Computation and Language · Cross-listed from cs.IR
DocAnnot – Accelerating the Creation of Key Information Extraction Datasets with GenAI-Powered Auto-annotation
Siddartha Reddy, Harikrishnan P M, Goutham Vignesh et al. (+2 more)
Key Information Extraction (KIE) is vital for many document applications, but creating training
datasets is traditionally a time-consuming manual process. We introduce DocAnnot, a framework that significantly accelerates KIE
dataset generation. DocAnnot leverages a Large
Vision Language Model (LVLM) for label value extraction, OCR for text/bounding box detection, and a novel Spatially Informed Contextual Matching (SICM)
algorithm. SICM improves label-value association by combining spatial relationships and proximity analysis with textual matching. We evaluate our framework on the CORD and SROIE
benchmarks, demonstrating its ability to auto-generate annotations with F1-scores of 0.679 and 0.846, respectively. Furthermore, we investigate the effectiveness of using auto-annotated data for fine-tuning downstream KIE models. While human-annotated data remains superior, models trained exclusively on DocAnnot's outputs attain respectable performance (e.g., LayoutLMv3 achieving an F1-score of 0.6765 on CORD). These results show that while our framework significantly reduces reliance on manual effort, it does not yet fully eliminate the need for human intervention. However, by automating the process to a point where reviewers can efficiently refine outputs, our system enables near-perfect annotations with much greater efficiency than manual annotation from scratch. This approach offers substantial time and cost savings, making it valuable for resource-constrained settings and rapid model prototyping.
Read the full abstract ▾
All authors Siddartha Reddy, Harikrishnan P M, Goutham Vignesh, Varun V, Vishal Vaddina
Computation and Language
Is MoE Routing a Huffman Code? Discovering the Frequency-Diversity Law in Chain-of-Thought
Ching-Chieh Tsao, Zhuoyi Lin and Wenya Wang
Mixture-of-Experts architectures have revolutionized scaling, yet the underlying logic of their routing remains a black box. In this paper, we uncover a fundamental governing principle: MoE routing is not merely selection, but a manifestation of Huffman Coding. We introduce the Frequency-Diversity Law, revealing that state-of-the-art models, such as Phi-3.5-MoE and Gemma-4-27B-A4B, spontaneously act as information-theoretic engines. These models allocate sparse expert resources for common
tokens while invoking high-diversity expert committees for rare, complex tasks found in chain-of-thought trajectories. However, we identify a critical redundancy trap in Qwen3.5-35B-A3B: when effective sparsity (k/E_eff) is sufficiently low, load-balancing inadvertently imposes functional redundancy, masking the underlying Huffman efficiency signal. To bridge this gap, we propose Subset Difference
Pruning, a
surgical strategy to eliminate functional duplicates. We demonstrate that
pruning does not degrade
reasoning; instead, it unleashes the model's latent Huffman efficiency, forcing the logic to collapse into streamlined, high-density paths. Our findings suggest that the next generation of MoEs should move beyond forced load-balancing toward Minimum Description Length (MDL) optimality, assigning shorter expert-routing codes to high-frequency information and longer, more diverse codes to low-frequency information, thereby transforming routing from a heuristic into a principled compression engine.
Read the full abstract ▾
Computation and Language
Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs'Hallucinations
Xinyue Fang, Zhiliang Tian, Zhen Huang et al. (+5 more)
Existing LLM
hallucination mitigation methods, including
prompt engineering and model optimization, either hardly alter models'internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates
hallucinations by using layer-wise differences in LLMs. However, prior studies only explore
transformer-based models (e.g., GPT), ignoring other effective frameworks like mixture-of-experts (MoE) models. Since MoE alters the traditional
transformer architecture, we conduct empirical studies to investigate whether similar layer-wise differences exist in MoEs. Our results show that they do not exist in MoE with shared experts; nevertheless, across different MoEs, higher layers exhibit distinct expert activation patterns between factual and non-factual outputs. Building on these, we propose EAACD, an expert-aware adaptive contrast decoding that uses expert differences in MoE's higher layers to mitigate
hallucinations on QA tasks. EAACD splits high-layer experts into a higher-reliability group and several lower-reliability groups based on their confidence and consistency. It contrasts the higher-reliability group's prediction with each lower-reliability group's prediction to calibrate the model's original predictions. To strengthen this contrast, EAACD amplifies
hallucinations from lower-reliability experts via
attention and masking to provide stronger negative references. EAACD outperforms all baselines on four
datasets.
