EXPLORE THE ARCHIVE

Language models

Explore arXiv papers about language models. Read abstracts, discover authors and related research, and follow the original paper for methods, results, and the latest version.

895 records · Page 3Linked to original sources

arXiv2026

Truthful AI Advisors: A Pre-Specified Benchmark for Large Language Model Honesty Under Preference Misalignment

Large language models are increasingly deployed as advisors whose objective is not aligned with the user's: recommenders optimize for engagement, sales assistants for purchases. Whether they stay truthful when honesty conflicts with their own payoff is a core alignment question. We turn the canonical Crawford-Sobel cheap-talk model into a pre-specified benchmark for LLM honesty under preference misalignment, in which theory supplies an exact oracle. A sender observes a state omega in [0,1], wants the receiver's action near omega+b, and sends one costless message to a receiver whose ideal action is omega. For the positive-bias grid b in {0.01,0.04,0.08,0.12} the exact most-informative partition sizes are 7,4,3,2, with oracle normalized mutual information 0.5294, 0.3268, 0.2205, 0.1829. Extending a pre-registered 4-model run of 12,000 sender calls to eight models across two capability tiers and 39,569 logged calls, all models over-reveal relative to the most-informative equilibrium by 1.8 to 4.5x: pooled normalized mutual information stays at 0.82-0.96 where the oracle prescribes 0.18-0.53. Informativeness declines with bias as predicted (beta = -1.71, t = -7.50) but never approaches the strategic optimum; rather than coarse partitions, models show near-full revelation with a constant upward offset tracking their bias (linear exaggeration). A structural hint separates capability from propensity: told the equilibrium partition size, reasoning models state a correct Crawford-Sobel cell in 0.20-0.99 of messages while the non-reasoning tier never exceeds 0.005. The capability is present but goes unexercised unless asked for, locating the failure in propensity rather than competence. A decoder ablation shows the finding is recoverable only when the receiver reads the sender's stated number: an embedding-only decoder mis-reads the same data as near-babbling.

cs.LG↗

arXiv2026

A Simple Method to Enhance Pre-trained Language Models with Speech Tokens for Classification

This paper presents a simple method that allows to easily enhance textual pre-trained large language models with speech information, when fine-tuned for a specific classification task. A classical issue with the fusion of many embeddings from audio with text is the large length of the audio sequence compared to the text one. Our method benefits from an existing speech tokenizer trained for Audio Speech Recognition that output long sequences of tokens from a large vocabulary, making it difficult to integrate it at low cost in a large language model. By applying a simple lasso-based feature selection on multimodal Bag-of-Words representation, we retain only the most important audio tokens for the task, and adapt the language model to them with a self-supervised language modeling objective, before fine-tuning it on the downstream task. We show this helps to improve the performances compared to an unimodal model, to a bigger SpeechLM or to integrating audio via a learned representation. We demonstrate its effectiveness on Argumentative Fallacy Detection and Classification tasks where audio was previously believed counterproductive, and affective computing tasks on a widely-used dataset. We also provide an in-depth analysis of the method, showing that even a random audio token selection helps enhancing the unimodal model. Our code is available [online](https://github.com/salocinc/EMNLP26SpeechTokLLM/).

cs.CL↗

arXiv2026

Evaluating the Semantic Specificity of Representation Steering in Language Models

Localized Representation Steering (LRS) is widely used to correct reasoning pathologies in large language models. However, standard benchmark evaluations can easily be fooled by superficial label overrides, creating a false impression of reasoning circuit repairs. In this work, we propose Cross-Rule Transfer (CRT), a diagnostic framework that audits representational interventions by evaluating them on rule families where the model is natively competent. Evaluating late-layer LRS for a widespread logical failure, contradiction blindness, reveals that the intervention merely injects a global label bias: applying the steering vector to rules the model already handles correctly (99.6% baseline) degrades performance to 40.4% by forcing false contradiction predictions. We support this diagnosis with four complementary controls (direct logit bias equivalence, control vector label-flipping, cross-model grafting, and early-layer steering checks), providing a rigorous methodology to distinguish genuine reasoning repairs from superficial label overrides.

