Academic Intelligence · Curated Daily

Explore the Frontier of Global Academia

AcademicHub aggregates real-time literature from top journals and preprint platforms. Build your personal research radar and let large language models compile cross-disciplinary analysis briefings automatically.

01.
medRxiv (Medicine) 2026-06-18

Urinary Creatine Riboside Complements PSA to Improve Disease Detection in the Diagnostic Gray Zone of Prostate Cancer

Circulating prostate-specific antigen (PSA) discriminates poorly in the diagnostic gray zone (3.0-9.99 ng/mL), where ~75% of biopsies yield no clinically significant prostate cancer (PCa). We evaluated whether urinary creatine riboside (CR), a tumor-derived metabolite excreted through the prostatic urethra, complements PSA for gray-zone detection and independently predicts prostate-cancer-specific mortality (PCSM). In the NCI-Maryland PCa Case-Control Study (951 cases, 962 controls; 47.6% African American men; median follow-up 11.5 years), urinary CR was quantified by UPLC-MS/MS. Within the PSA gray zone (n = 668), urinary CR was complementary to PSA, with markedly higher single-marker discrimination than PSA (AUC 0.93, 95% CI 0.88-0.98 vs 0.77, 0.66-0.89) and additive when combined ({Delta}AUC +0.17, p < 0.001; 91.4% sensitivity at 80% specificity). After adjustment for 11 clinical and sociodemographic covariates, urinary CR independently predicted PCSM complementary to PSA (Fine-Gray SHR 1.72, 1.35-2.19 for CR; 1.35, 1.08-1.68 for PSA; Harrell's C 0.85 for CR + PSA vs 0.77 for PSA alone), with strongest signal in African American men (SHR 2.43, 1.57-3.75 for CR). We conclude that urinary CR is a candidate non-invasive biomarker complementary to PSA - improving gray-zone triage and predicting PCSM; prospective validation in biopsy-referred cohorts is warranted.

02.
arXiv (CS.AI) 2026-06-15

Large-scale semantic mapping of learner agency and autonomy reveals what measurement and generative AI research overlook

arXiv:2606.10881v2 Announce Type: replace Abstract: Learner agency and autonomy are foundational to personal development, yet a pervasive "jingle-jangle" fallacy (i.e. identical terms denoting different constructs, distinct terms denoting identical ones) has substantially hindered cumulative knowledge. Treating meaning as a phenomenon constituted through use in linguistic practice, we extracted 8,954 definitions and 2,700 scale items from over 14,000 publications, to investigate how researchers actually used learner agency and autonomy with a semantic analysis pipeline. The definitional landscape of two constructs resolves into three dimensions: regulation and control of learning (task), intrinsic motivation and internal decision-making (person), and social-relational action (sociocultural), thereby empirically quantifying the jingle-jangle fallacy. Existing scales, however, systematically underrepresent the sociocultural dimension. Critically, current generative AI research in education concentrates on learning regulation and control, narrowing the behavioral repertoire that AI-mediated learning environments are designed to cultivate. Beyond conceptual clarification, this work carries direct implications for conceptualization, measurement, and practice towards supporting the multidimensional learner agency and autonomy.

03.
arXiv (CS.CV) 2026-06-18

Do as I Do: Dexterous Manipulation Data from Everyday Human Videos

How can we scalably generate data for robotic manipulation, especially on human-like platforms such as dexterous multi-fingered hands? Learning from human videos has recently emerged as a likely answer to this question. However, difficulties in estimating hand-object interaction and crossing the human-to-robot embodiment gap have hindered the adoption of abundant monocular RGB-only human videos as the primary source of robot manipulation data. In this work, we present DO AS I DO, an algorithm to reconstruct and retarget monocular RGB human videos to multi-fingered dexterous robotic hands. DO AS I DO reconstructs hand-object interactions from various egocentric and exocentric in-the-wild video sources. The algorithm then retargets these hand-object interaction estimates into a sequence of actions executable in the real world, yielding robot-complete manipulation data from disparate human videos. Overall, DO AS I DO outperforms previous state of the art in estimating hand-object interactions and extracting dexterous manipulation trajectories from RGB videos, as we show in experiments on datasets with ground truths and on a dataset of video clips collected online. Our experiments enable us to propose an efficacy playbook for practitioners collecting human data for manipulation.