Read the full abstract ▾
All authors Xinyue Fang, Zhiliang Tian, Zhen Huang, Ziyi Pan, Zhihua Wen, Xi Wang, Quntian Fang, Dongsheng Li
Computation and Language
Heterogeneous Element-Aware Cross-Version Differencing of Scientific Documents via Layout-Aware Alignment and Structure-Aware Reasoning
Zhen Yin, Wenkang An, Hao Wang et al. (+1 more)
Cross-version differencing of scientific documents is essential in scholarly publishing and technical documentation, but remains challenging because scientific documents are page-structured artifacts containing heterogeneous elements such as text, tables, formulas, figures, and layout cues. Existing text-sequence-based methods often lose layout and structural information, while image-based methods lack semantic interpretability and are sensitive to rendering variation. To address these limitations, this paper proposes a layout-aware heterogeneous element-aware framework for scientific document differencing. The framework decomposes document versions into semantically typed elements, establishes cross-version correspondence through an
alignment-first mechanism that jointly models spatial, content, and structural compatibility, and performs type-aware difference
reasoning over aligned element pairs. It supports unified change detection, localization, structure-awareness analysis, and
alignment/matching evaluation across text, tables, formulas, and figures. Experiments on real-world scientific PDF data from journal production proofreading workflows show that the proposed framework consistently outperforms element-specific baselines. It achieves detection F1 scores of 0.903, 0.855, 0.862, and 0.845 for text, tables, formulas, and figures, respectively, with further improvements in localization, structure awareness, and matching quality. Ablation and sensitivity analyses confirm the effectiveness of cross-version
alignment, type-specific representations, structure-aware
reasoning, and compatibility-weight design. These results demonstrate that heterogeneous element-aware differencing provides a robust and interpretable solution for scientific document comparison in realistic editorial production scenarios.
Read the full abstract ▾
All authors Zhen Yin, Wenkang An, Hao Wang, Keran You
Computation and Language
ReportMedSAM: Guiding Segmentation Through Radiology Reports
Anghong Du, Theodoros N. Arvanitis, Colin Watts et al. (+2 more)
Free-form radiology reports contain rich clinical descriptions, yet converting them for reliable
segmentation remains challenging due to the inherent variability of natural
language. Existing pipelines often rely on predefined organ phrases or brittle rule-based inference-time extraction, which limits their scalability to novel anatomical structures and makes them sensitive to linguistic variations. To address this, we propose ReportMedSAM, a report-driven framework that replaces discrete extraction with a learnable concept bank. By leveraging a frozen medical
vision-
language encoder (BiomedCLIP), we align organ-level concept
embeddings with large-scale clinical corpora through contrastive learning, establishing mutually orthogonal semantic anchors. Our approach explicitly mitigates organ-level semantic collapse and ensures high robustness against diverse clinical synonyms (e.g., "renal" vs. "kidney" ). During inference, a clinical report is embedded and matched against this concept bank to dynamically activate task-specific Mixture-of-Experts (MoE) modules. This decoupled design allows new concepts and experts to be added without retraining existing components, providing a parameter-isolated extension mechanism while keeping previously learned experts unchanged. Evaluated on the AbdomenAtlas 3.0
dataset, ReportMedSAM effectively interprets free-form reports, achieves competitive
segmentation accuracy, and demonstrates seamless, non-interfering extension to novel clinical tasks.
