Peer Revenue Past citations are no guarantee of future returns

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Computation and Language

cs.CL Watch Computation and Language

Wednesday 18 February

Computation and Language

Long-Tail Knowledge in Large Language Models: Taxonomy, Mechanisms, Interventions and Implications

Sanket Badhe, Deep Shah and Nehal Kathrotia

Large language models (LLMs) are trained on web-scale corpora that exhibit steep power-law distributions, in which the distribution of knowledge is highly long-tailed, with most appearing infrequently. While scaling has improved average-case performance, persistent failures on low-frequency, domain-specific, cultural, and temporal knowledge remain poorly characterized. This paper develops a structured taxonomy and analysis of long-Tail Knowledge in large language models, synthesizing prior work across technical and sociotechnical perspectives. We introduce a structured analytical framework that synthesizes prior work across four complementary axes: how long-Tail Knowledge is defined, the mechanisms by which it is lost or distorted during training and inference, the technical interventions proposed to mitigate these failures, and the implications of these failures for fairness, accountability, transparency, and user trust. We further examine how existing evaluation practices obscure tail behavior and complicate accountability for rare but consequential failures. The paper concludes by identifying open challenges related to privacy, sustainability, and governance that constrain long-Tail Knowledge representation. Taken together, this paper provides a unifying conceptual framework for understanding how long-Tail Knowledge is defined, lost, evaluated, and manifested in deployed language model systems.
Read the full abstract

Computation and Language

The Validity of Coreference-based Evaluations of Natural Language Understanding

Ian Porada

In this thesis, I refine our understanding as to what conclusions we can reach from coreference-based evaluations by expanding existing evaluation practices and considering the extent to which evaluation results are either converging or conflicting. First, I analyze standard coreference evaluations and show that their design often leads to non-generalizable conclusions due to issues of measurement validity - including contestedness (multiple, competing definitions of coreference) and convergent validity (evaluation results that rank models differently across benchmarks). Second, I propose and implement a novel evaluation focused on testing systems' ability to infer the relative plausibility of events, a key aspect of resolving coreference. Through this extended evaluation, I find that contemporary language models demonstrate strong performance on standard benchmarks - improving over earlier baseline systems within certain domains and types of coreference - but remain sensitive to the evaluation conditions: they often fail to generalize in ways one would expect a human to be capable of when evaluation contexts are slightly modified. Taken together, these findings clarify both the strengths, such as improved accuracy over baselines on widely used evaluations, and the limitations of the current NLP paradigm, including weaknesses in measurement validity, and suggest directions for future work in developing better evaluation methods and more genuinely generalizable systems.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

ModalImmune: Immunity Driven Unlearning via Self Destructive Training

Rong Fu, WeiZhi Tang, Ziming Wang et al. (+6 more)

Multimodal systems are vulnerable to partial or complete loss of input channels at deployment, which undermines reliability in real-world settings. This paper presents ModalImmune, a training framework that enforces modality immunity by intentionally and controllably collapsing selected modality information during training so the model learns joint representations that are robust to destructive modality influence. The framework combines a spectrum-adaptive collapse regularizer, an information-gain guided controller for targeted interventions, curvature-aware gradient masking to stabilize destructive updates, and a certified Neumann-truncated hyper-gradient procedure for automatic meta-parameter adaptation. Empirical evaluation on standard multimodal benchmarks demonstrates that ModalImmune improves resilience to modality removal and corruption while retaining convergence stability and reconstruction capacity.
Read the full abstract

All authors Rong Fu, WeiZhi Tang, Ziming Wang, Jia Yee Tan, Zijian Zhang, Zhaolu Kang, Muge Qi, Shuning Zhang, Simon Fong

Computation and Language

Beyond Learning: A Training-Free Alternative to Model Adaptation

Namkyung Yoon, Kyeonghyun Yoo, Wooyong Jung et al. (+2 more)

Despite the continuous research and evolution of language models, they sometimes underperform previous versions. Existing approaches to overcome these challenges are resource-intensive, highlighting the need for alternatives that enable immediate action. We assume that each language model has a local module inside that is suitable for a specific function. First, this work identifies a set of modules showing consistent and local activation changes under an inference workload through activation-based analysis. Subsequently, we transplant an internal module that is properly activated for a specific task into the target model, leading to immediate and measurable functional changes without additional training or fine-tuning. To experimentally demonstrate the effectiveness of the transplant technique, we quantify the relationship between transplant strength and performance improvement under different conditions for two language models. In the cross-generation setting, we find that transplanting activation-selected modules can substantially improve the underperforming model, reaching up to twice the target baseline and achieving gap-based recovery above 100%. Moreover, in transplant experiments between a base model and its instruction-tuned counterpart, transplantation improves the underperforming model toward the stronger baseline, yielding up to about 2.33 times the target baseline with gap-based recovery reaching up to 100% in the best case. These results show that meaningful capacity transfer can be realized through the implantation of highly localized modules implied by language models. Overall, this work provides empirical evidence for task-localized modularity in language models and presents a new research area: model transplantation.
Read the full abstract