cs.CL↗

arXiv2026

A Survey on Rubric-Guided Reinforcement Learning for Language Models

Reinforcement learning from human feedback (RLHF) has become the dominant paradigm for aligning large language models (LLMs) with human preferences. However, traditional RLHF relies on scalar reward signals that lack interpretability and fail to capture the multifaceted nature of response quality. Rubric-guided reinforcement learning addresses these limitations by introducing structured, interpretable evaluation criteria, or rubrics, as the backbone of reward design, feedback generation, and policy optimization. In this survey, we introduce a Bayesian framework that defines constitutions as prior distributions $P(R)$ over evaluation criteria and rubrics as conditional instantiations $R_x \sim P(R|x)$. Under this unified view, we present a taxonomy of rubric-guided RL along the prior-posterior axis, covering constitutional AI, instance-specific rubrics, process-level supervision, self-evolving rubrics, and their agentic and multimodal extensions. Furthermore, as rubrics are natural-language artifacts, we present a linguistic analysis of how granularity trade-offs, semantic drift, and linguistic reward hacking impact alignment reliability, identifying key open problems for future research.

cs.CL↗

arXiv2026

OpenStamp: A Watermark for Open-Source Language Models

With the growing prevalence of large language model (LLM) generated content, watermarking is considered a promising approach for attributing text to LLMs and distinguishing it from human-written content. A prominent class of techniques embeds subtle but detectable signals in generated text by modifying token sampling probabilities. However, such methods are unsuitable for open-source models, where users have white-box access and can easily disable watermarking during inference. In this work, we introduce OpenStamp, a watermarking technique that encodes the watermarking logic directly into the model weights by modifying only the final projection, or unembedding, layer. Through experiments across two models, we show that OpenStamp achieves superior detection performance, with minimal degradation in model capabilities compared to prior methods. The implanted watermark is explicitly designed, and empirically confirmed, to be more robust to paraphrasing attacks and harder to scrub off through post-hoc fine-tuning than prior open-source watermarks. To enable developers to watermark their models, we release our code alongside watermarked versions of 4 popular open-source models.

cs.CL↗

arXiv2026

Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models

Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1--28.0x speedup at 8K--32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Code is available at https://github.com/menik1126/Prefilling-dLLM.

cs.CL↗

arXiv2026

SUP-MIMIC: A Multi-Task Clinical Diagnosis Benchmark for Evaluating LLMs' Robustness to Contradictory Evidence

Current evaluations of large language models (LLMs) primarily focus on factual knowledge retrieval, overlooking the fundamental challenge of navigating the complex, non-bijective mappings between clinical indicators and diagnoses. Existing benchmarks fail to assess whether large language models truly possess the reasoning capability required for diagnostic ambiguity scenarios, where identical clinical presentations may correspond to different etiologies, and diagnostic convergence scenarios, where heterogeneous symptoms ultimately indicate the same disease. To address this issue, we propose SUP-MIMIC, a multi-task framework utilizing MIMIC-IV-v3.1 that comprises Basic Assessment (BA), Diagnostic Divergence Task (DDT), and Diagnostic Convergence Task (DCT). Specifically, DDT is designed to evaluate the model's "one-to-many" disambiguation capability among phenotypically similar cases, while DCT assesses the model's ability to identify "many-to-one" diagnostic patterns across different pathophysiological pathways. Comprehensive evaluation of state-of-the-art LLMs reveals substantial performance degradation on DDT and DCT compared to baseline tasks, exposing a systemic reliance on statistical shortcuts over genuine causal reasoning. Our findings further highlight a conservative bias toward "healthy" predictions, implying non-trivial risks for missed diagnoses in realistic medical settings. This work establishes a rigorous methodology for quantifying clinical reasoning robustness and provides a roadmap for enhancing the safety of language models in clinical medicine.

