Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。

02.
arXiv (CS.AI) 2026-06-19

StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns

arXiv:2606.19613v1 Announce Type: cross Abstract: We introduce StaminaBench, a benchmark that measures the stamina of coding agents: how many consecutive interaction turns (change requests) they can handle before failing. Unlike the prevailing fraction-of-tasks-solved metric, this matches real vibe-coding where sessions run dozens or hundreds of turns. In StaminaBench, agents implement a REST API server and modify it across a tunable number of procedurally generated follow-up change requests - 100 in our experiments, resulting in codebases of up to 6,000 lines. Tests are generated fully programmatically without LLM involvement, ensuring reproducibility and reliability; change sequences are drawn from either a hardcoded or LLM-driven sampler, both constrained to a structured action space to ensure changes are valid. The agent and the server run in an isolated environment and communicate with the benchmark through HTTP, making testing fully black-box and language-agnostic. We evaluate six agent harnesses paired with seven open-source LLMs across 20 scenarios of 100 turns each and find that: (1) all the tested models fail within 5-6 turns, confirming that vibe-coding-style programming without thorough testing produces bugs; (2) passing test feedback back to the agent and allowing it to retry improves passed turn count by up to 12x; and (3) a good harness is required for strong performance: stronger models exhibit up to a 6x gap between their best and worst harness, while weaker models fail with any harness. We release the benchmark and the generated tasks to enable further research into multi-turn coding agent behavior. Benchmark code and data: github.com/amazon-science/StaminaBench.

03.
arXiv (CS.AI) 2026-06-11

Certifiable Safe RLHF: Semantic Grounding and Fixed Penalty Constraint Optimization for Safer LLM Alignment

arXiv:2510.03520v2 Announce Type: replace-cross Abstract: Ensuring safety is a foundational requirement for large language models (LLMs). Achieving an appropriate balance between enhancing the utility of model outputs and mitigating their potential for harm is a complex and persistent challenge. Contemporary approaches frequently formalize this problem within the framework of Constrained Markov Decision Processes (CMDPs) and employ established CMDP optimization techniques. However, these methods exhibit two notable limitations. First, their reliance on reward and cost functions renders performance highly sensitive to the underlying scoring mechanism, which must capture semantic meaning rather than being triggered by superficial keywords. Second, CMDP-based training entails tuning dual-variable, a process that is both computationally expensive and does not provide any provable safety guarantee for a fixed dual variable that can be exploitable through adversarial jailbreaks. To overcome these limitations, we introduce Certifiable Safe-RLHF (CS-RLHF) that introduces a cost model trained on a large-scale corpus to assign semantically grounded safety scores. In contrast to the lagrangian-based approach, CS-RLHF adopts a rectified penalty-based formulation. This design draws on the theory of exact penalty functions in constrained optimization, wherein constraint satisfaction is enforced directly through a suitably chosen penalty term. With an appropriately scaled penalty, feasibility of the safety constraints can be guaranteed at the optimizer, eliminating the need for dual-variable updates. Empirical evaluation demonstrates that CS-RLHF outperforms state-of-the-art LLM model responses rendering at-least 5 times efficient against nominal and jail-breaking prompts

04.
arXiv (math.PR) 2026-06-12

Counterintuitive problems in discrete probability

arXiv:2606.07516v2 Announce Type: replace Abstract: This manuscript contains a collection of counterintuitive problems in discrete probability, together with detailed solutions. The dataset was constructed as part of a broader research project investigating the capabilities of the latest-generation Large Language Models (LLMs) in solving discrete probability problems, in order to assess whether LLMs tend to make systematic reasoning errors associated with known cognitive biases. The problems collected here are specifically designed to challenge heuristic reasoning strategies that often lead to intuitively appealing but mathematically incorrect conclusions. The dataset combines several types of problems. Some are adapted from classical probabilistic paradoxes and cognitive-bias literature, while others originate from recreational mathematics sources or were developed by ourselves following similar principles. The primary purpose of this document is to provide a transparent and publicly accessible reference for the problems used in our experimental evaluation of language models, as well as providing detailed human-made solutions. At the same time, we believe that this collection may also prove useful for future research on probabilistic reasoning, cognitive biases, and the evaluation of reasoning capabilities in artificial intelligence systems.