All authors Namkyung Yoon, Kyeonghyun Yoo, Wooyong Jung, Sanghong Kim, Hwangnam Kim

Computation and Language · Cross-listed from cs.AI

Learning Personalized Agents from Human Feedback

Kaiqu Liang, Julia Kruk, Shengyi Qian et al. (+9 more)

Modern AI agents are powerful but often fail to align with the idiosyncratic, evolving preferences of individual users. Prior approaches typically rely on static datasets, either training implicit preference models on interaction history or encoding user profiles in external memory. However, these approaches struggle with new users and with preferences that change over time. We introduce Personalized Agents from Human Feedback (PAHF), a framework for continual personalization in which agents learn online from live interaction using explicit per-user memory. PAHF operationalizes a three-step loop: (1) seeking pre-action clarification to resolve ambiguity, (2) grounding actions in preferences retrieved from memory, and (3) integrating post-action feedback to update memory when preferences drift. To evaluate this capability, we develop a four-phase protocol and two benchmarks in embodied manipulation and online shopping. These benchmarks quantify an agent's ability to learn initial preferences from scratch and subsequently adapt to persona shifts. Our theoretical analysis and empirical results show that integrating explicit memory with dual feedback channels is critical: PAHF learns substantially faster and consistently outperforms both no-memory and single-channel baselines, reducing initial personalization error and enabling rapid adaptation to preference shifts.
Read the full abstract

All authors Kaiqu Liang, Julia Kruk, Shengyi Qian, Xianjun Yang, Shengjie Bi, Yuanshun Yao, Shaoliang Nie, Mingyang Zhang, Lijuan Liu, Jaime Fernández Fisac, Shuyan Zhou, Saghar Hosseini

Computation and Language · Cross-listed from cs.LG

Discrete Stochastic Localization for Non-autoregressive Generation

Yunshu Wu, Jiayi Cheng, Longxuan Yu et al. (+4 more)

Continuous diffusion is a natural framework for non-autoregressive generation but has generally lagged behind masked discrete diffusion models (MDMs) on discrete sequence generation. We argue that the bottleneck is not continuity itself, but a representation in which denoising depends on timestep-indexed noise regimes. We introduce Discrete Stochastic Localization (DSL), a continuous-state framework with unit-sphere token embeddings whose Bayes-optimal denoiser is invariant to the nominal signal-to-noise ratio (SNR) under the localization channel. One trained network then supports an entire family of per-token SNR paths, with endpoint masked-diffusion paths as a special case. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves distributional faithfulness (MAUVE) on OpenWebText across all step budgets from $T{=}128$ to $T{=}1024$, and the same checkpoint supports random-order autoregressive sampling, as well as a hybrid continuous-then-discrete sampler using as few as T=48 total steps – without distillation or retraining.
Read the full abstract

All authors Yunshu Wu, Jiayi Cheng, Longxuan Yu, Partha Thakuria, Rob Brekelmans, Evangelos E. Papalexakis, Greg Ver Steeg

Computation and Language

LLMs Exhibit Significantly Lower Uncertainty in Creative Writing Than Professional Writers

Peiqi Sui

We argue that uncertainty is a key and understudied limitation of LLMs' performance in creative writing, which is often characterized as trite and cliché-ridden. Literary theory identifies uncertainty as a necessary condition for creative expression, while current alignment strategies steer models away from uncertain outputs to ensure factuality and reduce hallucination. We formalize this tension by quantifying the “uncertainty gap” between human-authored stories and model-generated continuations. Through a controlled information-theoretic analysis of 28 LLMs on high-quality storytelling datasets, we demonstrate that human writing consistently exhibits significantly higher uncertainty than model outputs. We find that instruction-tuned and reasoning models exacerbate this trend compared to their base counterparts; furthermore, the gap is more pronounced in creative writing than in functional domains, and shows a consistent correlation with writing quality. Achieving human-level creativity requires new uncertainty-aware alignment paradigms that can distinguish between destructive hallucinations and the constructive ambiguity required for literary richness.
Read the full abstract

Computation and Language · Cross-listed from cs.MM

Emotion Collider: Dual Hyperbolic Mirror Manifolds for Sentiment Recovery via Anti Emotion Reflection