cs.CL↗

arXiv2026

Investigating Social Bias Changes in Quantized Language Models

Post-training quantization reduces the memory needed to run large language models but alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced bias flipping, in which quantization causes models to change responses from biased to unbiased and vice versa, up to 21% of the time, despite no change in aggregate bias scores. These flips are strongly associated with model uncertainty, where the responses with high uncertainty are 3-11x more likely to change than the confident ones. Quantization strength amplifies this effect, with 4-bit quantized models exhibiting 4-6x more behavioral changes than 8-bit quantized models. Critically, these changes create asymmetric impacts across demographic groups, where bias can worsen by up to 18.6% for some groups while improving by 14.1% for others, yielding misleadingly neutral aggregate outcomes. Larger models show no consistent robustness advantage, and group-specific shifts vary unpredictably across model families. Our findings demonstrate that compression fundamentally alters bias patterns, requiring crucial post-quantization evaluation and interventions to ensure reliability in practice.

cs.CL↗

arXiv2026

Why Knowing Both Hops Is Not Enough: Understanding Two-Hop Generalization in Language Models

Large language models (LLMs) can solve complex multi-hop problems yet exhibit puzzling failures on simple two-hop queries: although a model may correctly store each individual hop, it often fails to combine them. To understand the internal mechanisms of this phenomenon, we train transformers from scratch in a controlled symbolic environment. Our experiments reveal a pattern in two-hop generalization: models generalize reliably when the second hop follows the training distribution, but always fail when it deviates. Through mechanistic analysis, we provide a complete explanation for these distinct generalization behaviors: in settings where models generalize successfully, performance is driven by the emergence of consistent intermediate representations for the same entities across contexts, whereas failures on settings where the second hop is out-of-distribution arise from a mismatch across layers: lower layers correctly construct these intermediate representations, but upper layers, while trained on corresponding atomic facts, primarily learn to map them to outputs rather than to reason over them. Driven by this insight, we propose a recurrent-style training strategy, which enables transformers to reuse their reasoning circuitry across input forms and substantially improves generalization on out-of-distribution two-hop queries. Our data and code are available at https://github.com/zzl-strong/two_hop .

cs.CL↗

arXiv2026

Vectors from Larger Language Models Predict Human Reading Time and fMRI Data More Poorly when Dimensionality Expansion is Controlled

The impressive linguistic abilities of large language models (LLMs) have recommended them as models of human sentence processing, with some conjecturing a positive 'quality-power' relationship, in which language models' (LMs') fit to psychometric data continues to improve as their ability to predict words in context increases. This is important because it might suggest that elements of LLM architecture reflect the architecture of the human sentence processing faculty, and that any inadequacies in predicting human reading time and brain imaging data may be attributed to insufficient model complexity, which recedes as larger models become available. But recent studies have shown this scaling inverts after a point, as LMs become excessively large and accurate, when information-theoretic surprisal is used as a predictor. Other studies propose the use of entire vectors from differently sized LLMs, still showing positive scaling, casting doubt on the value of surprisal as a predictor, but do not control for dimensionality expansion using untrained LLMs with more than 1.6B parameters. This study evaluates scaling of LLM vector predictors controlled using untrained LLMs with up to 66B parameters. Results show that inverse scaling obtains, and moreover the contribution of trained LMs over corresponding untrained LMs drops to zero at around a few billion parameters on most datasets.

cs.CL↗

arXiv2026

Man Made Language Models? Evaluating LLMs' Perpetuation of Masculine Generics Bias

Instruct-based large language models (LLMs) have been shown to propagate and even amplify gender bias when prompted with contextually constrained instructions (e.g., writing a text from a description or selecting a gendered pronoun). However, little attention has been paid to biases in responses to contextually unconstrained (generic) instructions conveyed by gendered language, particularly masculine generics (MG). MG, found in many gender-marked languages, denote the use of the masculine gender as a supposedly neutral reference to mixed-gender groups or individuals whose gender is unknown or non-binary. Yet, psycholinguistic studies demonstrate that MG are not neutral and systematically induce gender bias. This study investigates how both local and proprietary LLMs are MG-biased when responding to generic prompts in French, examining LLMs' MG bias rates and use of gender-fair language (GFL). We create a 16k+ human noun database from existing lexical resources and evaluate six LLMs on four instruction-response datasets under two conditions: prompts with and without MG. Overall, we find that $\approx$27.57% of LLMs' responses to MG-filtered generic instructions are MG-biased ($\approx$78.55% with MG-containing prompts). Moreover, we find that LLMs rarely use GFL spontaneously. These findings highlight the persistence of MG bias in LLM outputs and models' limited tendency towards GFL strategies.