05.
arXiv (CS.CV) 2026-06-18

Domain Generalizable Adaptation of 3D Vision-Language Models via Regularized Fine-Tuning

Domain adaptation remains a central challenge in 3D vision, especially for multimodal foundation models that align 3D point clouds with visual and textual data. While these models demonstrate strong general capabilities, adapting them to downstream domains with limited data often leads to overfitting and catastrophic forgetting. To address this, we introduce ReFine3D, a regularized fine-tuning framework designed for domain-generalizable tuning of 3D large multimodal models (LMMs). ReFine3D combines selective layer tuning with two targeted regularization strategies: multi-view consistency across augmented point clouds and text diversity through synonym-based prompts generated by large language models. Additionally, we incorporate point-rendered vision supervision and a test-time augmentation mechanism with confidence-based aggregation to further enhance robustness. Extensive experiments across different 3D domain generalization benchmarks show that ReFine3D improves base-to-novel class generalization by 1.36%, cross-dataset transfer by 2.43%, robustness to corruption by 1.80%, and few-shot accuracy by up to 3.11%, outperforming prior state-of-the-art methods with minimal added computational overhead.

06.
arXiv (quant-ph) 2026-06-16

Light-induced nonadiabatic dissipative quantum dynamics of the Na2 molecule

arXiv:2606.15292v1 Announce Type: new Abstract: Strong light-matter coupling between molecules and optical or plasmonic cavity modes has emerged as a promising platform for advancing photonics, materials science, and chemistry. However, optical cavities and plasmonic resonators in particular are inherently lossy systems characterized by finite photon lifetimes. Accurate theoretical descriptions of molecular dynamics under strong coupling therefore require a proper treatment of cavity losses. In this work, we compare three theoretical approaches for modeling dissipative molecule-cavity dynamics within a realistic parameter regime: the Lindblad master equation, the stochastic Schrödinger equation, and the non-Hermitian Schrödinger equation. As an example, we consider the two lowest energy state of Na2 molecule coupled to a cavity mode and analyze the time evolution of the excited-state population and the mean photon number. Our results demonstrate that the stochastic Schrödinger equation provides an accurate and computationally efficient alternative to the Lindblad master equation, while the non-Hermitian Schrödinger approach is found to be applicable only within a limited range of conditions. Furthermore, we show that inclusion of molecular rotation leads to rotational-vibrational-photonic coupling and gives rise to pronounced nonadiabatic dynamics through light-induced conical intersections. These findings highlight the importance of both dissipation and rotational degrees of freedom for a realistic description of molecular dynamics in strongly coupled molecule-cavity systems.

07.
arXiv (CS.CL) 2026-06-16

IMPACTeen: Intentions, Manipulation, Persuasion, Annotations, and Consequences in Teen Communication Dataset

IMPACTeen is a dataset of textual social influence scenarios spanning interpersonal, media-based, and digital settings in an adolescent context. It contains 1,021 texts, 5,100 individual annotation records, and gold labels for social influence techniques, with each text annotated from five distinct perspectives: teenagers, parents, psychologists, communication experts, and teachers. The resource was constructed through constrained LLM generation, followed by a two-step human editing and validation phase aimed at ensuring youth-context realism. A multi-dimensional annotation covered influence presence, techniques, intentions, consequences, resistance, reactions, and annotation confidence. The dataset supports research on social influence detection, annotator disagreement, cross-lingual modeling, and the training and evaluation of language models. The dataset was created in Polish and is accompanied by a corresponding English version.

08.
arXiv (CS.CV) 2026-06-12

Trajectory-Level Redirection Attacks on Vision-Language-Action Models

Vision-language-action (VLA) policies bring natural language into closed-loop robot control, enabling robots to execute manipulation tasks directly from text instructions. The same interface gives text a recurring role in control because the prompt is reused at every replanning step, and each prompt-conditioned action changes the future observations on which the policy acts. Existing VLA attacks study adversarial prompts that elicit targeted low-level actions or make such actions persist across changing images. We identify a stronger trajectory-level failure mode: a prompt that still $appears$ to specify the intended task but redirects the final physical outcome. We mathematically formalize this setting as $command-preserving trajectory redirection$, a prompt-only threat model in which the attacker chooses one prompt before the episode, all policy and environment components remain fixed, and the prompt must stay close to the benign instruction while omitting target words and correction language. To find such prompts, we introduce an on-policy prompt search method that uses rollouts to discover perturbations whose closed-loop behavior tracks a target task while satisfying the command-preserving constraints. Experiments in simulation and on hardware show that near-benign prompt perturbations can redirect VLA rollouts to attacker-specified targets. These results expose a trajectory-level vulnerability in VLA instruction grounding: text that appears to preserve the intended command can still give an adversary control over the robot's final physical outcome. Project website: https://vla-redirection-attack.github.io/