Rong Fu, Ziming Wang, Shuo Yin et al. (+3 more)

Emotional expression underpins natural communication and effective human-computer interaction. We present Emotion Collider (EC-Net), a hyperbolic hypergraph framework for multimodal emotion and sentiment modeling. EC-Net represents modality hierarchies using Poincare-ball embeddings and performs fusion through a hypergraph mechanism that passes messages bidirectionally between nodes and hyperedges. To sharpen class separation, contrastive learning is formulated in hyperbolic space with decoupled radial and angular objectives. High-order semantic relations across time steps and modalities are preserved via adaptive hyperedge construction. Empirical results on standard multimodal emotion benchmarks show that EC-Net produces robust, semantically coherent representations and consistently improves accuracy, particularly when modalities are partially available or contaminated by noise. These findings indicate that explicit hierarchical geometry combined with hypergraph fusion is effective for resilient multimodal affect understanding.
Read the full abstract

All authors Rong Fu, Ziming Wang, Shuo Yin, Kun Liu, Xianda Li, Simon Fong

Computation and Language

Balancing Faithfulness and Performance in Reasoning via Multi-Listener Soft Execution

Nithin Sivakumaran, Shoubin Yu, Hyunji Lee et al. (+4 more)

Chain-of-thought (CoT) reasoning sometimes fails to faithfully reflect the true computation of a large language model (LLM), hampering its utility in explaining how LLMs arrive at their answers. Moreover, optimizing for faithfulness and interpretability in reasoning often degrades task performance. To address this tradeoff and improve CoT faithfulness, we propose Reasoning Execution by Multiple Listeners (REMUL), a multi-party reinforcement learning approach. REMUL builds on the hypothesis that reasoning traces which other parties can follow will be more faithful. A speaker model generates a reasoning trace, which is truncated and passed to a pool of listener models who "execute" the trace, continuing the trace to an answer. Speakers are rewarded for producing reasoning that is clear to listeners, with additional correctness regularization via masked supervised finetuning to counter the tradeoff between faithfulness and performance. On multiple reasoning benchmarks (BIG-Bench Extra Hard, MuSR, ZebraLogicBench, and FOLIO), REMUL consistently and substantially improves three measures of faithfulness – hint attribution, early answering area over the curve (AOC), and mistake injection AOC – while also improving accuracy. Our analysis finds that these gains are robust across training domains, translate to legibility gains, and are associated with shorter and more direct CoTs.
Read the full abstract

All authors Nithin Sivakumaran, Shoubin Yu, Hyunji Lee, Yue Zhang, Ali Payani, Mohit Bansal, Elias Stengel-Eskin

Computation and Language

Missing-by-Design: Certifiable Modality Deletion for Revocable Multimodal Sentiment Analysis

Rong Fu, Ziming Wang, Chunlei Meng et al. (+4 more)

As multimodal systems increasingly process sensitive personal data, the ability to selectively revoke specific data modalities has become a critical requirement for privacy compliance and user autonomy. We present Missing-by-Design (MBD), a unified framework for revocable multimodal sentiment analysis that combines structured representation learning with a certifiable parameter-modification pipeline. Revocability is critical in privacy-sensitive applications where users or regulators may request removal of modality-specific information. MBD learns property-aware embeddings and employs generator-based reconstruction to recover missing channels while preserving task-relevant signals. For deletion requests, the framework applies saliency-driven candidate selection and a calibrated Gaussian update to produce a machine-verifiable Modality Deletion Certificate. Experiments on benchmark datasets show that MBD achieves strong predictive performance under incomplete inputs and delivers a practical privacy-utility trade-off, positioning surgical unlearning as an efficient alternative to full retraining.
Read the full abstract

All authors Rong Fu, Ziming Wang, Chunlei Meng, Jiekai Wu, Kangan Qian, Hao Zhang, Simon Fong

Tuesday 17 February

Computation and Language · Cross-listed from cs.CV

KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR

Henry Gagnier, Sophie Gagnier and Ashwin Kirubakaran

Kazakh is a Turkic language using the Arabic, Cyrillic, and Latin scripts, making it unique in terms of optical character recognition (OCR). Work on OCR for low-resource Kazakh scripts is very scarce, and no OCR benchmarks or images exist for the Arabic and Latin scripts. We construct a synthetic OCR dataset of 7,219 images for all three scripts with font, color, and noise variations to imitate real OCR tasks. We evaluated three multimodal large language models (MLLMs) on a subset of the benchmark for OCR and language identification: Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, and Llama-3.2-11B-Vision-Instruct. All models are unsuccessful with Latin and Arabic script OCR, and fail to recognize the Arabic script as Kazakh text, misclassifying it as Arabic, Farsi, and Kurdish. We further compare MLLMs with a classical OCR baseline and find that while traditional OCR has lower character error rates, MLLMs fail to match this performance. These findings show significant gaps in current MLLM capabilities to process low-resource Abjad-based scripts and demonstrate the need for inclusive models and benchmarks supporting low-resource scripts and languages.
Read the full abstract