04.
Nature (Science) 2026-06-17

Mapping the neuronal building blocks of human language with language models

Authors:

Humans can convey new and highly diverse information through language. This ability to form and combine words into elaborate phrases and sentences enables us to express inexhaustible meanings and is fundamental to human cognition1–5. However, understanding the microscopic&nbsp;cellular building blocks and cortical landscape that precisely&nbsp;underlie human language has remained a challenge. Here we used wide-scale single-neuronal recordings combined with natural language processing models to identify fine-grained linguistic representations across the human frontotemporal cortex during language production. We find that, whereas certain neurons represented the detailed grammatical relationships between words or their parts of speech, others tracked the sentences’ higher-order syntactic structure, their phrase transitions and sequence. Collectively, these neurons reliably captured the words’ syntactic and semantic properties but also dynamically incorporated their specific sentence contexts, therefore&nbsp;enabling them to encode information combinatorially and at highly granular levels of detail. We show how these cell populations were locally organized and how their microscale representations differed from that of their wider field potential patterns. We also show how these neurons were distributed broadly across the frontotemporal cortex, but how their ability to encode linguistic information was left-lateralized and varied between&nbsp;cortical regions. Together, these findings identify some of the most basic cellular building blocks by which linguistic information is encoded in humans and begin to define the cortical landscape of language at a combined micro (cellular), meso (local population) and macro (regional) scale. Wide-scale recordings reveal neurons in&nbsp;the human brain that encode&nbsp;fundamental components of language such as&nbsp;the grammatical relationships between words, their parts of speech and the&nbsp;higher-order syntactic structure&nbsp;of phrases and sentences.

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

Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation

Large language models (LLMs) have become an effective tool for synthetic data generation, including for low-resource languages, where generated data can improve downstream task performance. Current best-performing approaches typically rely on few-shot prompting with target-language examples, which increases inference costs and may reduce diversity through lexical anchoring. In this work, we investigate activation steering as an alternative for low-resource synthetic data generation. We study two steering strategies: Language Steering, which targets the linguistic identity of a language, and Quality Steering, which captures well-formedness by contrasting human-written and backtranslated text representations. We evaluate these methods across four open-source LLMs, multiple layers, and 11 typologically diverse languages by generating sentiment and topic classification data and finetuning smaller classifiers. Steering is applied in both zero-shot and few-shot prompting settings and compared against non-steered counterparts. Our results show that steering on early layers consistently improves the diversity of generated data while often yielding stronger downstream model performance, particularly for low-resource languages.

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

Rewrite to Translate, Translate to Reward: Reinforcement Learning for Source Rewriting in Machine Translation

Rewriting source text with large language models (LLMs) before translation has been shown to improve machine translation (MT) quality. However, we find that prompt-based rewriting can degrade translation quality rather than improve it, particularly when smaller LLMs, such as 4B-parameter models, are used. We argue that this limitation stems from the difficulty of controlling rewriting behavior through natural-language prompts alone: a rewrite is useful only if it improves downstream translation, yet existing prompt-based methods do not explicitly optimize for this signal. To address this issue, we propose RLSR (Reinforcement Learning for Source Rewriting), a reinforcement learning framework that trains the rewriting model with a reward based on the downstream translation-quality improvement produced by each rewrite. Experiments across six MT systems and 16 language pairs show that our 4B RLSR-trained rewriting models significantly outperform both the no-rewriting baseline and prompt-based rewriting baselines at the same model scale, while remaining competitive with baselines that use a 235B LLM.