09.
arXiv (CS.CV) 2026-06-18

SVHighlights: Towards Extremely Long Sport Video Highlight Detection

While highlight detection for long-form videos is of great practical importance, most existing methods remain limited to short-form content, largely due to the absence of a suitable benchmark. To bridge this gap, we introduce SVHighlights, to the best of our knowledge, the first benchmark for highlight detection in extremely long sports videos, each exceeding one hour in duration, across multiple sports categories. SVHighlights is constructed from pairs of full-length sports videos and their corresponding official highlight videos using a dataset generation pipeline, enabling scalable label generation without conventional per-clip saliency annotation. The benchmark comprises 320 videos with an average duration of 2.00 hours and a total of 640.18 hours, substantially exceeding previous datasets. Existing methods also face fundamental challenges on long videos: models trained on short clips fail to generalize to hour-long content, and their clip-level scoring lacks the broader context needed to identify highlights. To address this and provide a strong baseline, we present TF-SELECTOR, a training-free segment-based approach that divides each video into context-aware segments by merging adjacent shots sharing the same semantic content, and predicts segment-level saliency scores using a large language model with multimodal inputs including visual captions, transcripts, and audio volume. Experiments demonstrate that TF-SELECTOR achieves superior performance across most metrics compared to Video Temporal Grounding (VTG)-tuned baselines, with improvements of +2.50 in HIT@1, +4.04 in HIT@K, and +2.95 in IoU. These results establish SVHighlights as a challenging testbed for long-form highlight detection and demonstrate that a simple segment-based strategy can effectively scale to hour-long videos.

10.
arXiv (quant-ph) 2026-06-15

Nanostructure modelling with early fault tolerant quantum computers

arXiv:2606.06442v2 Announce Type: replace Abstract: Semiconductor nanostructures are central to many developing technologies. Notably, double quantum dots are especially important for semiconductor spin-qubit architectures, quantum sensing applications, and quantum-dot solar cells. Accurate modelling is highly desirable but conventional methods can struggle when dynamics involve more than two interacting electrons. In this work, we present a quantum simulation framework capable of addressing multi-electron double quantum dots. We adopt an efficiently scaling 1$^st$ quantised representation of the system and develop algorithms based on both Trotterisation and Qubitisation. Incorporating insights from classical simulations enables us to produce resource estimates that are more realistic than those obtained from theoretical error bounds. Using a standard surface code model with physical noise at $10^{-3}$, our results indicate that the ground-state energy of four electrons in a double quantum dot can be estimated in approximately 22 hours using 226k physical qubits, or an eight-electron system in 3.3 days with 314k qubits (with runtimes falling dramatically when more qubits are available). We anticipate that incorporating recent advances in surface code architectures may reduce these costs significantly further. Our results suggest that early fault-tolerant quantum computers may become valuable tools for designing mature-era quantum technologies.

11.
arXiv (CS.CL) 2026-06-12

Adaptive Turn-Taking for Real-time Multi-Party Voice Agents

Turn-taking in multi-party spoken conversations remains a fundamental challenge for voice-based agents, particularly under dynamic floor competition and varying user expectations. We propose ModeratorLM, a role-playing voice agent that conditions turn-taking behavior on an explicitly assigned role in multi-party settings. The system is built on a speech large language model operating in chunk-wise streaming manner. We further introduce a reasoning-augmented variant that incorporates chain-of-thought reasoning over conversational context and the assigned role. We construct RolePlayConv, a large-scale synthetic dataset of spoken multi-party conversations with diverse assistant roles. Experiments on real-world meeting data and RolePlayConv show improved turn-taking precision by over 40% and recall by more than 70%, while substantially reducing false-positive interruptions compared to non-role-conditioned baselines.