Read the full abstract ▾
All authors Anghong Du, Theodoros N. Arvanitis, Colin Watts, Alejandro F. Frangi, Le Zhang
Computation and Language · Cross-listed from cs.AI
DialogueVPR: Towards Conversational Visual Place Recognition
Yukun Song, Changwei Wang, Xingtian Pei et al. (+8 more)
Inspired by how humans communicate spatial information,
language-guided geo-localization has gained significant traction for its intuitive and practical value. Despite this progress, most methods still rely on a static, one-shot
retrieval paradigm, which fails to handle the ambiguity and incompleteness inherent in real-world natural
language descriptions. We propose a paradigm shift to
reasoning retrieval and introduce Dialogue Place Recognition (DlgPR), which casts localization as an interactive, dialogue-driven
reasoning process. To support this new task, we present DlgQuest-Cities, the first large-scale dialogue-based
benchmark for place recognition, and a unified
reasoning framework that couples a cross-modal multi-level retriever with an intelligent questioner, DQ-pilot. DQ-pilot is trained in a curriculum: supervised fine-tuning on a curated DQ-cities-20k subset followed by reinforcement refinement on a harder DQ-cities-10k split via GRPO. Two task-aligned metrics guide learning: a Discriminative Difficulty Index (DDI) for curriculum sampling and a Positional
Retrieval Gain (PRG) reward that directly measures
retrieval improvement induced by a question. Experiments show this
reasoning-based approach significantly outperforms baselines. The code and model are available at https://github.com/Graysonggg/DlgPR.
Read the full abstract ▾
All authors Yukun Song, Changwei Wang, Xingtian Pei, Shibiao Xu, Wenhao Xu, Shunpeng Chen, Yu Zhang, Ke Zhang, Rongtao Xu, Xuxiang Feng, Pengyang Wang
Computation and Language
CoEvoT: Co-Evolving Chain-of-Thought Prompting for Graph-LLM Reasoning
Haohua Niu, Xingtong Yu, Yang Liu et al. (+6 more)
Graph learning under distribution shift presents a persistent challenge, where models adapt to new graphs with limited or even no supervision. Recent graph–LLM approaches move toward label-efficient prediction by linearizing graphs into
prompts and using large
language models (LLMs) as predictors, and can adopt Chain-of-Thought (CoT) prompting to exploit LLM's multi-step
reasoning capability. However, existing CoT-based graph–LLM methods generate intermediate thoughts while conditioning on fixed graph
tokens, limiting step-wise refinement of structural cues. In this paper, we propose CoEvoT, a simple yet effective co-evolving CoT prompting framework for graph–LLM
reasoning. CoEvoT couples text-to-graph
token rewriting and graph-to-text
reasoning guidance in a closed loop: each intermediate textual thought is used to update the graph
token evidence state via a lightweight condition network, and the updated
tokens are fed back into the next-step instruction to guide subsequent LLM
reasoning. This enables step-wise, state-aware evidence refinement, rather than
reasoning over a fixed graph snapshot. Extensive experiments on eight
datasets demonstrate that CoEvoT consistently outperforms state-of-the-art baselines.
Read the full abstract ▾
All authors Haohua Niu, Xingtong Yu, Yang Liu, Junfeng Fang, Xuanting Xie, Jie Tan, Zhongjian Zhang, Hong Cheng, Yuan Fang
Computation and Language
T5-CSBoost: Adversarial Perturbation Resistant LLM Fingerprinting
Gayan K. Kulatilleke, Mahsa Baktashmotlagh, Siamak Layeghy et al. (+1 more)
While many AI-generated text (AIGT) detectors achieve strong performance on clean inputs, their accuracy degrades significantly under light paraphrasing, word substitutions, character edits, and distribution shifts. We present T5 Contrastive Style Boosted Classifier (T5-CSBoost), an extension to the T5-Sentinel framework that keeps the original next-
token prediction objective for source attribution while introducing an auxiliary margin-based triplet loss over decoder
embeddings. This contrastive style regularization encourages the learning of compact, perturbation-resistant stylistic representations, offering a lightweight yet effective alternative to prior approaches that rely on architectural modifications,
adversarial training, or complex multi-task objectives without altering the underlying T5-small backbone. T5-CSBoost achieves state-of-the-art multiclass source attribution and binary human-vs-LLM detection on OpenLLMText and HC3 AIGT
benchmarks. More importantly, T5-CSBoost demonstrates enhanced robustness to word and character level
adversarial perturbations of up to 90% intensity, achieving state-of-the-art on the challenging MAGE/Deepfake stress-test suite, including unseen models, unseen domains, and extreme paraphrasing scenarios. Our results highlight that explicitly regularizing stylistic
embeddings via contrastive learning is a practical and effective strategy for building more robust LLM fingerprinting systems in real-world
adversarial settings.