Computation and Language

Gemma Needs Help: Investigating and Mitigating Emotional Instability in LLMs

Anna Soligo, Vladimir Mikulik and William Saunders

Large language models can generate responses that resemble emotional distress, and this raises concerns around model reliability and safety. We introduce a set of evaluations to investigate expressions of distress in LLMs, and find that these surface emotional instability in Gemma and Gemini models, but not in other families. We find evidence that this difference arises in post-training. Base models from different families (Gemma, Qwen and OLMo) show similar propensities for expressing distress. However, instruct-tuned Gemma expresses substantially more distress than its base model, whereas instruct-tuned Qwen and OLMo express less. We find a simple mitigation for this: direct preference optimisation on just 280 preference pairs reduces Gemma's high-frustration responses from 35% to 0.3% in our evaluations, generalising across question types, user tones, and conversation lengths, without affecting capabilities. These findings show that emotional instability is an issue in some LLMs. We present (1) evaluations to track this behaviour, and (2) a mitigation without downsides in Gemma, with the caveat that upstream training modifications to improve emotional robustness would be significantly better than this post-hoc fix.
Read the full abstract

Computation and Language

FERRET: Framework for Expansion Reliant Red Teaming

Ninareh Mehrabi, Vitor Albiero, Maya Pavlova et al. (+1 more)

We introduce a multi-faceted automated red teaming framework in which the goal is to generate multi-modal adversarial conversations that would break a target model and introduce various expansions that would result in more effective and efficient adversarial conversations. The introduced expansions include: 1. Horizontal expansion in which the goal is for the red team model to self-improve and generate more effective conversation starters that would shape a conversation. 2. Vertical expansion in which the goal is to take these conversation starters that are discovered in the horizontal expansion phase and expand them into effective multi-modal conversations and 3. Meta expansion in which the goal is for the red team model to discover more effective multi-modal attack strategies during the course of a conversation. We call our framework FERRET (Framework for Expansion Reliant Red Teaming) and compare it with various existing automated red teaming approaches. In our experiments, we demonstrate the effectiveness of FERRET in generating effective multi-modal adversarial conversations and its superior performance against existing state of the art approaches.
Read the full abstract

All authors Ninareh Mehrabi, Vitor Albiero, Maya Pavlova, Joanna Bitton

Computation and Language · Cross-listed from cs.LG

Personalized Group Relative Policy Optimization for Heterogenous Preference Alignment

Jialu Wang, Heinrich Peters, Asad A. Butt et al. (+6 more)

Despite their sophisticated general-purpose capabilities, Large Language Models (LLMs) often fail to align with diverse individual preferences because standard post-training methods, like Reinforcement Learning with Human Feedback (RLHF), optimize for a single, global objective. While Group Relative Policy Optimization (GRPO) is a widely adopted on-policy reinforcement learning framework, its group-based normalization implicitly assumes that all samples are exchangeable, inheriting this limitation in personalized settings. This assumption conflates distinct user reward distributions and systematically biases learning toward dominant preferences while suppressing minority signals. To address this, we introduce Personalized GRPO (P-GRPO), a novel alignment framework that decouples advantage estimation from immediate batch statistics. By normalizing advantages against preference-group-specific reward histories rather than the concurrent generation group, P-GRPO preserves the contrastive signal necessary for learning distinct preferences. We evaluate P-GRPO across diverse tasks and find that it consistently achieves faster convergence and higher rewards than standard GRPO, thereby enhancing its ability to recover and align with heterogeneous preference signals. Our results demonstrate that accounting for reward heterogeneity at the optimization level is essential for building models that faithfully align with diverse human preferences without sacrificing general capabilities.
Read the full abstract

All authors Jialu Wang, Heinrich Peters, Asad A. Butt, Navid Hashemi, Alireza Hashemi, Pouya M. Ghari, Joseph Hoover, James Rae, Morteza Dehghani

Computation and Language

GATech at AbjadMed: Bidirectional Encoders vs. Causal Decoders: Insights from 82-Class Arabic Medical Classification