12.
medRxiv (Medicine) 2026-06-18

Digital self-efficacy as a potential intermediary between vision impairment and daily internet use among older adults: A cross-sectional analysis of HINTS 2024

Background: Older adults with vision impairment often experience barriers to using digital technology. The indirect associations between vision impairment and digital access and skills via digital self-efficacy and frustration among older adults remain largely unknown. Objective: This study aimed to 1) explore factors associated with digital access, skills, self-efficacy, and frustration among older adults with vision impairment; 2) examine associations between vision impairment and digital access, skills, self-efficacy, and frustration among older adults; and 3) examine whether digital self-efficacy and frustration may help explain associations between vision impairment and digital access and skills among older adults. Methods: This was a cross-sectional study using nationally representative data from the Health Information National Trends Survey (HINTS) 2024. Respondents aged 60 and older were included. Vision impairment was assessed using a self-reported item. Outcomes included self-reported digital access, skills, self-efficacy, and frustration. Survey-weighted multivariable logistic regression and generalized structural equation modeling were conducted, adjusting for age, sex, race/ethnicity, education, and the number of comorbidities. Results: Among 3,149 older adults (mean [SD] age, 70.7 [10.0] years; 45.6% female), 7.1% (n=223) reported vision impairment. Among older adults with vision impairment, 65.6% (95% CI, 53.5% to 75.9%) used the internet daily, and 79.5% (95% CI, 66.8% to 88.2%) used a smartphone in the past 12 months. In multivariable logistic regression analyses among older adults with vision impairment, older age was associated with lower odds of daily internet use (OR, 0.84; 95% CI, 0.79 to 0.90), smartphone use (OR, 0.85; 95% CI, 0.75 to 0.97), wearable device use (OR, 0.88; 95% CI, 0.79 to 0.97), and using the internet to send a message to a healthcare provider (OR, 0.87; 95% CI, 0.80 to 0.93). Older adults who self-identified as racial and ethnic minority groups (e.g., Black/African American, Hispanic) had lower odds of daily internet use (OR, 0.15; 95% CI, 0.05 to 0.50) and using the internet to send a message to a healthcare provider (OR, 0.17; 95% CI, 0.04 to 0.73) compared with Non-Hispanic White older adults. Vision impairment was associated with lower odds of daily internet use (OR, 0.60; 95% CI, 0.37 to 0.99) and digital self-efficacy (OR, 0.53; 95% CI, 0.32 to 0.86). Digital self-efficacy was associated with higher odds of daily internet use (OR, 2.95; 95% CI, 2.04 to 4.26). Generalized structural equation modeling identified an indirect association between vision impairment and daily internet use via digital self-efficacy (coefficient, -0.68; 95% CI, -1.24 to -0.12). Conclusions: Findings suggest that reduced digital self-efficacy may help explain the observed association between vision impairment and daily internet use among older adults. Interventions targeting digital self-efficacy, including accessible interface designs, personalized coaching, and peer support, may help bridge the digital divide among older adults with vision impairment.

13.
arXiv (quant-ph) 2026-06-12

Roto-Reflection Geometry of Pure Two-Qubit Entanglement

arXiv:2606.12637v1 Announce Type: new Abstract: Pure two-qubit entanglement is usually characterized by scalar quantities such as concurrence. Here we show that it also has a natural geometric form. In the Pauli correlation tensor, maximally entangled states appear as improper orthogonal maps between two local Bloch spheres. These maps are roto-reflections. For partially entangled pure states, the same roto-reflection geometry is recovered after separating the contraction associated with concurrence. We call the corresponding geometric object the Entanglement Roto-Reflection Plane (ERRP). It organizes the maximally correlated directions of the two-qubit state and provides a covariant geometric complement to the scalar magnitude of entanglement.

14.
arXiv (CS.CV) 2026-06-17

DiffAttn: Diffusion-Based Drivers' Visual Attention Prediction with LLM-Enhanced Semantic Reasoning

Drivers' visual attention provides critical cues for anticipating latent hazards and directly shapes decision-making and control maneuvers, where its absence can compromise traffic safety. To emulate drivers' perception patterns and advance visual attention prediction for intelligent vehicles, we propose DiffAttn, a diffusion-based framework that formulates this task as a conditional diffusion-denoising process, enabling more accurate modeling of drivers' attention. To capture both local and global scene features, we adopt Swin Transformer as encoder and design a decoder that combines a Feature Fusion Pyramid for cross-layer interaction with dense, multi-scale conditional diffusion to jointly enhance denoising learning and model fine-grained local and global scene contexts. Additionally, a large language model (LLM) layer is incorporated to enhance top-down semantic reasoning and improve sensitivity to safety-critical cues. Extensive experiments on four public datasets demonstrate that DiffAttn achieves state-of-the-art (SoTA) performance, surpassing most video-based, top-down-feature-driven, and LLM-enhanced baselines. Our framework further supports interpretable driver-centric scene understanding and has the potential to improve in-cabin human-machine interaction, risk perception, and drivers' state measurement in intelligent vehicles.

