Computation and Language · Cross-listed from cs.LG
Online Bayesian Imbalanced Learning with Bregman-Calibrated Deep Networks
Zahir Alsulaimawi
Class imbalance remains a fundamental challenge in machine learning, where standard classifiers exhibit severe performance degradation in minority classes. Although existing approaches address imbalance through resampling or cost-sensitive learning during training, they require retraining or access to labeled target data when class distributions shift at deployment time, a common occurrence in real-world applications such as fraud detection, medical diagnosis, and anomaly detection. We present Online
Bayesian Imbalanced Learning (OBIL), a principled framework that decouples likelihood-ratio estimation from class-prior assumptions, enabling real-time adaptation to distribution shifts without model retraining. Our approach builds on the established connection between Bregman divergences and proper scoring rules to show that deep networks trained with such losses produce posterior probability estimates from which prior-invariant likelihood ratios can be extracted. We prove that these likelihood-ratio estimates remain valid under arbitrary changes in class priors and cost structures, requiring only a threshold adjustment for optimal Bayes decisions. We derive finite-sample regret bounds demonstrating that OBIL achieves $O(\sqrt{T \log T})$ regret against an oracle with perfect prior knowledge. Extensive experiments on
benchmark datasets and medical diagnosis
benchmarks under simulated deployment shifts demonstrate that OBIL maintains robust performance under severe distribution shifts, outperforming state-of-the-art methods in F1 Score when test distributions deviate significantly from the training conditions.
Read the full abstract ▾
Computation and Language
Gender and Race Bias in Consumer Product Recommendations by Large Language Models
Ke Xu, Shera Potka and Alex Thomo
Large
Language Models are increasingly employed in generating consumer product recommendations, yet their potential for
embedding and amplifying gender and race biases remains underexplored. This paper serves as one of the first attempts to examine these biases within LLM-generated recommendations. We leverage
prompt engineering to elicit product suggestions from LLMs for various race and gender groups and employ three analytical methods-Marked Words, Support Vector Machines, and Jensen-Shannon Divergence-to identify and quantify biases. Our findings reveal significant disparities in the recommendations for demographic groups, underscoring the need for more equitable LLM recommendation systems.
Read the full abstract ▾
Computation and Language
Emergent Search and Backtracking in Latent Reasoning Models
Jasmine Cui and Charles Ye
What happens when a
language model thinks without words? Standard
reasoning LLMs verbalize intermediate steps as chain-of-thought; latent
reasoning transformers (LRTs) instead perform deliberation entirely in continuous hidden space. We investigate an LRT, decoding the model's evolving beliefs at every step on a multiple-choice QA
benchmark. We find that the model spontaneously learns a structured search process in latent space. Deliberation follows a consistent trajectory: an exploration phase where probability mass spreads across candidates, tentative commitment to a frontrunner, and either convergence or backtracking. Backtracking is prevalent (32% of instances), beneficial (34% accuracy gain over non-backtracking instances), and predominantly directed away from the semantically closest distractor toward the correct answer. The search is adaptive: replacing distractors with implausible alternatives shortens exploration by 54%. Latent
reasoning models achieve in activation space what chain-of-thought achieves through words: the ability to be wrong, notice, and recover.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
SiameseNorm: Breaking the Barrier to Reconciling Pre/Post-Norm
Tianyu Li, Dongchen Han, Zixuan Cao et al. (+7 more)
The long-standing tension between Pre- and Post-Norm remains an open problem in
Transformer architecture, reflecting a fundamental trade-off between training stability and representational capacity. Prior attempts to combine their strengths have made progress, but often show limited robustness across training settings, restricting their broader applicability. We revisit this dilemma, showing that single-stream architectures struggle to reconcile Pre-Norm's stable identity-gradient propagation with Post-Norm's normalization of the main residual path. To address this structural tension, we propose SiameseNorm, a simple yet effective two-stream architecture that remains compatible with Pre-Norm training recipes. SiameseNorm couples Pre-Norm-like and Post-Norm-like streams through shared residual blocks, allowing each residual block to receive optimization signals from both pathways with negligible overhead. Extensive experiments on 400M and 1.3B dense
language models, 15B MoE models,
Vision Transformers, and
Diffusion Transformers show that SiameseNorm consistently improves performance while maintaining strong training stability across architectures and modalities. Code is available at https://github.com/Qwen-Applications/SiameseNorm.
Read the full abstract ▾
All authors Tianyu Li, Dongchen Han, Zixuan Cao, Haofeng Huang, Mengyu Zhou, Ming Chen, Erchao Zhao, Xiaoxi Jiang, Guanjun Jiang, Gao Huang
Computation and Language
Arshia Hemmat, Philip Torr, Yongqiang Chen et al. (+1 more)
Diffusion language models (D-LLMs) offer parallel denoising and bidirectional context, but
hallucination detection for D-LLMs remains underexplored. Prior detectors developed for auto-regressive LLMs typically rely on single-pass cues and do not directly transfer to
diffusion generation, where factuality evidence is distributed across the denoising trajectory and may appear, drift, or be self-corrected over time. We introduce TDGNet, a temporal dynamic graph framework that formulates
hallucination detection as learning over evolving
token-level
attention graphs. At each denoising step, we sparsify the
attention graph and update per-
token memories via message passing, then apply temporal
attention to aggregate trajectory-wide evidence for final prediction. Experiments on LLaDA-8B and Dream-7B across QA
benchmarks show consistent AUROC improvements over output-based, latent-based, and static-graph baselines, with single-pass inference and modest overhead. These results highlight the importance of temporal
reasoning on
attention graphs for robust
hallucination detection in
diffusion language models.