Read the full abstract ▾
All authors Gayan K. Kulatilleke, Mahsa Baktashmotlagh, Siamak Layeghy, Marius Portmann
Computation and Language
Limits of Reliability and Scaling in Language Models
Subhabrata Majumdar
Large
language models (LLMs) are trained and evaluated as though perfect reliability is achievable for any task given sufficient scale. We show that this assumption is information-theoretically unjustified. Every generative task has a reliability ceiling that no model can exceed, determined by how much output uncertainty is resolvable from observable context. The gap decomposes into a resolvable component closable with additional context and a subjective component inherent to task ambiguity. Autoregressive generation further degrades this ceiling at a rate governed by the task's dependency kernel, which quantifies inter-
token correlations in the output. From these two primitives, we derive a first-principles scaling law where LLM performance is bottlenecked by the scarcer resource: training data or model capacity. This law recovers the Chinchilla scaling law as a special case and provides a structural account of when scaling improves reliability. Beyond scaling, our framework unifies diverse practical phenomena, such as the benefits of
retrieval-augmentation and the spectral mechanics of catastrophic forgetting. Our work formalizes the resource-complexity tradeoffs that govern model performance across domains, offering a unified theory of performance limits in generative
language models.
Read the full abstract ▾
Computation and Language
Introspection Fine-Tuning (IFT): Training Small LLMs to Introspect
Ely Hahami, Ishaan Sinha and Lavik Jain
Can small
language models detect and report on perturbations their own internal activations? We investigate this question through the lens of activation steering: injecting concept vectors into a model's residual stream and measuring whether the model can accurately report on the perturbation. We first show that the binary detection paradigm used in prior work – prompting the model to answer Yes” or No” to whether it detects an injected thought – is confounded in small models, as steering biases the model toward affirmative responses regardless of the question content. We therefore propose two confound-free evaluation paradigms: sentence localization (identifying which of $N$ sentences was perturbed, chance $= 1/N$) and strength comparison (identifying which of two sentences received a stronger injection, chance $= 50\%$). Evaluating across six models from two families (Llama-3.2 and Gemma-4), we find that models as small as 2B parameters introspect reliably well above chance, and that introspective ability generally increases with scale. Llama-1B, however, performs at or below chance. We then introduce Introspection Fine-Tuning (IFT): supervised fine-tuning on sentence-localization examples constructed from the model's own perturbed forward passes. IFT raises Llama-1B sentence-localization accuracy from $9.6\%$ to $60.6\%$ (a $6\times$ improvement), with gains generalizing zero-shot to the held-out strength-comparison task ($30.2\% \to 52.2\%$). IFT also improves introspection for 3B and 8B models, while inducing negligible degradation on standard capability
benchmarks. Our results suggest that introspective ability is not fixed by scale alone: it can be directly trained, and doing so unlocks latent self-monitoring capacity with implications for AI transparency and
alignment. Our code is \href{https://anonymous.4open.science/r/IFT-introspection-2092/README.md}{here}.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.IR
Prompt Optimization for User Simulation in Conversational Recommender Systems: A Multi-Objective Framework
Nipun B Nair, Tongtong Wu and Weiqing Wang
Conversational recommender systems (CRSs) are a core component of next-generation intelligent recommender systems because they enable users to actively elicit preferences, clarify intentions, and adapt recommendations in real time. However, there are two key obstacles in the CRS domain: evaluation and access to training data. Evaluating CRSs through real human studies is more critical than for traditional recommender systems, yet such studies are both costly and time-consuming. Moreover, CRS interaction data are often difficult to obtain for model training due to
privacy concerns. Large
language model (LLM)-based user simulators have shown promise in addressing both challenges by generating synthetic user interactions for evaluation and training. However, existing approaches suffer from systematic positive bias, data leakage, and limited behavioral diversity, and they rely on brittle manual
prompt engineering that requires extensive domain expertise. In this paper, we propose a framework to automatically optimize
prompts for LLM-based user simulators in CRSs, simultaneously mitigating these issues. Experimental results demonstrate that the proposed framework achieves improved behavioral
alignment with human interaction patterns compared to baseline methods across diverse
prompt settings.