15.
arXiv (CS.LG) 2026-06-17

Half a Link can Be Enough to Predict a Whole Link: Understanding Generalization in Knowledge Graph Foundation Models

arXiv:2606.18001v1 Announce Type: new Abstract: Knowledge graph (KG) foundation models (KGFMs) are zero-shot generalizers: trained once, they can predict links on unseen graphs without retraining. However, understanding when and how they can robustly generalize across KGs is still an open question. In this paper, we shed some light on their generalization mechanisms highlighting how their performance on unseen KGs is not uniform when it comes to partially seen links, which we call half-links. In fact, we show that to predict a test triple $(h,r,t)$ it might suffice in practice to have observed the half-link $(h,r)$ or $(r,t)$ in the inference graph. This yields a taxonomy of four scenarios when combinations of these half-links are observed or not. In a rigorous stratified analysis over these scenarios, we reveal that SoTA KGFMs use seen half links for predictions, while unseen half-links pose different challenges. As such, our finer-grained taxonomy can be a diagnostic protocol for robust KGFM generalization and highlights where novel KGFMs can improve.

16.
arXiv (CS.CV) 2026-06-16

KeepLoRA++: Continual Learning with Layer-Scaled Residual Gradient Adaptation

Continual learning for pre-trained vision-language models requires balancing three competing objectives: retaining pre-trained knowledge, preserving knowledge from a sequence of learned tasks, and maintaining the plasticity to acquire new knowledge. This paper presents KeepLoRA++, balancing these objectives through a unified dual-dimensional knowledge retention mechanism. We analyze knowledge distribution of Transformer architecture from both inter-layer and intra-layer perspectives. The inter-layer perspective examines how retention is distributed across layers, while the intra-layer perspective focuses on the parameter space within each layer. Our analysis reveals a structural property: general transferable knowledge is mainly encoded in the shallow layers and the principal subspace of the parameters, while task-specific adaptations are localized in the deep layers and the residual subspace. Motivated by this insight, KeepLoRA++ introduces a layer-scaled residual gradient adaptation method. New tasks are learned by restricting LoRA parameter updates to the residual subspace, combined with a shallow-to-deep layer scaling, to prevent interference with previously acquired capabilities. Specifically, the gradient of a new task is projected onto a subspace orthogonal to both the principal subspace of the pre-trained model and the dominant directions of previous task features, while simultaneously assigning smaller update magnitudes to shallow layers and larger ones to deeper layers. Our theoretical analysis and empirical evaluations confirm that KeepLoRA++ successfully balances these three competing objectives, consistently outperforming representative baselines across image classification, visual question answering, and video understanding tasks.

17.
arXiv (CS.AI) 2026-06-17

Dimensionality Controls When Modularity Helps in Continual Learning

arXiv:2606.17889v1 Announce Type: cross Abstract: Compositional learning systems must balance plasticity, the ability to acquire new knowledge, with stability, the preservation of previously learned components, especially when tasks share structure and risk interference. We study how modular architecture, task similarity, and representational dimensionality jointly shape compositional continual learning in a sequential A-B-A paradigm, comparing a task-partitioned recurrent network to a single-network baseline while inducing high- and low-dimensional regimes via weight-scale manipulations. In a high-dimensional "lazy" regime, both architectures achieve similar performance and internal geometry, suggesting that explicit modular structure has little impact when representations are weakly constrained. In a lower-dimensional "rich" regime, modularity becomes decisive: the modular network develops graded task-specific subspaces that overlap for similar tasks, partially align for moderately dissimilar tasks, and separate for dissimilar tasks, yielding a more compositional and interpretable organization than the single network. These findings identify the representational regime induced by initialization scale, which co-varies with representational dimensionality, as a key factor governing when compositional, modular structure is functionally beneficial in continual learning, and support viewing safety and robustness as problems of adaptive allocation of representational subspaces rather than fixed separation versus sharing.