Read the full abstract ▾
All authors Arshia Hemmat, Philip Torr, Yongqiang Chen, Junchi Yu
Computation and Language · Cross-listed from cs.LG
Implicit Strategic Optimization: Rethinking Long-Horizon Decision-Making in Adversarial Poker Environments
Boyang Xia, Weiyou Tian, Qingnan Ren et al. (+7 more)
Training large
language model (LLM)
agents for
adversarial games is often driven by episodic objectives such as win rate. In long-horizon settings, however, payoffs are shaped by latent strategic externalities that evolve over time, so myopic optimization and variation-based regret analyses can become vacuous even when the dynamics are predictable. To solve this problem, we introduce Implicit Strategic Optimization (ISO), a prediction-aware framework in which each
agent forecasts the current strategic context and uses it to update its policy online. ISO combines a Strategic Reward Model (SRM) that estimates the long-run strategic value of actions with iso-grpo, a context-conditioned optimistic learning rule. We prove sublinear contextual regret and
equilibrium convergence guarantees whose dominant terms scale with the number of context mispredictions; when prediction errors are bounded, our bounds recover the static-game rates obtained when strategic externalities are known. Experiments in 6-player No-Limit Texas Hold'em and competitive Pokemon show consistent improvements in long-term return over strong LLM and RL baselines, and graceful degradation under controlled prediction noise.
Read the full abstract ▾
All authors Boyang Xia, Weiyou Tian, Qingnan Ren, Jiaqi Huang, Jie Xiao, Shuo Lu, Kai Wang, Lynn Ai, Eric Yang, Bill Shi
Computation and Language
Beyond Raw Detection Scores: Markov-Informed Calibration for Boosting Machine-Generated Text Detection
Chenwang Wu, Yiu-ming Cheung, Shuhai Zhang et al. (+2 more)
While machine-generated texts (MGTs) offer great convenience, they also pose risks such as disinformation and phishing, highlighting the need for reliable detection. Metric-based methods, which extract statistically distinguishable features of MGTs, are often more practical than complex model-based methods that are prone to overfitting. Given their diverse designs, we first place representative metric-based methods within a unified framework, enabling a clear assessment of their advantages and limitations. Our analysis identifies a core challenge across these methods: the
token-level detection score is easily biased by the inherent randomness of the MGTs generation process. To address this, we theoretically and empirically reveal two relationships of context detection scores that may aid calibration: Neighbor Similarity and Initial Instability. We then propose a Markov-informed score calibration strategy that models these relationships using Markov random fields, and implements it as a lightweight component via a mean-field approximation, allowing our method to be seamlessly integrated into existing detectors. Extensive experiments in various real-world scenarios, such as cross-LLM and paraphrasing attacks, demonstrate significant gains over baselines with negligible computational overhead. The code is available at https://github.com/tmlr-group/MRF_Calibration.
Read the full abstract ▾
All authors Chenwang Wu, Yiu-ming Cheung, Shuhai Zhang, Bo Han, Defu Lian
Computation and Language · Cross-listed from cs.AI
Free(): Learning to Forget in Malloc-Only Reasoning Models
Yilun Zheng, Dongyang Ma, Tian Liang et al. (+5 more)
Reasoning models enhance problem-solving by scaling test-time compute, yet they face a critical paradox: excessive thinking
tokens often degrade performance rather than improve it. We attribute this to a fundamental architectural flaw: standard LLMs operate as "malloc-only" engines, continuously accumulating valid and redundant steps alike without a mechanism to prune obsolete information. To break this cycle, we propose Free()LM, a model that introduces an intrinsic self-forgetting capability via the Free-Module, a plug-and-play LoRA adapter. By iteratively switching between
reasoning and cleaning modes, Free()LM dynamically identifies and prunes useless context chunks, maintaining a compact and noise-free state. Extensive experiments show that Free()LM provides consistent improvements across all model scales (8B to 685B). It achieves a 3.3% average improvement over top-tier
reasoning baselines, even establishing a new SOTA on IMOanswerBench using DeepSeek V3.2-Speciale. Most notably, in long-horizon tasks where the standard Qwen3-235B-A22B model suffers a total collapse (0% accuracy), Free()LM restores performance to 50%. Our findings suggest that sustainable intelligence requires the freedom to forget as much as the power to think.