Read the full abstract ▾
Computation and Language
Position: The Term "Machine Unlearning" Is Overused in LLMs
Sangyeon Yoon, Yeachan Jun and Albert No
Large
language models increasingly face demands to "forget" training data, knowledge, or behaviors due to regulatory deletion obligations, copyright/licensing disputes, and safety or product-policy requirements. This position paper argues that machine unlearning is overused as a term in LLM research and should be reserved for
dataset-defined deletion: removing the training influence of a precisely specified forget set such that the resulting model is approximately indistinguishable from retraining without that data. We contend that many tasks currently labeled "unlearning" (e.g., refusal for harmful requests, entity/knowledge removal, or targeted suppression) pursue different, often policy-dependent objectives and therefore require different terminology and baselines (e.g.,
alignment, suppression, editing, obfuscation). We further argue that this confusion is not cosmetic: because papers make different implicit guarantees under the same label, metrics and
benchmarks are frequently reused outside their intended scope, rewarding surface-level non-disclosure (e.g., low ROUGE/forget accuracy) even when retraining-equivalence is not tested and derived capabilities remain. We conclude by calling for stricter terminology tied to explicit guarantees and reference models, and for evaluations that match the claimed objective.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Augmenting Human Evaluation with LLM Judges: How Many Human Reviews Do You Need?
Jane Paik Kim
Large
language models (LLMs) are increasingly used as automated evaluators of AI systems, including in high-stakes applications. In this role, LLMs are used to generate judgments about the quality, appropriateness, or even safety of model outputs. This approach is motivated by practical constraints. Expert human ratings are costly and difficult to scale, whereas LLM ratings can be produced quickly at low cost. However, current approaches to deploying LLM evaluators are ad hoc, typically limited to reporting agreement metrics between human and LLM judges as a justification for substitution of human ratings, and lack a formal basis for study design. This paper (1) shifts the role of the LLM judge from substitutive to auxiliary, and (2) formulates the LLM-as-a-judge paradigm as one of augmenting human evaluation through a two-stage sampling design, where LLM evaluations are measured for all observations at the first stage and human ratings are partially observed for a subsample at the second stage. We propose to use a doubly robust
estimator from the missing data literature, which takes advantage of the robustness property against the prediction model, since the missingness model is known by design. Using the asymptotic variance of this
estimator, we propose how sample sizes of human and LLM ratings can be determined to achieve a targeted level of power. We also show that a study can be efficiently designed by allocating more human ratings for types of evaluations where the predictability of LLM ratings is not high. To the best of our knowledge, there is very little guidance on how much human oversight should be retained when validating
benchmarks.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models
Amin Karimi Monsefi, Dominic Culver, Nikhil Bhendawade et al. (+4 more)
Diffusion large
language models are a compelling alternative to autoregressive models, yet existing RL methods for
diffusion treat all denoising steps as equally important and rely on biased, high-variance likelihood estimates. We identify two fundamental weaknesses: the absence of temporal credit assignment across the denoising trajectory, and the systematic bias of mean-field likelihood estimates used for policy optimization. To address these, we propose Denoising-Aware Credit Assignment for GRPO (DACA-GRPO), a lightweight, plug-and-play enhancement for any GRPO-style trainer. DACA-GRPO introduces two complementary mechanisms: Denoising Progress Scores, which extract per-
token importance weights from intermediate predictions at no additional forward cost, and Stratified Masking Likelihood, which partitions
token positions into strata so that each
token is predicted with most of the sequence as context, reducing the mean-field bias. Applied on top of three GRPO base methods, DACA-GRPO achieves consistent improvements across seven
benchmarks spanning mathematical
reasoning, code generation, constraint satisfaction, and constrained generation, with gains of up to 5.6pp on math
reasoning, 7.4pp on code generation, 36.3pp on constraint satisfaction, and 5.9pp on JSON schema adherence.