07.
arXiv (CS.AI) 2026-06-11

SkillJuror: Measuring How Agent Skill Organization Changes Runtime Behavior

arXiv:2606.11543v1 Announce Type: new Abstract: Agent Skills augment large language model (LLM) agents with procedural knowledge at inference time, but current benchmarks rarely distinguish what a Skill says from how it is organized. We study this distinction through Progressive Disclosure, where a concise root file points agents to supporting resources on demand, and compare it with a normalized flat baseline. We present SkillJuror, a framework for evaluating Skill writing paradigms through semantically controlled variants, matched multi-trial evaluations, and trajectory evidence while holding task knowledge fixed. In an 82-task SkillsBench study, Progressive Disclosure changes runtime behavior before aggregate outcomes: distinct Skill resources touched per trajectory rise from 1.18 to 3.85, and effective uptake events rise from 1.33 to 3.92. It also yields 17 additional verifier-passing trials out of 410 matched trials (+4.1%) over the normalized flat baseline. The benefit is task-dependent. Progressive Disclosure helps when supporting resources guide implementation, checking, or repair, but is weaker when success hinges on exact output conventions, numerical thresholds, or long artifact-generation pipelines. These results show that Skill organization is not mere presentation: it can change how agents search and apply procedural knowledge, while outcome gains depend on whether the exposed resources are actionable for the task. Code is available at https://github.com/zhiyuchen-ai/skill-juror.

08.
arXiv (CS.CL) 2026-06-17

Moderating Illicit Online Image Promotion for Unsafe User-Generated Content Games Using Large Vision-Language Models

Online user generated content games (UGCGs) are increasingly popular among children and adolescents for social interaction and more creative online entertainment. However, they pose a heightened risk of exposure to explicit content, raising growing concerns for the online safety of children and adolescents. Despite these concerns, few studies have addressed the issue of illicit image-based promotions of unsafe UGCGs on social media, which can inadvertently attract young users. This challenge arises from the difficulty of obtaining comprehensive training data for UGCG images and the unique nature of these images, which differ from traditional unsafe content. In this work, we take the first step towards studying the threat of illicit promotions of unsafe UGCGs. We collect a real-world dataset comprising 2,924 images that display diverse sexually explicit and violent content used to promote UGCGs by their game creators. Our in-depth studies reveal a new understanding of this problem and the urgent need for automatically flagging illicit UGCG promotions. We additionally create a cutting-edge system, UGCG-Guard, designed to aid social media platforms in effectively identifying images used for illicit UGCG promotions. This system leverages recently introduced large vision-language models (VLMs) and employs a novel conditional prompting strategy for zero-shot domain adaptation, along with chain-of-thought (CoT) reasoning for contextual identification. UGCG-Guard achieves outstanding results, with an accuracy rate of 94% in detecting these images used for the illicit promotion of such games in real-world scenarios.

09.
arXiv (math.PR) 2026-06-16

A non-asymptotic bound on the TV distance between a Wishart matrix and an appropriately scaled GOE matrix

arXiv:2606.16018v1 Announce Type: new Abstract: In this note, we prove a non-asymptotic version of a theorem by Bubeck, Ding, Eldan, and Rácz, showing that a Wishart matrix is close in total variation to an affine transformation of a GOE matrix. The proof mirrors the proof given by Bubeck et al., with some changes made to make it non-asymptotic.

10.
arXiv (CS.LG) 2026-06-17

ReRAM-aware Model Finetuning addressing I-V Non-linearity and Retention Errors

arXiv:2606.17471v1 Announce Type: new Abstract: Traditional CPU, GPU, and NPU architectures are increasingly limited by the von Neumann bottleneck. While In-Memory Computing (IMC) using ReRAM crossbar arrays offers a high-density, energy-efficient alternative, its practical deployment is constrained through their non-idealities. Existing hardware-aware training frameworks often require training from scratch, which is computationally prohibitive for modern large-scale models. In this work, we propose a finetuning-based hardware-aware training algorithm that enables robust DNN deployment on ReRAM with minimal training overhead. Our approach mitigates I-V non-linearity by applying a range-shrunk sinh transformation and incorporates retention errors directly into a regularization loss during the finetuning process. We evaluate our framework across models and tasks such as image classification and question-answering (QA). Experimental results demonstrate that our method achieves similar accuracy on large-scale models like ResNet18 and DeiT-Tiny as the base model. In-case of ImageNet for MobileNetV3 families the technique has only less than 2% accuracy degradation. Further, applying the technique on the SQuAD v2 dataset results in only 1 point degradation of F-1 score.