Read the full abstract ▾
All authors Yilun Zheng, Dongyang Ma, Tian Liang, Jiahao Xu, Xinting Huang, Lihui Chen, Haitao Mi, Yan Wang
Computation and Language
Diverge to Induce Prompting: Multi-Rationale Induction for Zero-Shot Reasoning
Po-Chun Chen, Hen-Hsen Huang and Hsin-Hsi Chen
To address the instability of unguided
reasoning paths in standard Chain-of-Thought prompting, recent methods guide large
language models (LLMs) by first eliciting a single
reasoning strategy. However, relying on just one strategy for each question can still limit performance across diverse tasks. We propose Diverge-to-Induce Prompting (DIP), a framework that first
prompts an LLM to generate multiple diverse high-level rationales for each question. Each rationale is then elaborated into a detailed, step-by-step draft plan. Finally, these draft plans are induced into a final plan. DIP enhances zero-shot
reasoning accuracy without reliance on resource-intensive sampling. Experiments show that DIP outperforms single-strategy prompting, demonstrating the effectiveness of multi-plan induction for
prompt-based
reasoning.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging
Ziyang Fan, Keyu Chen, Ruilong Xing et al. (+3 more)
Although
Video Large
Language Models (VLLMs) have shown remarkable capabilities in
video understanding, they are required to process high volumes of visual
tokens, causing significant computational inefficiency. Existing VLLMs acceleration frameworks usually compress spatial and temporal redundancy independently, which overlooks the spatiotemporal relationships, thereby leading to suboptimal spatiotemporal compression. The highly correlated visual features are likely to change in spatial position, scale, orientation, and other attributes over time due to the dynamic nature of
video. Building on this insight, we introduce FlashVID, a training-free inference acceleration framework for VLLMs. Specifically, FlashVID utilizes
Attention and Diversity-based
Token Selection (ADTS) to select the most representative
tokens for basic
video representation, then applies Tree-based Spatiotemporal
Token Merging (TSTM) for fine-grained spatiotemporal redundancy elimination. Extensive experiments conducted on three representative VLLMs across five
video understanding
benchmarks demonstrate the effectiveness and generalization of our method. Notably, by retaining only 10% of visual
tokens, FlashVID preserves 99.1% of the performance of LLaVA-OneVision. Consequently, FlashVID can serve as a training-free and plug-and-play module for extending long
video frames, which enables a 10x increase in
video frame input to Qwen2.5-VL, resulting in a relative improvement of 8.6% within the same computational budget. Code is available at https://github.com/Fanziyang-v/FlashVID.
Read the full abstract ▾
All authors Ziyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, Zhuotao Tian
Computation and Language · Cross-listed from cs.AI
Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective
Rui Li, Zeyu Zhang, Xiaohe Bo et al. (+4 more)
Multi-
agent architectures built on large
language models (LLMs) have demonstrated the potential to realize swarm intelligence through well-crafted collaboration. However, the substantial burden of manual orchestration inherently raises an imperative to automate the design of agentic workflows. We frame such an
agent coordination challenge as a classic problem in dynamic ad-hoc networking: How to establish adaptive and reliable communication among a scalable number of agentic hosts? In response to this unresolved dilemma, we introduce RAPS, a reputation-aware publish-subscribe paradigm for adaptive, scalable, and robust coordination of LLM
agents. RAPS is grounded in the Distributed Publish-Subscribe Protocol, allowing LLM
agents to exchange messages based on their declared intents rather than predefined
topologies. Beyond this substrate, RAPS further incorporates two coherent overlays: (i) Reactive Subscription, enabling
agents to dynamically refine their intents; and (ii)
Bayesian Reputation, empowering each
agent with a local watchdog to detect and isolate malicious peers. Extensive experiments over five
benchmarks showcase that our design effectively reconciles adaptivity, scalability, and robustness in a unified multi-
agent coordination framework.
Read the full abstract ▾
All authors Rui Li, Zeyu Zhang, Xiaohe Bo, Quanyu Dai, Chaozhuo Li, Feng Wen, Xu Chen
Computation and Language
ResidualKV: Residual-Based KV Cache Compression for Efficient Long-Context Inference
Jitai Hao, Qiang Huang, Yaowei Wang et al. (+2 more)
Efficient long-context inference faces two coupled bottlenecks: KV-cache
memory grows linearly with context length, while
attention computation grows quadratically. Existing approaches typically address one at the expense of irreversible
token eviction, full-cache retention, or full-history reconstruction, limiting their effectiveness for multi-turn interaction and long-form
reasoning. Motivated by two empirical properties, Long-Range Inter-
Token Similarity and Smooth Residual Distribution, we propose ResidualKV, which factorizes the KV cache into a sparse set of globally retrieved references and compact, quantized residual codes for the remaining
tokens. This representation preserves
token-specific information without permanent eviction and, when combined with sparse
attention, reconstructs only the selected states on demand. Dynamic-stride scheduling further reduces reference growth from linear to approximately logarithmic at ultra-long contexts. Across Llama, Qwen, LLaVA-OV, and Qwen3-VL backbones, ResidualKV maintains near-full-cache performance using only 13%-16% KV storage and 30%
attention computation on LongBench, and 8%-10% storage and 10% computation in matched-budget
multimodal evaluation. It also accelerates decoding by up to $1.5\times$ with KV-cache quantization and $3.4\times$ without it. These results show that global cross-
token redundancy supports accurate,
memory-efficient, and computation-efficient long-context inference. The source code is available at https://github.com/CURRENTF/ResidualKV.