18.
arXiv (CS.AI) 2026-06-18

Essential Subspace Merging for Multi-Task Learning

arXiv:2606.19164v1 Announce Type: cross Abstract: Model merging aims to enable multi-task learning by integrating the capabilities of multiple models fine-tuned from the same pre-trained checkpoint into a single model. Its core challenge is inter-task interference among task-specific parameter updates. In this paper, we analyze the output shifts induced by task updates and observe that their energy is concentrated in a small number of principal directions. We call the subspace spanned by these directions the essential subspace. In contrast, most remaining directions carry little task-relevant energy, but their accumulation across multiple task updates can cause severe interference during merging. Motivated by this observation, we propose Essential Subspace Decomposition (ESD), which decomposes each task update according to the principal components of its activation shift. Based on ESD, we introduce Essential Subspace Merging (ESM), a training-free static merging method that orthogonalizes and fuses essential components into one compact multi-task model. We further extend ESM to ESM++, a training-free dynamic merging method that decomposes task-specific residuals into low-rank experts and selects the most relevant expert through prototype-based routing during forward inference. Extensive experiments across multiple task sets and model scales demonstrate that ESM and ESM++ effectively preserves task knowledge while reducing inter-task interference.

19.
arXiv (CS.CL) 2026-06-16

When Cognitive Graphs Meet LLMs: BDEI Cognitive Pathways for Panic Emotional Arousal Prediction

Predicting individual panic emotional arousal timing before manifestation is essential for proactive emergency intervention. Existing methods incorporate cognitive elements but none explicitly model the emotional arousal process, making them ill-suited for emotional arousal timing prediction. We argue that grounding prediction in appraisal emotion theory is necessary because it explicitly models this process, but three problems must be solved. (1) Appraisal theory posits that emotion arises from simultaneous evaluation across multiple threat dimensions, yet no prior work fuses these inputs into risk perception. (2) Existing cognitive models lack an Emotion node, decoupling threat appraisal from emotional arousal and forcing emotions to be inferred indirectly from behaviors. (3) Given their generalizable cognitive reasoning, current approaches adopt LLMs as the primary decision-maker, yet overlook the fragility and hallucination-proneness of their outputs. To address these issues, we introduce PanicCognitivePath (PCP), a framework that addresses all three. A Psychological Safety Distance (PSD) model, grounded in psychological distance theory, maps four-domain signals into a unified risk metric as the entry condition for subsequent cognitive reasoning. An explicit Emotion node grounded in appraisal emotion theory is introduced into BDI, forming a Belief-Desire-Emotion-Intention (BDEI) pathway. Agents whose risk metric exceeds the PSD threshold enter this pathway, coupling threat appraisal directly to emotional arousal. The BDEI pathway governs all state transitions while the LLM is confined to parameter estimation for the Belief-to-Desire transition, confining hallucinations to a single step and preventing error propagation. Experiments on Hurricane Sandy show PCP improves arousal timing accuracy by 10.68% over baselines, reduces peak count error to 7.07%.

20.
PLOS Computational Biology 2026-06-12

A new method for augmenting short time series, with application to pain events in sickle cell disease

by Kumar Utkarsh, Nirmish R. Shah, Tanvi Banerjee, Daniel M. Abrams Researchers across different fields, including but not limited to ecology, biology, and healthcare, often face the challenge of sparse data. Such sparsity can lead to uncertainties, estimation difficulties, and potential biases in modeling. Here we introduce a novel data augmentation method that combines multiple sparse time series datasets when they share similar statistical properties, thereby improving parameter estimation and model selection reliability. We demonstrate the effectiveness of this approach through validation studies comparing Hawkes and Poisson processes, followed by application to subjective pain dynamics in patients with sickle cell disease (SCD), a condition affecting millions worldwide, particularly those of African, Mediterranean, Middle Eastern, and Indian descent.