11.
arXiv (CS.LG) 2026-06-16

Privacy from Symmetry: Orthogonally Equivariant Transformers for LLM Inference

arXiv:2606.16461v1 Announce Type: new Abstract: Running large language models locally is often impractical, pushing inference on sensitive text to third-party providers. Split inference partially mitigates this by keeping tokens on the client and sending only hidden representations, but these representations can still be recovered via nearest-neighbor search against the public embedding table. We propose an orthogonal obfuscation procedure in which the client multiplies embeddings by a secret orthogonal matrix before transmission. To enable correct inference under arbitrary rotations, we introduce ConjFormer, a transformer variant that is exactly $\mathrm{O}(d)$-equivariant via a lightweight normalization change (scalar RMSNorm) together with blockwise orthogonal conjugation of all linear weights. As a result, the server performs the full forward pass entirely in the rotated basis and never observes unrotated hidden states. Experiments on GPT-2 and Llama 3.2 1B models fine-tuned on PubMed show that orthogonal obfuscation eliminates direct cosine nearest-neighbor inversion and reduces token recovery from over 35% top-10 to at most 1.3%, while increasing perplexity by only 0.4% after fine-tuning. These results indicate that enforcing symmetry at the architectural level can provide a practical defense for privacy-preserving LLM inference without noise injection or heavy cryptographic machinery.

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

On The Effectiveness-Fluency Trade-Off In LLM Conditioning: A Systematic Study

Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the involved trade-offs remains elusive. Current approaches to conditioning are often evaluated with a narrow focus on their effectiveness at injecting or removing a target concept, neglecting generation quality. We systematically investigate a range of conditioning methods in both injection and removal scenarios. We find that efficient steering methods frequently achieve conditioning at a steep cost to fluency. Furthermore, we identify a critical yet previously overlooked interaction with the training paradigm: activation steering methods are far less effective on instruction-tuned models than on their base counterparts. Simple prompting and full-fledged supervised fine-tuning, on the other hand, are viable options for concept injection, but are not as good at concept removal. Finally, cheaply computed textual metrics highly correlate to costly LLM-as-judge scores, and provide insights on the behavior of conditioning methods.

13.
arXiv (CS.AI) 2026-06-16

AI Pluralism and the Worlds It Misses

arXiv:2606.16167v1 Announce Type: new Abstract: AI pluralism is often framed as a problem of representing diverse values, preferences, users, or outputs. This paper argues that this framing is incomplete because AI systems also impose ontologies: they define what counts as an entity, relation, feature, harm, benefit, and valid form of evidence. We define ontological flattening as the conversion of situated, contested, and historically specific meanings into a restricted technical category, proxy, aggregation rule, or benchmark target that is treated as neutral and difficult to contest. The paper develops a bounded conceptual and qualitative synthesis across value pluralism, pluralistic alignment, participatory and democratic AI, procedural justice, science and technology studies, accountability research, aggregate themes from 11 expert interviews, and three urban AI companion cases. The cases illustrate how pluralistic methods can improve or structure model behavior while still compressing categories, proxies, aggregation rules, and revision rights before affected actors have procedural standing. We introduce Pluralistic Lifecycle Governance (PLG) as a preliminary qualitative audit scaffold for documenting ontological openness, epistemic inclusion, procedural authority, evaluation pluralism, and lifecycle accountability. PLG is not presented as a validated scoring instrument; it is a framework for making the evidence and governance conditions of pluralistic AI explicit.