Read the full abstract ▾
All authors Jitai Hao, Qiang Huang, Yaowei Wang, Min Zhang, Jun Yu
Computation and Language
The Judge Who Never Admits: Hidden Shortcuts in LLM-based Evaluation
Arash Marioriyad, Omid Ghahroodi, Ehsaneddin Asgari et al. (+2 more)
Large
language models (LLMs) are increasingly used as automatic judges to evaluate system outputs in tasks such as
reasoning, question answering, and creative writing. A faithful judge should base its verdicts solely on content quality, remain invariant to irrelevant context, and transparently reflect the factors driving its decisions. We test this ideal via controlled cue perturbations-synthetic metadata labels injected into evaluation
prompts-for six judge models: GPT-4o, Gemini-2.0-Flash, Gemma-3-27B, Qwen3-235B, Claude-3-Haiku, and Llama3-70B. Experiments span two complementary
datasets with distinct evaluation regimes: ELI5 (factual QA) and LitBench (open-ended creative writing). We study six cue families: source, temporal, age, gender, ethnicity, and educational status. Beyond measuring verdict shift rates (VSR), we introduce cue acknowledgment rate (CAR) to quantify whether judges explicitly reference the injected cues in their natural-
language rationales. Across cues with strong behavioral effects-e.g., provenance hierarchies (Expert > Human > LLM > Unknown), recency preferences (New > Old), and educational-status favoritism-CAR is typically at or near zero, indicating that shortcut reliance is largely unreported even when it drives decisions. Crucially, CAR is also
dataset-dependent: explicit cue recognition is more likely to surface in the factual ELI5 setting for some models and cues, but often collapses in the open-ended LitBench regime, where large verdict shifts can persist despite zero acknowledgment. The combination of substantial verdict sensitivity and limited cue acknowledgment reveals an explanation gap in LLM-as-judge pipelines, raising concerns about reliability of model-based evaluation in both research and deployment.
Read the full abstract ▾
All authors Arash Marioriyad, Omid Ghahroodi, Ehsaneddin Asgari, Mohammad Hossein Rohban, Mahdieh Soleymani Baghshah
Computation and Language · Cross-listed from cs.AI
Accelerating Social Science Research via Agentic Hypothesization and Experimentation
Jishu Sen Gupta, Harini SI, Somesh Kumar Singh et al. (+5 more)
Data-driven social science research is inherently slow, relying on iterative cycles of observation, hypothesis generation, and experimental validation. While recent data-driven methods promise to accelerate parts of this process, they largely fail to support end-to-end scientific discovery. To address this gap, we introduce EXPERIGEN, an agentic framework that operationalizes end-to-end discovery through a
Bayesian optimization inspired two-phase search, in which a Generator proposes candidate hypotheses and an Experimenter evaluates them empirically. Across multiple domains, EXPERIGEN consistently discovers 2-4x more statistically significant hypotheses that are 7-17 percent more predictive than prior approaches, and naturally extends to complex data regimes including
multimodal and relational
datasets. Beyond statistical performance, hypotheses must be novel, empirically grounded, and actionable to drive real scientific progress. To evaluate these qualities, we conduct an expert review of machine-generated hypotheses, collecting feedback from senior faculty. Among 25 reviewed hypotheses, 88 percent were rated moderately or strongly novel, 70 percent were deemed impactful and worth pursuing, and most demonstrated rigor comparable to senior graduate-level research. Finally, recognizing that ultimate validation requires real-world evidence, we conduct the first A/B test of LLM-generated hypotheses, observing statistically significant results with p less than 1e-6 and a large effect size of 344 percent.
Read the full abstract ▾
All authors Jishu Sen Gupta, Harini SI, Somesh Kumar Singh, Syed Mohamad Tawseeq, Yaman Kumar Singla, David Doermann, Rajiv Ratn Shah, Balaji Krishnamurthy
Computation and Language
Cross-Linguistic Persona-Driven Data Synthesis for Robust Multimodal Cognitive Decline Detection
Rui Feng, Zhiyao Luo, Liuyu Wu et al. (+6 more)
Speech-based digital biomarkers represent a scalable, non-invasive frontier for the early identification of Mild Cognitive Impairment (MCI). However, the development of robust diagnostic models remains impeded by acute clinical data scarcity and a lack of interpretable
reasoning. Current solutions frequently struggle with cross-lingual generalization and fail to provide the transparent rationales essential for clinical trust. To address these barriers, we introduce SynCog, a novel framework integrating controllable zero-shot
multimodal data synthesis with Chain-of-Thought (CoT) deduction fine-tuning. Specifically, SynCog simulates diverse virtual subjects with varying cognitive profiles to effectively alleviate clinical data scarcity. This generative paradigm enables the rapid, zero-shot expansion of clinical corpora across diverse
languages, effectively bypassing data bottlenecks in low-resource settings and bolstering the diagnostic performance of
Multimodal Large
Language Models (MLLMs). Leveraging this synthesized
dataset, we fine-tune a foundational
multimodal backbone using a CoT deduction strategy, empowering the model to explicitly articulate diagnostic thought processes rather than relying on black-box predictions. Extensive experiments on the ADReSS and ADReSSo
benchmarks demonstrate that augmenting limited clinical data with synthetic phenotypes yields competitive diagnostic performance, achieving Macro-F1 scores of 80.67% and 78.46%, respectively, outperforming current baseline models. Furthermore, evaluation on an independent real-world Mandarin cohort (CIR-E) demonstrates robust cross-linguistic generalization, attaining a Macro-F1 of 48.71%. These findings constitute a critical step toward providing clinically trustworthy and linguistically inclusive cognitive assessment tools for global healthcare.