Ahmed Khaled Khamis

This paper presents system description for Arabic medical text classification across 82 distinct categories. Our primary architecture utilizes a fine-tuned AraBERTv2 encoder enhanced with a hybrid pooling strategies, combining attention and mean representations, and multi-sample dropout for robust regularization. We systematically benchmark this approach against a suite of multilingual and Arabic-specific encoders, as well as several large-scale causal decoders, including zero-shot re-ranking via Llama 3.3 70B and feature extraction from Qwen 3B hidden states. Our findings demonstrate that specialized bidirectional encoders significantly outperform causal decoders in capturing the precise semantic boundaries required for fine-grained medical text classification. We show that causal decoders, optimized for next-token prediction, produce sequence-biased embeddings that are less effective for categorization compared to the global context captured by bidirectional attention. Despite significant class imbalance and label noise identified within the training data, our results highlight the superior semantic compression of fine-tuned encoders for specialized Arabic NLP tasks. Final performance metrics on the test set, including Accuracy and Macro-F1, are reported and discussed.
Read the full abstract

Computation and Language

GATech at AbjadGenEval Shared Task: Multilingual Embeddings for Arabic Machine-Generated Text Classification

Ahmed Khaled Khamis

We present our approach to the AbjadGenEval shared task on detecting AI-generated Arabic text. We fine-tuned the multilingual E5-large encoder for binary classification, and we explored several pooling strategies to pool token representations, including weighted layer pooling, multi-head attention pooling, and gated fusion. Interestingly, none of these outperformed simple mean pooling, which achieved an F1 of 0.75 on the test set. We believe this is because complex pooling methods introduce additional parameters that need more data to train properly, whereas mean pooling offers a stable baseline that generalizes well even with limited examples. We also observe a clear pattern in the data: human-written texts tend to be significantly longer than machine-generated ones.
Read the full abstract

Computation and Language

Adaptive Engram Memory System for Indonesian Language Model: Generative AI Based on TOBA LM for Batak and Minang Language

Hokky Situngkir, Kevin Siringoringo and Andhika Bernard Lumbantobing

This study presents TOBA-LM, a trilingual language model based on GPT-2 architecture with 1.2 billion parameters, trained on a corpus encompassing Indonesian, Batak, and Minangkabau using syllabic-agglutinative tokenization. The architecture integrates an Engram Memory mechanism, an adaptive n-gram-based memory system with a 500,000 x 768 embedding table that captures morphological dependencies through bigram and trigram pathways. Empirical results demonstrate a training efficiency of 80%, with the loss value dropping from 6.4 to 1.7996 in only 12,973 steps – significantly faster than the conventional transformer architecture, which required over 70,000 steps to achieve comparable convergence. These findings confirm that the integration of external statistical memory substantially reduces computational requirements for developing regional language models under limited resources.
Read the full abstract

Computation and Language

SENS-ASR: Semantic Embedding injection in Neural-transducer for Streaming Automatic Speech Recognition

Youness Dkhissi, Valentin Vielzeuf, Elys Allesiardo et al. (+1 more)

Many Automatic Speech Recognition (ASR) applications require streaming processing of the audio data. In streaming mode, ASR systems need to start transcribing the input stream before it is complete, i.e., the systems have to process a stream of inputs with a limited (or no) future context. Compared to offline mode, this reduction of the future context degrades the performance of Streaming-ASR systems, especially while working with low-latency constraint. In this work, we present SENS-ASR, an approach to enhance the transcription quality of Streaming-ASR by reinforcing the acoustic information with semantic information. This semantic information is extracted from the available past frame-embeddings by a context module. This module is trained using knowledge distillation from a sentence embedding Language Model fine-tuned on the training dataset transcriptions. Experiments on standard datasets show that SENS-ASR significantly improves the Word Error Rate on small-chunk streaming scenarios.
Read the full abstract

All authors Youness Dkhissi, Valentin Vielzeuf, Elys Allesiardo, Anthony Larcher

Computation and Language

What Is Missing: Interpretable Ratings for Large Language Model Outputs

Nicholas Stranges and Yimin Yang

Current Large Language Model (LLM) preference learning methods such as Proximal Policy Optimization and Direct Preference Optimization learn from direct rankings or numerical ratings of model outputs, these rankings are subjective, and a single numerical rating chosen directly by a judge is a poor proxy for the quality of natural language, we introduce the What Is Missing (WIM) rating system to produce rankings from natural-language feedback, WIM integrates into existing training pipelines, can be combined with other rating techniques, and can be used as input to any preference learning method without changing the learning algorithm, to compute a WIM rating, a human or LLM judge writes feedback describing what the model output is missing, we embed the output and the feedback with a sentence embedding model and compute the cosine similarity between the resulting vectors, we empirically observe that, compared to discrete numerical ratings, WIM yields fewer ties and larger rating deltas, which improves the availability of a learning signal in pairwise preference data, we use interpretable in the following limited sense: for each scalar rating, we can inspect the judge's missing-information text that produced it, enabling qualitative debugging of the preference labels.
Read the full abstract

