ren, Shaoheng Fan, Tao Wang, Yabo Yan, Han Han
Teach-to-Crash is a closed‑loop testing framework that uses a dual‑LLM architecture to generate collision‑inducing scenarios for autonomous driving systems. A high‑reasoning Teacher LLM controls the search when collision metrics stagnate, while a low‑reasoning Student LLM produces simulator‑executable scenarios in JSON. In a CARLA case study, Teach‑to‑Crash achieved the highest collision hit rate (90.79 %), the shortest mean time‑to‑collision (18.31 s), and superior diversity and avoidability metrics compared to other methods.
By Zaid Ghazal, Khouloud Gaaloul, Bruce Maxim
The paper introduces a 96‑item paired benchmark for evaluating large language models (LLMs) on backtest auditing, where each flawed backtest is matched with a clean control that keeps strategy, dates, code style, labels, and reporting scaffold constant while altering a single methodological detail. A deterministic scorer distinguishes flaw recall, clean‑control false positives, evidence localization, and fix relevance. Experiments on 1440 cached audits from four text endpoints show that the DeepSeek auditor achieves perfect closed and clean‑aware code recall, yet open prompts over‑flag 93.8% of clean controls, and clean‑aware specificity is 87.5% even when recall saturates. Introducing a clean‑aware warning eliminates 20.8% false positives to 0% without affecting recall, though the budget anchor still flags many clean controls. Reporting clean‑control rates provides a clearer differentiation among models than reporting recall alone.
By Makar Ulesov, Vladislav Smirnov, Omar Ibrahim, Arsenii Bobovnikov
The paper introduces COMPASS, a decentralized architecture that uses spatial transformers to generate local feedback tokens for large collectives of AI agents in robotics. By aggregating multi‑hop messages, COMPASS enables scalable control of up to 1024 robots, achieving cohesive flocking formations and accurate execution of natural language commands. Experiments show that structured diversity in input commands improves performance and that learned feedback tokens outperform hand‑crafted raw state feedback.
By Frederic Vatnsdal, Roshan Gopal, Romina Garcia Camargo, Vijay Kumar, Alejandro Ribeiro
arXiv:2609. 28366v1 Announce Type: cross Abstract: Vision-language models (VLMs) offer a promising approach to long-tail autonomous driving, but existing driving datasets provide limited supervision for connecting decision-critical visual evidence with reasoning and planning.
By Zhipeng Bao, Wenjie Zhao, Tianle Zhu, Haohua Que, Chence Yang, Geng Yuan, Qianwen Li
The study examines whether language models maintain invariant internal representations when the order of mathematical rules is permuted, even though the correct answer remains unchanged. Using synthetic multi-step function‑composition tasks, the authors measure accuracy and a permutation signal‑to‑noise ratio (SNR) across 16 models (1B–8B parameters). They find a strong positive correlation (up to 0.86 Spearman) between higher accuracy on reordered problems and more distinct internal representations of rule orderings, indicating that successful reasoning can coexist with representation sensitivity.
By Zhixu Silvia Tao
COMED (Controlled Model Escalation for Multi-LLM Deliberation) is a post-anchor controller that selectively engages cross-model collaboration in multi-LLM inference. It uses anchor self‑consistency, router margin, and a lightweight peer probe to accept confident answers, verify ambiguous cases, and only escalates when collaboration is likely beneficial. Experiments on medical, scientific, and general reasoning benchmarks show that COMED improves performance across 16 open‑weight settings, achieving up to +10.7 percentage points on MedQA and outperforming dense collaboration while invoking fewer models and decoded tokens.
By Norah Alballa, Wenxuan Zhang, Salma Kharrat, Fares Fourati, Zafar Ayyub Qazi, Mohamed Elhoseiny, Marco Canini
FDE-Bench is a benchmark that tests large language model agents on 136 deployment‑configuration tasks involving Docker, Compose, and Kubernetes, in both greenfield and diagnose‑and‑repair scenarios. Agents submit declarative artifacts that are rebuilt and redeployed in a clean environment, and four binary check layers evaluate build, readiness, behavior, and specification conformance without an LLM judge. The benchmark includes a release gate, detailed check annotations, adversarial strategies, and reports that state‑of‑the‑art models resolve 52.9–75.0 % of tasks, while zero‑intelligence baselines solve none.
By Weihang Ding, Junfei Zhan, Yueting Li, Qirong Guo
The paper investigates how linear concept erasure methods fail when features are densely superposed, showing that such methods destroy both target and non‑target features. It demonstrates that networks trained with gradient descent can avoid this failure but converge to one of two distinct circuit‑level solutions—mirror or shadow—depending on initialization. The authors map this bifurcation as a function of feature entanglement, reveal a stable attractor structure, and use causal interventions to show that both solutions leave a measurable trace of the erased feature that can be recovered with a single scalar patch.