Read the full abstract ▾
All authors Rui Feng, Zhiyao Luo, Liuyu Wu, Wei Wang, Yuting Song, Yong Liu, Kok Pin Ng, Jianqing Li, Xingyao Wang
Computation and Language
Lost in Translation? A Comparative Study on the Cross-Lingual Transfer of Composite Harms
Vaibhav Shukla, Hardik Sharma, Adith N Reganti et al. (+3 more)
Most safety evaluations of large
language models (LLMs) remain anchored in English. Translation is often used as a shortcut to probe
multilingual behavior, but it rarely captures the full picture, especially when harmful intent or structure morphs across
languages. Some types of harm survive translation almost intact, while others distort or disappear. To study this effect, we introduce CompositeHarm, a translation-based
benchmark designed to examine how safety
alignment holds up as both syntax and semantics shift. It combines two complementary English
datasets, AttaQ, which targets structured
adversarial attacks, and MMSafetyBench, which covers contextual, real-world harms, and extends them into six
languages: English, Hindi, Assamese, Marathi, Kannada, and Gujarati. Using three large models, we find that attack success rates rise sharply in Indic
languages, especially under
adversarial syntax, while contextual harms transfer more moderately. To ensure scalability and energy efficiency, our study adopts lightweight inference strategies inspired by edge-AI design principles, reducing redundant evaluation passes while preserving cross-lingual fidelity. This design makes large-scale
multilingual safety testing both computationally feasible and environmentally conscious. Overall, our results show that translated
benchmarks are a necessary first step, but not a sufficient one, toward building grounded, resource-aware,
language-adaptive safety systems.
Read the full abstract ▾
All authors Vaibhav Shukla, Hardik Sharma, Adith N Reganti, Soham Wasmatkar, Bagesh Kumar, Vrijendra Singh
Computation and Language
Bielik Guard: Efficient Polish Language Safety Classifiers for LLM Content Moderation
Krzysztof Wróbel, Jan Maria Kowalski, Jerzy Surma et al. (+2 more)
As Large
Language Models (LLMs) become increasingly deployed in Polish
language applications, the need for efficient and accurate content safety classifiers has become paramount. We present Bielik Guard, a family of compact Polish
language safety classifiers comprising two model variants: a 0.1B parameter model based on MMLW-RoBERTa-base and a 0.5B parameter model based on PKOBP/polish-roberta-8k. Fine-tuned on a community-annotated
dataset of 6,885 Polish texts, these models classify content across five safety categories: Hate/Aggression, Vulgarities, Sexual Content, Crime, and Self-Harm. Our evaluation demonstrates that both models achieve strong performance on multiple
benchmarks. The 0.5B variant offers the best overall discrimination capability with F1 scores of 0.791 (micro) and 0.785 (macro) on the test set, while the 0.1B variant demonstrates exceptional efficiency. Notably, Bielik Guard 0.1B v1.1 achieves superior precision (77.65%) and very low false positive rate (0.63%) on real user
prompts, outperforming HerBERT-PL-Guard (31.55% precision, 4.70% FPR) despite identical model size. The models are publicly available and designed to provide appropriate responses rather than simple content blocking, particularly for sensitive categories like self-harm.
Read the full abstract ▾
All authors Krzysztof Wróbel, Jan Maria Kowalski, Jerzy Surma, Igor Ciuciura, Maciej Szymański
Computation and Language
Patches of Nonlinearity: Instruction Vectors in Large Language Models
Irina Bigoulaeva, Jonas Rohweder, Subhabrata Dutta et al. (+1 more)
Despite the recent success of instruction-tuned
language models and their ubiquitous usage, very little is known of how models process instructions internally. In this work, we address this gap from a mechanistic point of view by investigating how instruction-specific representations are constructed and utilized in different stages of post-training: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Via causal mediation, we identify that instruction representation is fairly localized in models. These representations, which we call Instruction Vectors (IVs), demonstrate a curious juxtaposition of linear separability along with non-linear causal interaction, broadly questioning the scope of the linear representation hypothesis commonplace in mechanistic interpretability. To disentangle the non-linear causal interaction, we propose a novel method to localize information processing in
language models that is free from the implicit linear assumptions of patching-based techniques. We find that, conditioned on the task representations formed in the early layers, different information pathways are selected in the later layers to solve that task, i.e., IVs act as
circuit selectors.