Computation and Language · Cross-listed from cs.CR

Intent Laundering: AI Safety Datasets Are Not What They Seem

Shahriar Golchin and Marc Wetter

We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three defining properties: being driven by ulterior intent, well-crafted, and out-of-distribution. We find that these datasets overrely on "triggering cues": words or phrases with overt negative/sensitive connotations that are intended to trigger safety mechanisms explicitly, which is unrealistic compared to real-world attacks. In practice, we evaluate whether these datasets genuinely measure safety risks or merely provoke refusals through triggering cues. To explore this, we introduce "intent laundering": a procedure that abstracts away triggering cues from adversarial attacks (data points) while strictly preserving their malicious intent and all relevant details. Our results show that current adversarial safety datasets fail to faithfully represent real-world adversarial behavior due to their overreliance on triggering cues. Once these cues are removed, all previously evaluated "reasonably safe" models become unsafe, including Gemini 3 Pro and Claude Sonnet 3.7/4. Moreover, when intent laundering is adapted as a jailbreaking technique, it consistently achieves high attack success rates, ranging from 90.00% to 100.00%, under fully black-box access. Overall, our findings expose a significant disconnect between how existing datasets evaluate model safety and how real-world adversaries behave.
Read the full abstract

Computation and Language

Updating Parametric Knowledge with Context Distillation Retains Post-Training Capabilities

Shankar Padmanabhan, Mustafa Omer Gul and Tanya Goyal

Post-training endows pretrained LLMs with a variety of desirable skills, including instruction-following, reasoning, and others. However, these post-trained LLMs only encode knowledge up to a cut-off date, necessitating continual adaptation. Unfortunately, existing solutions cannot simultaneously learn new knowledge from an adaptation document corpora and mitigate the forgetting of earlier learned capabilities. To address this, we introduce Distillation via Split Contexts (DiSC), a simple context-distillation based approach for continual knowledge adaptation. \methodname~derives student and teacher distributions by conditioning on distinct segments of the training example and minimizes the KL divergence between the shared tokens. This allows us to efficiently apply context-distillation without requiring explicit generation steps during training. We run experiments on four post-trained models and two adaptation domains. Compared to prior finetuning and distillation methods for continual adaptation, DiSC consistently reports the best trade-off between learning new knowledge and mitigating forgetting of previously learned skills like instruction-following, reasoning, and factual knowledge.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Why Any-Order Autoregressive Models Need Two-Stream Attention: A Structural-Semantic Tradeoff

Patrick Pynadath and Ruqi Zhang

Any-order autoregressive models (AO-ARMs) offer a promising path toward efficient masked diffusion by enabling native key-value caching, but competitive performance has so far required two-stream attention, typically motivated as a means of decoupling token content from position. In this work, we argue that two-stream attention may be serving a more subtle role. We identify a structural-semantic tradeoff in any-order generation: the hidden representation at each step must simultaneously attend to semantically informative tokens for prediction and structurally recent tokens for summarization, objectives that compete for attention capacity in a single stream but can specialize across two streams. To isolate this tradeoff from position-content separation, we propose Decoupled RoPE, a modification to rotary position embeddings that provides target position information without revealing target content. Decoupled RoPE performs competitively at short sequence lengths–where semantic and structural proximity coincide–but degrades as sequence length increases and the two orderings diverge. These results suggest that the success of two-stream attention stems not merely from separating position from content, but from circumventing the deeper structural-semantic tradeoff inherent to any-order generation.
Read the full abstract

Computation and Language

Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs

Sean Trott, Samuel Taylor, Cameron Jones et al. (+2 more)

Research on mental state reasoning in language models (LMs) has the potential to inform theories of human social cognition–such as the theory that mental state reasoning emerges in part from language exposure–and our understanding of LMs themselves. Yet much published work on LMs relies on a relatively small sample of closed-source LMs, limiting our ability to rigorously test psychological theories and evaluate LM capacities. Here, we replicate and extend published work on the false belief task by assessing LM mental state reasoning behavior across 41 open-weight models (from distinct model families). We find sensitivity to implied knowledge states in 34% of the LMs tested; however, consistent with prior work, none fully “explain away” the effect in humans. Larger LMs show increased sensitivity and also exhibit higher psychometric predictive power. Finally, we use LM behavior to generate and test a novel hypothesis about human cognition: both humans and LMs show a bias towards attributing false beliefs when knowledge states are cued using a non-factive verb (“John thinks...”) than when cued indirectly (“John looks in the...”). Unlike the primary effect of knowledge states, where human sensitivity exceeds that of LMs, the magnitude of the human knowledge cue effect falls squarely within the distribution of LM effect sizes-suggesting that distributional statistics of language can in principle account for the latter but not the former in humans. These results demonstrate the value of using larger samples of open-weight LMs to test theories of human cognition and evaluate LM capacities.
Read the full abstract