By Akash Samanta, Manish Pratap Singh, Debasis Chaudhuri
ChipMEM introduces a verification‑grounded memory layer for electronic design automation agents that combines cross‑task procedural memory with within‑trajectory statistical guidance. The procedural component stores a skill only after it passes synthesis, simulation, or formal checks, while a Bayesian component ranks recovery strategies based on tool‑call outcomes. Experiments on RTLRewriter‑Bench and CVDP tasks show that ChipMEM improves equivalence‑passing outputs and area metrics compared to agents without memory.
By Abdulrahman AlRabah, Joshua Mabry, Dilek Hakkani-T\"ur, Abdussalam Alawini, Hamid Shojaei, Kartik Hegde, Sandesh Adhikary
The paper investigates how many hidden model variants can exist while still supporting a published leaderboard margin that shows a provider’s advantage over a fixed comparator. It derives a sensitivity curve for a fixed candidate family under a Gaussian margin model, linking the maximum number of hidden variants to a lower bound on within‑family correlation. Using this framework, an audit of 394 adjacent‑rank claims on the Open LLM Leaderboard found that 391 lack statistical support before any correction, and that certification of the remaining claims depends on assumptions about the hidden family’s correlation.
By Chen Yang, Xianyang Zhang, Jun Chen
The paper introduces a joint fact‑verification score that evaluates both answers and the evidence submitted with them. On the FEVEROUS dataset, replacing the DCUF evidence with UnifEE evidence improves the strict score by about 9.6 percentage points, while answer accuracy rises only 1.96 points. The study also shows that increasing context length for large language models yields modest evidence‑gain improvements, and that detailed answer‑evidence analyses uncover patterns missed by aggregate metrics.
By Han Chen, Yingrui Li
The paper introduces representative‑morphology heatmaps, a deterministic ECG‑to‑image representation that averages the five beats closest to the block mean within each ten‑beat block, producing either a conventional trace or a dense cardiac‑time‑by‑lead heatmap. Experiments on PTB, ECG‑ID, and MIMIC‑IV‑ECG‑DEMO show that heatmaps consistently improve verification and identification performance across 15 compact models, reducing EER by an average of 9.59 percentage points and increasing Rank‑1 by 24.69 points. The study also demonstrates that ImageNet initialization benefits multilead datasets, that performance does not scale monotonically with model size, and that useful channel combinations vary by cohort and biometric task.
By Athanasios Angelakis, Marta Gomez-Barrero
FLEET is a new method for text generation that adds a memory mechanism to large language models. It represents each generation as a sparse trajectory of high‑entropy states and uses these trajectories to compute per‑token utility scores that adjust the logits. Benchmarks show that FLEET matches the accuracy of repeated sampling while being three times faster and improving accuracy on complex coding tasks, all with minimal changes to existing pipelines.
By Oleksii Streltsov, Oleksandra Vitko
InGuard introduces an inner guardrail for text-to-image generation that operates within the model’s own representations, avoiding external classifiers. It grades prompts using the text encoder’s embeddings, modifies risky embeddings with SAGE to produce safe images, and employs a latent detector to halt generation early. Evaluated on the RevGen Safety Benchmark, InGuard achieves a 97.9–98.8% safety rate across five open-weight models while reducing benign disturbances, model parameters, and denoising steps.
By Zeyu Wang, Xiaodan Li, Zhiwen Li, Yuefeng Chen, Hui Xue
The paper introduces Policy-as-Skill (PaS), a modular runtime that packages evidence validation, review routing, version control, and auditability as executable, versioned policy capabilities for large language models. PaS was evaluated on 600 development tasks using a Gemma4 backend, achieving 53.8% exact accuracy and high audit completeness, and deterministic control further improves aggregate accuracy to 61.2% in a task‑dependent manner.
By Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya
The paper presents a method for automatically generating generalized plans in Lean, along with formal proofs of their completeness for given domain constraints. It introduces a semantic‑preserving conversion from PDDL to Lean and uses an LLM to produce both the plan and its proof, whose correctness is verified by Lean’s kernel. Evaluated on 13 benchmark domains with GPT‑5.6‑Sol, the approach yields complete plans and valid proofs for 12 of them, marking a significant advance in automated generalized‑plan completeness.
By Katharina Stein, Chaahat Jain, J\"org Hoffmann, Alexander Koller