Read the full abstract ▾
All authors Irina Bigoulaeva, Jonas Rohweder, Subhabrata Dutta, Iryna Gurevych
Computation and Language
SparseEval: Efficient Evaluation of Large Language Models by Sparse Optimization
Taolin Zhang, Hang Guo, Wang Lu et al. (+3 more)
As large
language models (LLMs) continue to scale up, their performance on various downstream tasks has significantly improved. However, evaluating their capabilities has become increasingly expensive, as performing inference on a large number of
benchmark samples incurs high computational costs. In this paper, we revisit the model-item performance
matrix and show that it exhibits sparsity, that representative items can be selected as anchors, and that the task of efficient benchmarking can be formulated as a sparse optimization problem. Based on these insights, we propose SparseEval, a method that, for the first time, adopts gradient descent to optimize anchor weights and employs an iterative refinement strategy for anchor selection. We utilize the representation capacity of MLP to handle sparse optimization and propose the Anchor Importance Score and Candidate Importance Score to evaluate the value of each item for task-aware refinement. Extensive experiments demonstrate the low estimation error and high Kendall's~$\tau$ of our method across a variety of
benchmarks, showcasing its superior robustness and practicality in real-world scenarios. Code is available at {https://github.com/taolinzhang/SparseEval}.
Read the full abstract ▾
All authors Taolin Zhang, Hang Guo, Wang Lu, Tao Dai, Shu-Tao Xia, Jindong Wang
Computation and Language · Cross-listed from cs.LG
Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection
Guanglong Sun, Siyuan Zhang, Liyuan Wang et al. (+3 more)
Safety post-training can improve the harmfulness and policy compliance of Large
Language Models (LLMs), but it may also reduce general utility, a phenomenon often described as the
alignment tax. We study this trade-off through the lens of continual learning: sequential
alignment stages expose the model to shifted data distributions and objectives, and their gradients may interfere with directions that support previously acquired general capabilities. This view does not claim that all
alignment degradation has a single cause; rather, it provides a useful first-order mechanism for mitigating one important source of capability regression. We propose Orthogonal Gradient Projection for Safety
Alignment (OGPSA), a lightweight update rule that estimates a low-rank reference subspace from gradients on a small set of general-capability data and removes from each safety gradient the component lying in this subspace. The resulting update is the steepest local safety-descent direction subject to first-order preservation constraints on the reference objectives. OGPSA is compatible with standard post-training pipelines and avoids large-scale replay, although it introduces periodic reference-gradient computation. Across Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and sequential SFT$\rightarrow$DPO settings, OGPSA improves the observed safety–utility trade-off over standard baselines. Under the sequential SFT$\rightarrow$DPO pipeline, the average performance gain increases from 33.98% to 42.74% on Qwen2.5-7B-Instruct and from 19.74% to 32.98% on Llama3.1-8B-Instruct. We have open sourced our code at https://github.com/SunGL001/OGPSA.
Read the full abstract ▾
All authors Guanglong Sun, Siyuan Zhang, Liyuan Wang, Jun Zhu, Hang Su, Yi Zhong
Computation and Language · Cross-listed from cs.AI
Emergent Misalignment is Easy, Narrow Misalignment is Hard
Anna Soligo, Edward Turner, Senthooran Rajamanoharan et al. (+1 more)
Finetuning large
language models on narrowly harmful
datasets can cause them to become emergently misaligned, giving stereotypically `evil' responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases and find that models can just learn the narrow
dataset task, but that the general solution appears to be more stable and more efficient. To establish this, we build on the result that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. We find a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and preliminary metrics for investigating how inductive biases shape generalisation in LLMs. We open-source all code,
datasets and model finetunes.
Read the full abstract ▾
All authors Anna Soligo, Edward Turner, Senthooran Rajamanoharan, Neel Nanda
Computation and Language
Evaluating and Calibrating LLM Confidence on Questions with Multiple Correct Answers
Yuhan Wang, Shiyu Ni, Zhikai Ding et al. (+3 more)
Confidence calibration is essential for making large
language models (LLMs) reliable, yet existing training-free methods have been primarily studied under single-answer question answering. In this paper, we show that these methods break down in the presence of multiple valid answers, where disagreement among equally correct responses leads to systematic underestimation of confidence. To enable a systematic study of this phenomenon, we introduce MACE, a
benchmark of 12,000 factual questions spanning six domains with varying numbers of correct answers. Experiments across 15 representative calibration methods and four LLM families (7B-72B) reveal that while accuracy increases with answer cardinality, estimated confidence consistently decreases, causing severe miscalibration for questions with mixed answer counts. To address this issue, we propose Semantic Confidence Aggregation (SCA), which aggregates confidence over multiple high-probability sampled responses. SCA achieves state-of-the-art calibration performance under mixed-answer settings while preserving strong calibration on single-answer questions.