All authors Sean Trott, Samuel Taylor, Cameron Jones, James A. Michaelov, Pamela D. Rivière

Computation and Language

Activation Steering via Generative Causal Mediation

Aruna Sankaranarayanan, Amir Zur, Atticus Geiger et al. (+1 more)

Where should we intervene in a language model (LM) to localize and control behaviors that are diffused across many tokens of a long-form response? We introduce Generative Causal Mediation (GCM), a procedure for selecting model components (e.g., attention heads) from contrastive long-form responses, to steer such diffuse concepts (e.g., talk in verse vs. talk in prose). In GCM, we first construct a dataset of contrasting behavioral inputs and long-form responses. Then, we quantify how model components mediate the concept and select the strongest mediators for steering. We evaluate GCM on three behaviors–refusal, sycophancy, and style transfer–across three language models. GCM successfully localizes concepts expressed in long-form responses and outperforms correlational probe-based baselines when steering with a sparse set of attention heads. Together, these results demonstrate that GCM provides an effective approach for localizing from and controlling the long-form responses of LMs.
Read the full abstract

All authors Aruna Sankaranarayanan, Amir Zur, Atticus Geiger, Dylan Hadfield-Menell

Computation and Language

CLAA: Cross-Layer Attention Aggregation for Accelerating LLM Prefill

Bradley McDanel, Steven Li and Harshit Khaitan

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39% compared to the Full KV Cache baseline.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Multi-Objective Alignment of Language Models for Personalized Psychotherapy

Mehrab Beikzadeh, Yasaman Asadollah Salmanpour, Ashima Suvarna et al. (+4 more)

Mental health disorders affect over 1 billion people worldwide, yet access to care remains limited by workforce shortages and cost constraints. While AI systems show therapeutic promise, current alignment approaches optimize objectives independently, failing to balance patient preferences with clinical safety. We survey 335 individuals with lived mental health experience to collect preference rankings across therapeutic dimensions, then develop a multi-objective alignment framework using direct preference optimization. We train reward models for six criteria – empathy, safety, active listening, self-motivated change, trust/rapport, and patient autonomy – and systematically compare multi-objective approaches against single-objective optimization, supervised fine-tuning, and parameter merging. Multi-objective DPO (MODPO) achieves superior balance (77.6% empathy, 62.6% safety) compared to single-objective optimization (93.6% empathy, 47.8% safety), and therapeutic criteria outperform general communication principles by 17.2%. Blinded clinician evaluation confirms MODPO is consistently preferred, with LLM-evaluator agreement comparable to inter-clinician reliability.
Read the full abstract

All authors Mehrab Beikzadeh, Yasaman Asadollah Salmanpour, Ashima Suvarna, Sriram Sankararaman, Matteo Malgaroli, Majid Sarrafzadeh, Saadia Gabriel

Computation and Language · Cross-listed from cs.AI

Evidence-Grounded Subspecialty Reasoning: Evaluating a Curated Clinical Intelligence Layer on the 2025 Endocrinology Board-Style Examination

Amir Hosseinian, MohammadReza Zare Shahneh, Umer Mansoor et al. (+3 more)

Background: Large language models have demonstrated strong performance on general medical examinations, but subspecialty clinical reasoning remains challenging due to rapidly evolving guidelines and nuanced evidence hierarchies. Methods: We evaluated January Mirror, an evidence-grounded clinical reasoning system, against frontier LLMs (GPT-5, GPT-5.2, Gemini-3-Pro) on a 120-question endocrinology board-style examination. Mirror integrates a curated endocrinology and cardiometabolic evidence corpus with a structured reasoning architecture to generate evidence-linked outputs. Mirror operated under a closed-evidence constraint without external retrieval. Comparator LLMs had real-time web access to guidelines and primary literature. Results: Mirror achieved 87.5% accuracy (105/120; 95% CI: 80.4-92.3%), exceeding a human reference of 62.3% and frontier LLMs including GPT-5.2 (74.6%), GPT-5 (74.0%), and Gemini-3-Pro (69.8%). On the 30 most difficult questions (human accuracy less than 50%), Mirror achieved 76.7% accuracy. Top-2 accuracy was 92.5% for Mirror versus 85.25% for GPT-5.2. Conclusions: Mirror provided evidence traceability: 74.2% of outputs cited at least one guideline-tier source, with 100% citation accuracy on manual verification. Curated evidence with explicit provenance can outperform unconstrained web retrieval for subspecialty clinical reasoning and supports auditability for clinical deployment.
Read the full abstract