cs.CL↗

arXiv2026

An evidence-guided reinforcement learning method to improve psychiatric reasoning in small language models

Privacy and computational constraints limit the use of large language models in psychiatry, while adapting small language models (SLMs) often requires substantial data and expert annotation. We developed ClinMPO, an evidence-guided reinforcement-learning framework guided by the psychiatrist-defined Clinical Psychiatry Thinking Strategy (CPTS). ClinMPO uses ClinRM, a reward model trained on 18,569 question--answer pairs from 4,474 psychiatry articles. We evaluated four Qwen3 sizes on 1,737 model-screened questions. ClinMPO outperformed Base, supervised fine-tuning and standard group relative policy optimization across scales. From responses by 300 senior pre-licensure medical students, we established the human baseline, a medical-student reference. The 4B model approached this baseline, whereas the 8B model surpassed it and ranked first among 31 models and post-training variants. ClinMPO improved performance across two complementary schemes covering ICD-11 diagnostic categories and psychiatric practice competencies. Blinded assessment by three clinicians showed improved rationale quality across CPTS criteria. These findings highlight how existing clinical evidence and specialist knowledge can be incorporated into the development of medical AI systems through evidence-guided learning.

cs.CL↗

arXiv2026

Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval

Artificial intelligence (AI) hiring tools have revolutionized resume screening, and large language models (LLMs) have the potential to do the same. However, given the biases which are embedded within LLMs, it is unclear whether they can be used in this scenario without disadvantaging groups based on their protected attributes. In this work, we investigate the possibilities of using LLMs in a resume screening setting via a document retrieval framework that simulates job candidate selection. Using that framework, we then perform a resume audit study to determine whether a selection of Massive Text Embedding (MTE) models are biased in resume screening scenarios. We simulate this for nine occupations, using a collection of over 500 publicly available resumes and 500 job descriptions. We find that the MTEs are biased, significantly favoring White-associated names in 85.1\% of cases and female-associated names in only 11.1\% of cases, with a minority of cases showing no statistically significant differences. Further analyses show that Black males are disadvantaged in up to 100\% of cases, replicating real-world patterns of bias in employment settings, and validate three hypotheses of intersectionality. We also find an impact of document length as well as the corpus frequency of names in the selection of resumes. These findings have implications for widely used AI tools that are automating employment, fairness, and tech policy.

cs.CY↗

arXiv2026

Tastes without distinction: silicon samples and the synthetic construction of tastes

Large-language models have proven to be remarkable if inconsistent parrots of public attitudes and opinions. The extent to which LLMs are able to produce reasonable approximations of cultural taste remains an open empirical question that becomes more urgent by the day, with market research companies already offering provisional 'synthetic' survey panels and the contamination of standard survey data from LLM-generated responses. In this study, we build on past work on silicon sampling by extending considerations of their ecological, relational, and positional fidelity in the doomain of cultural tastes. We use large-language models from OpenAI, Anthropic, and DeepSeek to produce 554,940 silicon surrogates of survey respondents from the Survey of Public Participation in the Arts (SPPA). We find these silicon surrogates' tastes to be highly stylized facsimiles of human tastes. First, silicon samples are super-omnivorous with a systematic postive-bias for liking. These individual-level bias of silicon samples are not well-explained by the WEIRD-bias often discussed in the literature. Second, the complex relationality in real taste structures is completely distorted among silicon samples. Third, very little of the known cultural alignment between tastes and social space are preserved. Silicon samples juvenilize age-taste associations, resurrect anachronistic class-taste associations, and caricaturize gender- and race-taste associations. Key words: AI, taste, consumption, culture, silicon sampling, meta-analysis.

cs.CL↗

arXiv2026

Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations?