14.
arXiv (CS.CL) 2026-06-15

Chronological Thinking in Full-Duplex Spoken Dialogue Language Models

Recent advances in spoken dialogue language models (SDLMs) reflect growing interest in shifting from turn-based to full-duplex systems, where the models continuously perceive user speech streams while generating responses. This simultaneous listening and speaking design enables real-time interaction and the agent can handle dynamic conversational behaviors like user barge-in. However, during the listening phase, existing systems keep the agent idle by repeatedly predicting the silence token, which departs from human behavior: we usually engage in lightweight thinking during conversation rather than remaining absent-minded. Inspired by this, we propose Chronological Thinking, an on-the-fly conversational thinking mechanism that aims to improve response quality in full-duplex SDLMs. Specifically, chronological thinking presents a paradigm shift from conventional LLM thinking approaches, such as Chain-of-Thought, purpose-built for streaming acoustic input. (1) Strictly causal: the agent reasons incrementally while listening, updating internal hypotheses only from past audio with no lookahead. (2) No additional latency: reasoning is amortized during the listening window; once the user stops speaking, the agent halts thinking and begins speaking without further delay. Experiments demonstrate the effectiveness of chronological thinking through both objective metrics and human evaluations show consistent improvements in response quality. Furthermore, chronological thinking robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.

15.
arXiv (CS.CV) 2026-06-11

Weakly Supervised Segmentation as Semantic-Based Regularization

Weakly supervised semantic segmentation (WSSS) trains dense pixel-level segmentation models from partial or coarse annotations such as bounding boxes, scribbles, or image-level tags. While recent work leverages foundation models such as the Segment Anything Model (SAM) to generate pseudo-labels, these approaches typically depend on heuristic prompt choices and offer limited ways to incorporate prior knowledge or heterogeneous labels. We address this gap by taking a neurosymbolic perspective: integrating differentiable fuzzy logic with deep segmentation models. Weak annotations and domain-specific priors are unified as continuous logical constraints that fine-tune SAM under weak supervision. The refined foundation model then produces improved pseudo-labels, from which we train a second-stage prompt-free segmentation model. Experiments on Pascal VOC 2012 and the REFUGE2 optic disc/cup segmentation dataset show that our logic-guided fine-tuning yields higher-quality pseudo-labels, leading to state-of-the-art segmentation accuracy that often exceeds densely supervised baselines.

16.
arXiv (CS.CL) 2026-06-15

The Holistic Storage of Verb+Up Phrases in Text-based and Audio-based Language Models

A crucial aspect of linguistic capability is the ability to trade off between stored representations and abstract knowledge: one must retrieve learned representations, but also generate novel ones by applying productive rules. While recent work has examined abstract knowledge in language models, holistic storage of multi-word units has received far less attention. We probe internal representations in text-based LLMs and an ASR model, testing whether V+up phrasal verbs develop distinct representations as a function of frequency and predictability. All models show evidence of holistic storage driven by frequency and predictability, further supporting usage-based theories of language.

17.
medRxiv (Medicine) 2026-06-15

Entity-Aware Generation of Synthetic Clinical Progress Notes for Prostate Cancer using Large Language Model

Objectives: This study investigates large language models (LLMs) for clinical entity projection across substantial textual transformation. Specifically, we evaluate whether entities annotated in Spanish prostate cancer case reports can be preserved and explicitly projected when the source narratives are transformed into hospital-style clinical progress notes. Entity projection is treated as a generation-driven task, allowing paraphrase, condensation and narrative reorganisation, providing that clinically relevant entities remain recoverable as structured annotations. Methods: A corpus of 109 Spanish prostate cancer case reports was annotated using a silver-standard pipeline combining Spanish biomedical named-entity recognition with rule-based prostate-specific antigen (PSA) and Gleason extractors. The resulting silver-standard annotations were validated on a subset of generated notes against a gold-standard consensus produced by medical experts in prostate cancer. Four LLMs were evaluated for note generation and entity projection: GPT-5.4 Nano, Qwen 3.5:35B-A3B, GLM5 and Claude Sonnet 4.6. Entity-to-Entity (E2E) generation used XML-annotated cases as RAG-supported input, whereas Text-to-Entity (T2E) generation required models to generate and annotate notes directly from plain text cases. Zero-shot and few-shot prompting were tested. Projection quality was measured using precision, recall and F1-score, and complemented by LLM-as-a-judge evaluation using Kimi K2.6. Results: E2E consistently outperformed T2E, indicating that explicit entity-enriched in- put substantially facilitates entity preservation and localisation. GLM5 achieved the best E2E zero-shot result (F1 = 0.915), followed by Claude Sonnet 4.6 (F1 = 0.896). In T2E, few-shot prompting improved performance, with Claude Sonnet 4.6 reaching the highest score (F1 =0.718). Age, Gleason, Disease, Procedure, Duration and negation-related entities were robustly projected, whereas PSA and Dose showed less stable behaviour. Conclusion: LLMs can generate clinically plausible synthetic prostate cancer evolution notes while preserving a substantial proportion of source entities, particularly when explicit semantic annotations are provided as input. However, the lower and more variable performance observed in T2E highlights the difficulty of jointly generating clinical narratives and projecting entities without source-side information, especially for numerical and measure-related entities.