All authors Amir Hosseinian, MohammadReza Zare Shahneh, Umer Mansoor, Gilbert Szeto, Kirill Karlin, Nima Aghaeepour

Computation and Language

A Curious Class of Adpositional Multiword Expressions in Korean

Junghyun Min, Na-Rae Han, Jena D. Hwang et al. (+1 more)

Multiword expressions (MWEs) have been widely studied in cross-lingual annotation frameworks such as PARSEME. However, Korean MWEs remain underrepresented in these efforts. In particular, Korean multiword adpositions lack systematic analysis, annotated resources, and integration into existing multilingual frameworks. In this paper, we study a class of Korean functional multiword expressions: postpositional verb-based constructions (PVCs). Using data from Korean Wikipedia, we survey and analyze several PVC expressions and contrast them with non-MWEs and light verb constructions (LVCs) with similar structure. Building on this analysis, we propose annotation guidelines designed to support future work in Korean multiword adpositions and facilitate alignment with cross-lingual frameworks.
Read the full abstract

All authors Junghyun Min, Na-Rae Han, Jena D. Hwang, Nathan Schneider

Computation and Language · Cross-listed from cs.SD

MAEB: Massive Audio Embedding Benchmark

Adnan El Assadi, Isaac Chung, Chenghao Xiao et al. (+15 more)

We introduce the Massive Audio Embedding Benchmark (MAEB), a large-scale benchmark covering 30 tasks across speech, music, environmental sounds, and cross-modal audio-text reasoning in 100+ languages. We evaluate 50+ models and find that no single model dominates across all tasks: contrastive audio-text models excel at environmental sound classification (e.g., ESC50) but score near random on multilingual speech tasks (e.g., SIB-FLEURS), while speech-pretrained models show the opposite pattern. Clustering remains challenging for all models, with even the best-performing model achieving only modest results. We observe that models excelling on acoustic understanding often perform poorly on linguistic tasks, and vice versa. We also show that the performance of audio encoders on MAEB correlates highly with their performance when used in audio large language models. MAEB is derived from MAEB+, a collection of 98 tasks. MAEB is designed to maintain task diversity while reducing evaluation cost, and it integrates into the MTEB ecosystem for unified evaluation across text, image, and audio modalities. We release MAEB and all 98 tasks along with code and a leaderboard at https://github.com/embeddings-benchmark/mteb.
Read the full abstract

All authors Adnan El Assadi, Isaac Chung, Chenghao Xiao, Roman Solomatin, Animesh Jha, Rahul Chand, Silky Singh, Kaitlyn Wang, Ali Sartaz Khan, Marc Moussa Nasser, Sufen Fong, Pengfei He, Alan Xiao, Ayush Sunil Munot, Aditya Shrivastava, Artem Gazizov, Niklas Muennighoff, Kenneth Enevoldsen

Computation and Language · Cross-listed from cs.LG

The Geometric Anatomy of Capability Acquisition in Transformers

Jayadev Billa

Neural networks gain capabilities during training, but the internal changes that precede capability acquisition are not well understood. In particular, the relationship between geometric change and behavioral change, and the effect of task difficulty and model scale on that relationship, is unclear. We track geometric measures and linear probes across six transformer sizes (405K–151M parameters), eight algorithmic tasks (144 task$\times$level$\times$model combinations), and three Pythia language models (160M–2.8B). Across all settings, representations first collapse to a low-dimensional state, then recover, and only then does behavioral performance improve. Linear probes show that the model's hidden states already contain task-relevant information before the model can act on it. The collapse floor is task-specific, the collapse propagates top-down through the network, and of the geometric measures tested, only \rankme reliably precedes capability acquisition for hard tasks. Whether this precursor is detectable depends on task difficulty relative to model capacity. For hard tasks, there is a clear gap: geometry changes first, behavior follows. For easy tasks, the model learns so quickly that both happen simultaneously and no precursor is detectable. On Pythia-2.8B, a logical deduction task that is genuinely hard for the model shows a precursor gap of ${\sim}$49K training steps, while easy benchmarks show none. This suggests that geometric patterns observed in small proxy models can persist at larger scale when the task remains difficult relative to model capacity.
Read the full abstract