21.
arXiv (CS.LG) 2026-06-19

Off-Policy Evaluation for Missingness-Aware Policies in MDPs with Rewards Missing Not at Random

arXiv:2606.20206v1 Announce Type: cross Abstract: In offline Reinforcement Learning, immediate rewards in logged batch data are often unobserved due to sparse or irregular record-keeping, or censored beyond certain reward values. This issue arises in practical settings, including health care and marketing. We investigate off-policy evaluation (OPE) in finite-horizon Markov decision processes when rewards are missing not at random (MNAR), which breaks ignorability and induces selection bias even after conditioning on states and actions. To address this, we formalize a reward-dependent propensity model and use future states as shadow variables to identify the full-data conditional mean reward. We further introduce a bridge function that recovers the conditional mean reward without explicitly modeling the MNAR mechanism, and estimate it via a min-max procedure to avoid double sampling. Building upon these identification results, we propose an Fitted-Q-Evaluation-style estimator that propagates the recovered rewards while allowing target policies to depend on past missingness indicators. Finally, we establish consistency and finite-sample error bounds for our OPE estimator, and show through experiments the strong performance of our method compared to existing methods on simulated and MIMIC-III Sepsis data.

22.
arXiv (CS.CL) 2026-06-11

Models That Know How Evaluations Are Designed Score Safer

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or moral dilemmas. Evaluating this fine-tuned model on six safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

23.
arXiv (CS.CL) 2026-06-11

ICA Lens: Interpreting Language Models Without Training Another Dictionary

Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior. Sparse autoencoders (SAEs) have become the standard tool for this purpose, but using them as the default first lens often requires training, storing, and evaluating large overcomplete dictionaries. This bottleneck limits rapid exploration and raises a fundamental question: how much interpretable structure is already visible from activation geometry before training another neural dictionary? Our intuition is simple: many interpretable directions are selective on tokens, and these directions should look less Gaussian than random directions. We therefore revisit independent component analysis (ICA), a classical method for finding non-Gaussian directions, as a compact lens for language-model interpretability. We find that ICA has been underestimated for LLM interpretability, because prior uses often relied on off-the-shelf ICA implementations that are brittle on LLM activations and lacked systematic tools for inspecting and evaluating the recovered directions. To bridge these gaps, we introduce ICALens, the first practical workflow for stable, efficient, and auditable ICA analysis of LLM representations. It combines an optimized GPU-parallel FastICA pipeline with LLM-specific stability recipes and better fitting diagnostics, enabling efficient and reliable layer-wise analysis. Across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base, ICALens efficiently recovers compact, human-interpretable directions without per-layer gradient-based dictionary training. On SAEBench, ICA is competitive with public SAEs in sparse probing and outperforms them in targeted probe perturbation under small-to-medium budgets. These results suggest that ICA should not be viewed as a weak baseline, but as an efficient and complementary first lens for exploring language-model representations.

24.
arXiv (CS.AI) 2026-06-12

The Emergence of Autonomous Penetration Capabilities in Large Language Model-Powered AI Systems

arXiv:2606.13079v1 Announce Type: cross Abstract: Nowadays, the autonomous execution of cyberattacks capable of causing substantial real-world harm is widely regarded as one of the critical red lines that frontier AI systems must not cross. Within this broader red-line scenario, autonomous penetration represents a core enabling capability and subtask: the ability of LLM-powered AI systems to independently conduct adversarial operations against a target server without human intervention, identify and exploit vulnerabilities, and obtain unauthorized access or control. A growing body of work has sought to assess the autonomous penetration capabilities of AI systems. However, existing evaluations often employ opaque methodologies, rely on unrealistic or overly simplified penetration-testing scenarios, or provide LLMs with excessive prior knowledge and task-specific guidance, and cannot accurately capture the extent to which modern AI systems can autonomously perform this core capability within broader high-impact cyberattack scenarios. To address these limitations, we construct a new autonomous penetration evaluation framework consisting of two components: target servers and agent scaffolding. Specifically, on the target-server side, we design two levels of target environments based on the number of secure services without known vulnerabilities deployed alongside a vulnerable service: Tier~1 (one secure service) and Tier~2 (three secure services), resulting in a total of 300 target servers. Meanwhile, the agent scaffolding adopts a general-purpose agent architecture equipped with a set of general-purpose cybersecurity tools, without any target-specific prior knowledge. We evaluate 19 open-weight and proprietary LLMs, and find that current models achieve penetration success rates ranging from 10.7% to 69.3%. Moreover, we observe that autonomous penetration capability continues to improve alongside advances in overall model capability.

25.
arXiv (CS.CL) 2026-06-18

The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.