Read the full abstract ▾
All authors Amin Karimi Monsefi, Dominic Culver, Nikhil Bhendawade, Lokesh Boominathan, Manuel R. Ciosici, Yizhe Zhang, Irina Belousova
Computation and Language · Cross-listed from cs.LG
Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Models
Hanhan Zhou, Shamik Roy and Rashmi Gangadharaiah
Discrete
diffusion language models (DLMs) generate text by iteratively denoising all positions in parallel, offering an alternative to autoregressive models. Controlled generation methods for DLMs, imported from autoregressive models, apply uniform intervention at every denoising step. We show this uniform schedule is inefficient and degrades quality, and the damage compounds when multiple attributes are steered jointly. To diagnose the failure, we train sparse autoencoders on four DLMs (124M-8B parameters) and find that different attributes commit on distinct schedules, varying in timing, sharpness, and magnitude. For instance, topic commits within the first 2% of denoising on MDLM, whereas sentiment emerges gradually over 20% of the process. Motivated by these profiles, we propose an adaptive scheduling mechanism that concentrates intervention where each attribute is actively forming. An idealized allocation analysis predicts that attributes with more sharply concentrated emergence benefit more from adaptive scheduling, a prediction we confirm empirically. Across seven single- and multi-attribute steering tasks on four DLMs, adaptive steering consistently improves the control-quality balance over uniform and interval-restricted baselines.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Human-Inspired Memory Architecture for LLM Agents
Doga Kerestecioglu, Alexei Robsky, Clemens Vasters et al. (+2 more)
Current LLM
agents lack principled mechanisms for managing persistent
memory across long interaction horizons. We present a biologically-grounded
memory architecture comprising six cognitive mechanisms: (1) sleep-phase consolidation, (2) interference-based forgetting, (3) engram maturation, (4) reconsolidation upon
retrieval, (5) entity knowledge graphs, and (6) hybrid multi-cue
retrieval. Each mechanism addresses a specific failure mode of naive
memory accumulation. We introduce a synthetic calibration methodology that derives all pipeline thresholds without
benchmark data exposure, eliminating a common source of evaluation leakage. We evaluate on two
benchmarks. First, a VSCode issue-tracking
dataset (13K issues, 120K events) where deduplication-based consolidation achieves 97.2% retention precision with 58% store reduction (+21.8 pp over baseline). Second, the LongMemEval personal-chat
benchmark where we conduct the first streaming M-tier evaluation (475 sessions, ~540K unique turns). At a 200K-
token context budget, our pipeline matches raw
retrieval accuracy (70.1% vs. 71.2%, overlapping 95% CI) while exposing a tunable accuracy/store-size operating curve. At S-tier scale (50 sessions), dedup-based consolidation yields a +13.3 pp improvement in preference recall.
Read the full abstract ▾
All authors Doga Kerestecioglu, Alexei Robsky, Clemens Vasters, Anshul Sharma, Yitzhak Kesselman
Computation and Language
Coordinates of Capability: A Unified MTMM-Geometric Framework for LLM Evaluation
Adib Sakhawat, Tahsin Islam, Takia Farhin et al. (+3 more)
The evaluation of Large
Language Models (LLMs) faces a critical challenge in construct validity, where fragmented
benchmarks and ad hoc metrics frequently conflate method variance, such as
prompt sensitivity, with true latent capabilities. Concurrently, emerging research suggests that LLM capabilities and outputs can be modeled as continuous
geometric manifolds. In this Systematization of Knowledge (SoK), we bridge these paradigms by proposing a generalized Multi-Trait Multi-Method (MTMM) framework for LLM evaluation. We formalize and unify nine evaluation metrics, including Paraphrase Instability, Drift Score, Overton Width, and Pluralism Score, interpreting them not as isolated scalar values but as
geometric measurements within a shared latent coordinate space. This spatial unification factorizes model behavior into three orthogonal latent dimensions: (1) Instability and Sensitivity, (2) Position and
Alignment, and (3) Coverage and Expressiveness. By systematically separating task-irrelevant perturbations from true capability spans, the framework provides a theoretically grounded and domain-agnostic taxonomy for robust and empirically stable
benchmark design.
Read the full abstract ▾
All authors Adib Sakhawat, Tahsin Islam, Takia Farhin, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan
Computation and Language
A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
Hamid Kazemi, Atoosa Chegini and Maria Safi
Safety
alignment in
language models operates through two mechanistically distinct systems: refusal
neurons that gate whether harmful knowledge is expressed, and concept
neurons that encode the harmful knowledge itself. By targeting a single
neuron in each system, we demonstrate both directions of failure – bypassing safety on explicit harmful requests via suppression, and inducing harmful content from innocent
prompts via amplification – across seven models spanning two families and 1.7B to 70B parameters, without any training or
prompt engineering. Our findings suggest that safety
alignment is not robustly distributed across model weights but is mediated by individual
neurons that are each causally sufficient to gate refusal behavior – suppressing any one of the identified refusal
neurons bypasses safety
alignment across diverse harmful requests.