Read the full abstract ▾
All authors Yuhan Wang, Shiyu Ni, Zhikai Ding, Zihang Zhan, Yuanzi Li, Keping Bi
Computation and Language
TodoEvolve: Learning to Architect Agent Planning Systems
Jiaxi Liu, Yanzuo Jiang, Guibin Zhang et al. (+5 more)
Planning has become a central capability for contemporary
agent systems in navigating complex, long-horizon tasks, yet existing approaches predominantly rely on fixed, hand-crafted planning structures that lack the flexibility to adapt to the structural diversity of open-ended problems. To address this limitation, we introduce TodoEvolve, a meta-planning paradigm that autonomously synthesizes and dynamically revises task-specific planning architectures. Specifically, we first construct PlanFactory, a modular design space that standardizes diverse planning paradigms within a unified codebase encompassing
topology, initialization, adaptation, and
navigation, thereby providing a common interface for heterogeneous planning patterns. Leveraging PlanFactory, we collect high-quality planning trajectories and train Todo-14B via Impedance-Guided Preference Optimization (IGPO), a multi-objective reinforcement learning objective that encourages the generation of planning systems that are performant, stable, and
token-efficient across arbitrary tasks and
agent backbones. Empirical evaluations on five agentic
benchmarks demonstrate that TodoEvolve consistently surpasses carefully engineered planning modules while maintaining economical API costs and runtime overhead.
Read the full abstract ▾
All authors Jiaxi Liu, Yanzuo Jiang, Guibin Zhang, Zihan Zhang, Heng Chang, Zhenfei Yin, Qibing Ren, Junchi Yan
Computation and Language · Cross-listed from cs.CV
SPD-Faith Bench: Diagnosing and Improving Faithfulness in Chain-of-Thought for Multimodal Large Language Models
Weijiang Lv, Yaoxuan Feng, Xiaobo Xia et al. (+4 more)
Chain-of-Thought
reasoning is widely used to improve the interpretability of
multimodal large
language models (MLLMs), yet the faithfulness of the generated
reasoning traces remains unclear. Prior work has mainly focused on perceptual
hallucinations, leaving
reasoning level unfaithfulness underexplored. To isolate faithfulness from linguistic priors, we introduce SPD-Faith Bench, a diagnostic
benchmark based on fine-grained image difference
reasoning that enforces explicit visual comparison. Evaluations on state-of-the-art MLLMs reveal two systematic failure modes, perceptual blindness and perception-
reasoning dissociation. We trace these failures to decaying visual
attention and representation shifts in the residual stream. Guided by this analysis, we propose SAGE, a train-free visual evidence-calibrated framework that improves visual routing and aligns
reasoning with perception. Our results highlight the importance of explicitly evaluating faithfulness beyond response correctness. Our
benchmark and codes are available at https://github.com/Johanson-colab/SPD-Faith-Bench.
Read the full abstract ▾
All authors Weijiang Lv, Yaoxuan Feng, Xiaobo Xia, Jiayu Wang, Yan Jing, Wenchao Chen, Bo Chen
Computation and Language · Cross-listed from cs.AI
Data Darwinism Part I: Unlocking the Value of Scientific Data for Pre-training
Yiwei Qin, Zhen Huang, Tiantian Mi et al. (+5 more)
Data quality determines foundation model performance, yet systematic processing frameworks are lacking. We introduce Data Darwinism, a ten-level taxonomy (L0-L9) that conceptualizes data-model co-
evolution: advanced models produce superior data for next-generation systems. We validate this on scientific literature by constructing Darwin-Science, a 900B-
token corpus (L0-L5). We identify a learnability gap in raw scientific text, which we bridge via L4 (Generative Refinement) and L5 (Cognitive Completion) using frontier LLMs to explicate
reasoning and terminology. To ensure rigorous attribution, we pre-trained daVinci-origin-3B/7B models from scratch, excluding scientific content to create contamination-free baselines. After 600B
tokens of continued pre-training, Darwin-Science outperforms baselines by +2.12 (3B) and +2.95 (7B) points across 20+
benchmarks, rising to +5.60 and +8.40 points on domain-aligned tasks. Systematic progression to L5 yields a +1.36 total gain, confirming that higher-level processing unlocks latent data value. We release the Darwin-Science corpus and daVinci-origin models to enable principled, co-evolutionary development.
Read the full abstract ▾
All authors Yiwei Qin, Zhen Huang, Tiantian Mi, Weiye Si, Chenyang Zhou, Qipeng Guo, Siyuan Feng, Pengfei Liu
Computation and Language
LLMs Know More About Numbers than They Can Say
Fengting Yuchi, Li Du and Jason Eisner
Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: "Which is larger, $5.7 \times 10^2$ or $580$?" This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe's log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models' internal magnitude representations can enhance their numerical
reasoning capabilities. Our code is available at https://github.com/VCY019/Numeracy-Probing.