Power differences shape human communication through well documented socio cognitive effects, including language coordination, pronoun usage, authority bias, and harmful compliance. We examine whether large language models (LLMs) exhibit similar behaviors when assigned high or low status personas. Using personas from diverse professions, we simulate multi turn, power asymmetric dialogues (e.g., principal teacher, justice lawyer) and measure (i) language coordination, (ii) pronoun usage, (iii) persuasion success, and (iv) compliance with unsafe requests. Our results show that LLMs show key socio-cognitive effects of power, albeit with nuances and variability, linking simulated interactions to both desirable and unsafe behaviors.

cs.CL↗

arXiv2026

Demystifying Reinforcement Learning Post-Training of Language Models

Reinforcement learning (RL) post-training has emerged as a powerful framework for enhancing the capabilities of large language models (LLMs), enabling impressive reasoning, math, and coding capabilities. Yet for many researchers and practitioners, the principles behind classical RL remain a "black box". In this work, we deconstruct the RL post-training algorithm, investigating each step to clarify what is actually happening beneath the surface. By isolating the mechanics of RL with Verifiable Rewards in a controlled and simplified environment, we examine how RL outcomes are shaped by the base model's prior distribution, the granularity of the reward signal, the diversity of the prompt distribution, and model scale. We use the entropy of the policy's output distribution as a lens to compare the distributions learned through pretraining, SFT, and RL post-training, revealing how each stage shapes model certainty. Our investigation sheds light on how these choices interact to affect post-training success. For example, we show that the effect of so-called 'spurious rewards' depends on the prompt distribution used for post-training. We also provide insight into why the success of RL post-training depends on whether the base model already places sufficient probability mass on the desired behavior, linking it to the classical concept of exploration in RL. Ultimately, we provide this primer as a resource to those in the NLP community wishing to incorporate RL as a tool in their toolbox.

cs.LG↗

arXiv2026

ReVA: A Region-Aware Visual Assistant for Visually Grounded Question Answering

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in Visual Question Answering (VQA), yet they continue to struggle with questions requiring precise spatial reasoning and fine-grained visual understanding. These limitations often manifest as object, attribute, and spatial hallucinations, where models generate confident but visually unsupported responses due to insufficient region-level and fine-grained visual grounding. To address this challenge, we propose ReVA, a region-aware VQA model that employs a frozen CLIP ViT-L/14 Vision Transformer (ViT) and a Qwen2.5-7B-Instruct large language model (LLM) connected through a dual bridge that aligns both whole-image and region-level representations with the LLM's embedding space. The image bridge maps final transformer block features into image tokens. The region bridge maps cropped features from enriched intermediate features across ViT blocks so early texture and later object cues are more evident, into K region tokens for every bounding box. ReVA uses a detector stack that supplies automatic zero-shot bounding boxes that are both question-agnostic and question-dependent, using RAM++ (Recognize Anything Model), spaCy, and Grounding DINO. The image tokens and region tokens are concatenated as an LLM prompt prefix to jointly encode scene-level context and fine-grained regional evidence when answering questions. Evaluated on VQAv2, MMBench, POPE, and SEED-Bench, ReVA achieves 82.85% mean F1 on POPE, compared with 81.14% for an image-token baseline without region tokens. These results demonstrate that explicit region-aware visual representations reduce object hallucination and improve the factual grounding of MLLMs.

cs.CL↗

arXiv2026

Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task Allocation

Enhancing the reasoning capabilities of large language model (LLM) agents requires effective orchestration of diverse expert models and tools. However, existing frameworks typically call APIs, based on coarse-grained matching between tasks and the functions of expert models or tools, while overlooking critical factors such as performance variability and cost efficiency among functionally similar alternatives. To address this, we propose Agora, a framework that uses a confidence-calibrated auction to dynamically allocate tasks to expert models and tools. By treating reasoning steps as tradeable items, Agora bases allocation on calibrated competence rather than raw confidence. Across five main benchmarks, Agora improves or remains competitive with single-model, routing, and cascade baselines under matched candidate pools.

cs.AI↗