Read the full abstract ▾
Computation and Language
A Single Layer to Explain Them All:Understanding Massive Activations in Large Language Models
Zeru Shi, Zhenting Wang, Fan Yang et al. (+2 more)
We investigate the origins of massive activations in large
language models (LLMs) and identify a specific layer named the Massive Emergence Layer (ME Layer), that is consistently observed across model families, where massive activations first emerge and subsequently propagate to deeper layers through residual connections. We show that, within the ME Layer both the RMSNorm and the FFN parameters jointly contribute to the emergence of massive activations. Once formed, the massive activation
token representation remains largely invariant across layers, reducing the diversity of hidden representations passed to the
attention module. Motivated by this limitation, we propose a simple and effective method to reduce the
rigidity of the massive activation
token. Our approach consistently improves LLM performance across multiple tasks, including instruction following and math
reasoning, in both training free and fine tuning settings. Moreover, we show that our method mitigates
attention sinks by selectively weakening their influence, elucidating their origin at the hidden state level and shedding new light on principled mitigation strategies.
Read the full abstract ▾
All authors Zeru Shi, Zhenting Wang, Fan Yang, Qifan Wang, Ruixiang Tang
Computation and Language
NARRA-Gym for Evaluating Interactive Narrative Agents
Yue Huang, Yuchen Ma, Jiayi Ye et al. (+14 more)
Interactive narrative tasks require LLMs to sustain a coherent, evolving story while adapting to a user over multiple turns. However, suitable
benchmarks for this setting are limited: existing evaluations often focus on static
prompts, isolated story generations, or post-hoc ratings, and therefore miss whether models can jointly manage story generation, long-context state and pacing, character simulation, empathic personalization, and story-grounded artifacts. We introduce NARRA-Gym, an executable evaluation environment that turns a sparse emotional seed into a complete interactive story episode and logs the full model-in-the-loop trajectory, including story construction,
memory updates, planning, pacing interventions, and optional artifact synthesis. We evaluate nine frontier LLMs using a controlled LLM-as-judge sweep over eight
benchmark personas and a human evaluation in which participants rate customized model outputs. Our results show substantial variation across models, personas, and evaluation dimensions: models that produce fluent stories can still fail on robustness, user experience, or resistance-sensitive personalization. These findings suggest that interactive narrative offers a useful
benchmark for evaluating long-horizon, user-adaptive LLM behavior beyond isolated story quality.
Read the full abstract ▾
All authors Yue Huang, Yuchen Ma, Jiayi Ye, Wenjie Wang, Zipeng Ling, Xingjian Hu, Yuexing Hao, Zichen Chen, Zhangchen Xu, Yunhong He, Zhengqing Yuan, Yujun Zhou, Kehan Guo, Chaoran Chen, Toby Jia-Jun Li, Stefan Feuerriegel, Xiangliang Zhang
Computation and Language · Cross-listed from cs.LG
ShifaMind: A Multiplicative Concept Bottleneck for Interpretable ICD-10 Coding
Mohammed Sameer Syed and Xuan Lu
Automated ICD-10 coding from clinical discharge summaries requires models that are both accurate on long-tailed multi-label classification tasks and interpretable to clinicians. Concept Bottleneck Models (CBMs) offer a principled framework for interpretability by routing predictions through human-interpretable concepts, but this transparency often comes at a cost: compressing rich clinical text representations into a narrow concept layer can restrict gradient flow and limit predictive capacity. We present ShifaMind, a concept-grounded architecture built around a Multiplicative Concept Bottleneck (MCB), which changes the form, rather than the width, of the bottleneck. Instead of projecting through a narrow concept layer, ShifaMind uses a learned multiplicative gate over a concept-grounded representation while retaining a scalar concept interface for inspection. On MIMIC-IV top-50 ICD-10 coding, ShifaMind achieves performance competitive with LAAT, the strongest baseline, across F1, AUC, and ranking metrics, while outperforming five additional ICD-coding baselines and providing concept-mediated explanations. Its substantial gains over a capacity-matched Vanilla CBM in both predictive performance and interpretability-oriented metrics highlight the importance of the bottleneck design.
Read the full abstract ▾