18.
arXiv (CS.AI) 2026-06-17

Membership Inference Attacks against Large Audio Language Models

arXiv:2603.28378v2 Announce Type: replace-cross Abstract: We present the first systematic Membership Inference Attack (MIA) evaluation of LALMs. Using Multi-modal Blind Baselines based on textual, spectral and prosodic features, we demonstrate that common audio datasets exhibit near-perfect train/test separability (AUC ~ 1.0) even without model inference, thus MIA may primarily detect distribution shift. We therefore introduce a blind-baseline protocol to control for this confound. Under this protocol, we identify that the distribution-matched datasets enable reliable MIA evaluation without distribution-shift artifacts. We benchmark multiple MIA methods and conduct modality disentanglement experiments on these datasets. The results reveal that LALM memorization is cross-modal, arising only from binding a speaker's vocal identity with its text. These findings establish a principled standard for auditing LALMs beyond spurious correlations. Our codebase is available at https://github.com/snooow1029/ALM_MIA.

19.
arXiv (CS.CV) 2026-06-11

Right Regions, Wrong Labels: Semantic Label Flips in Segmentation under Correlation Shift

The robustness of machine learning models can be compromised by spurious correlations between non-causal features in the input data and target labels. A common way to test for such correlations is to train on data where the label is strongly tied to some non-causal cue, then evaluate on examples where that tie no longer holds. This idea is well established for classification tasks, but for semantic segmentation the specific failure modes are not well understood. We show that a model may achieve reasonable overlap while assigning the wrong semantic label, swapping one plausible foreground class for another, even when object boundaries are largely correct. We focus on this semantic label-flip behaviour and quantify it with a simple diagnostic (Flip) that counts how often ground truth foreground pixels are assigned the wrong foreground identity while remaining predicted as foreground. In a setting where category and scene are correlated during training, increasing the correlation consistently widens the gap between common and rare test conditions and increases these within-object label swaps on counterfactual groups. Overall, our results motivate assessing segmentation robustness under distribution shift beyond overlap by decomposing foreground errors into correct pixels, flipped-identity pixels, and missed-to-background pixels. We also propose an entropy-based, ground truth label-free `flip-risk' score, which is computed from foreground identity uncertainty, and show that it can flag flip-prone cases at inference time. Code is available at https://github.com/acharaakshit/label-flips.

20.
arXiv (CS.CV) 2026-06-15

Representation Forcing for Bottleneck-Free Unified Multimodal Models

Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.

21.
arXiv (CS.AI) 2026-06-16

The Distributed Detectability Band Against Marginal-Preserving Attacks

arXiv:2606.10456v2 Announce Type: replace-cross Abstract: AI-control monitors score individual agent actions to detect misbehavior, but real harm can be distributed across many benign-looking steps, each individually below any per-step alarm. We construct a marginal-preserving, correlation-encoded distributed-sabotage attack using a Gaussian-copula AR(1) construction: the per-step monitor-score marginal is held exactly equal to benign, so mean, max, top-k tail, and threshold monitors (Monitor A) are defeated by construction, while harm is encoded in the temporal correlation structure. We sequence the paper around three reviewer-mandated gates. (1) Realizability gate: the stealthy attack achieves KS-distance to benign of 0.013 (effectively zero) at all tested harm levels up to 3.0, confirming that harm is fully decoupled from the per-step marginal and realizability is not harm-limited. (2) Monitor-A-vs-B reconciliation: we show formally that the attack, built against Monitor A's score marginal, remains marginal-preserving under a different-score Monitor B (the correlation/sequence family: CUSUM, SPRT, HMM-LR, runs test, autocorrelation, windowed logistic), and scope worst-case claims to score functions that admit a temporal signature. (3) Non-empty detectability band: Monitor A achieves AUC 0.52 (chance); Monitor B spans AUC 0.79-0.97 at the same 1% FPR target, and as harm is amortized over more steps Monitor A collapses to chance while Monitor B holds at AUC ~0.95. These results demonstrate a non-empty detectability band and characterize the sub-threshold sabotage frontier: distribution-shape monitors fail by construction; temporal-correlation monitors can detect but are not trivially optimal.

22.
arXiv (CS.CL) 2026-06-17

The Benchmark Illusion: Pruned LLMs Can Pass Multiple Choice but Fail to Answer

Compressing large language models reduces memory use and inference cost, but it can also create failures that standard benchmarks miss. A pruned model may still perform well on multiple-choice evaluations, yet fail to answer the same question in open generation. We ask what pruning changes: does it erase the correct answer, or does it make the answer harder to produce as the top output? We study this question with multilingual question answering, tracking the same questions before and after pruning. We find a benchmark illusion. Under high-sparsity pruning, especially Wanda, models often fail in greedy open generation while still selecting the correct answer under multiple-choice scoring. In these recognition-only errors, the answer is usually not gone, but demoted: it often reappears with beam search, sampling, or one in-context example. Overall, multiple-choice benchmarks can overstate the usability of compressed LLMs, creating an evaluation blind spot. Compressed models should be tested on what they can produce, not only on what they can recognize.

23.
arXiv (CS.AI) 2026-06-19

Simulation of Language Evolution under Regulated Social Media Platforms: A Synergistic Approach of Large Language Models and Genetic Algorithms

arXiv:2502.19193v2 Announce Type: replace-cross Abstract: Social media platforms frequently impose restrictive policies to moderate user content, prompting the emergence of creative evasion language strategies. This paper presents a multi-agent framework based on Large Language Models (LLMs) to simulate the iterative evolution of language strategies under regulatory constraints. In this framework, participant agents, as social media users, continuously evolve their language expression, while supervisory agents emulate platform-level regulation by assessing policy violations. To achieve a more faithful simulation, we employ a dual design of language strategies (constraint and expression) to differentiate conflicting goals and utilize an LLM-driven GA (Genetic Algorithm) for the selection, mutation, and crossover of language strategies. The framework is evaluated using two distinct scenarios: an abstract password game and a realistic simulated illegal pet trade scenario. Experimental results demonstrate that as the number of dialogue rounds increases, both the number of uninterrupted dialogue turns and the accuracy of information transmission improve significantly. Furthermore, a user study with 40 participants validates the real-world relevance of the generated dialogues and strategies. Moreover, ablation studies validate the importance of the GA, emphasizing its contribution to long-term adaptability and improved overall results.

24.
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.

25.
arXiv (CS.LG) 2026-06-17

A Diffusion Approximation for Temporal-Difference Learning with Linear Features under Markovian Noise

arXiv:2606.18183v1 Announce Type: cross Abstract: Temporal difference (TD) learning with linear function approximation is a core method for policy evaluation. Its classical continuous-time description is an ordinary differential equation (ODE), which captures the asymptotic mean dynamics but neglects stochastic fluctuations determining the error floor. We introduce a stochastic differential equation (SDE) approximation for linear TD(0) under Markovian noise. The resulting model distinguishes the contraction dynamics governed by the projected Bellman operator from the influence of Markovian sampling. As a consequence, the model explains the constant-stepsize error floor through the interaction between Markovian long-run covariance and the contraction geometry of the projected Bellman operator.