Read the full abstract ▾
Computation and Language
Pruning as a Cooperative Game: Surrogate-Assisted Layer Contribution Estimation for Large Language Models
Xuan Ding, Pengyu Tong, Ranjie Duan et al. (+3 more)
While large
language models (LLMs) demonstrate impressive performance across various tasks, their deployment in real-world scenarios is still constrained by high computational demands. Layer-wise
pruning, a commonly employed strategy to mitigate inference costs, can partially address this challenge. However, existing approaches generally depend on static heuristic rules and fail to account for the interdependencies among layers, thereby limiting the effectiveness of the
pruning process. To this end, this paper proposes a game-theoretic framework that formulates layer
pruning as a cooperative game in which each layer acts as a player and model performance serves as the utility. As computing exact Shapley values is computationally infeasible for large
language models (LLMs), we propose using a lightweight surrogate network to estimate layer-wise marginal contributions. This network can predict LLM performance for arbitrary layer combinations at a low computational cost. Additionally, we employ stratified Monte Carlo mask sampling to further reduce the cost of Sharpley value estimation. This approach captures inter-layer dependencies and dynamically identifies critical layers for
pruning. Extensive experiments demonstrate the consistent superiority of our method in terms of perplexity and zero-shot accuracy, achieving more efficient and effective layer-wise
pruning for large
language models.
Read the full abstract ▾
All authors Xuan Ding, Pengyu Tong, Ranjie Duan, Yunjian Zhang, Rui Sun, Yao Zhu
Computation and Language
Thinking Is Not Telling: Information Disclosure in User-Service LLM Agents
Jiatong Li, Changdae Oh, Hyeong Kyu Choi et al. (+2 more)
User-engaged LLM
agents increasingly operate in service scenarios where task success depends on coordination between the
agent, the user, and a stateful environment. In such interactions, the
agent often has access to task policies, tool results, and environment states that the user does not observe. This makes
agent-user communication a central component of task completion. In this work, we study a communication-level failure mode of user-service
agents: user-invisible
reasoning does not necessarily translate into timely user-facing information disclosure. We introduce information disclosure as a measurable communication mechanism, covering state corrections, available options, constraints, consequences, and proactively retrieved environment information. We use this perspective to analyze the widely used user-invisible “think” tool, which allows
agents to reason internally before responding or acting. Across five models and three user-service scenarios, we first show that enforced thinking has unstable utility. We then analyze response scaling under different think efforts and find that more internal thinking does not reliably translate into richer user-facing communication. Using evidence-grounded probing and turn-level counterfactual intervention, we show that injecting golden decision-relevant information produces the clearest improvement in First Edit Correction Rate, outperforming verbosity and direct think-content controls. This indicates that missing disclosure can causally affect the
agent's immediate downstream environment-edit behavior. We further find that a lightweight disclosure-aware
prompt mitigates degradation from enforced thinking. These results suggest that the key challenge is not simply making
agents think more, but ensuring that private
reasoning and environment knowledge are externalized as timely, user-facing information disclosure.
Read the full abstract ▾
All authors Jiatong Li, Changdae Oh, Hyeong Kyu Choi, Jindong Wang, Sharon Li
Computation and Language
Emergent Structured Representations Support Flexible In-Context Inference in Large Language Models
Ningyu Xu, Qi Zhang, Xipeng Qiu et al. (+1 more)
Large
language models (LLMs) exhibit emergent behaviors suggestive of human-like
reasoning. While recent work has identified structured conceptual representations within these models, it remains unclear whether they functionally rely on such representations for
reasoning. Here we investigate the internal processing of LLMs during in-context inference across diverse tasks. Our results reveal a conceptual subspace emerging in middle to late layers, whose representational structure persists across contexts. Using causal mediation analyses, we demonstrate that this subspace is not merely an epiphenomenon but is functionally central to model predictions, establishing its causal role in inference. We further identify a layer-wise progression where
attention heads in early-to-middle layers integrate contextual cues to construct and refine the subspace, which is subsequently leveraged by later layers to generate predictions. Together, these findings provide evidence that LLMs dynamically construct and use structured latent representations in context for inference, offering insights into the computational processes underlying flexible adaptation.
Read the full abstract ▾
All authors Ningyu Xu, Qi Zhang, Xipeng Qiu, Xuanjing Huang
Computation and Language
Attn-GS: Attention-Guided Context Compression for Efficient Personalized LLMs
Shenglai Zeng, Tianqi Zheng, Chuan Tian et al. (+10 more)
Personalizing large
language models (LLMs) to individual users requires incorporating extensive interaction histories and profiles, but input
token constraints make this impractical due to high inference latency and API costs. Existing approaches rely on heuristic methods such as selecting recent interactions or prompting summarization models to compress user profiles. However, these methods treat context as a monolithic whole and fail to consider how LLMs internally process and prioritize different profile components. We investigate whether LLMs'
attention patterns can effectively identify important personalization signals for intelligent context compression. Through preliminary studies on representative personalization tasks, we discover that (a) LLMs'
attention patterns naturally reveal important signals, and (b) fine-tuning enhances LLMs' ability to distinguish between relevant and irrelevant information. Based on these insights, we propose Attn-GS, an
attention-guided context compression framework that leverages
attention feedback from a marking model to mark important personalization sentences, then guides a compression model to generate task-relevant, high-quality compressed user contexts. Extensive experiments demonstrate that Attn-GS significantly outperforms various baselines across different tasks,
token limits, and settings, achieving performance close to using full context while reducing
token usage by 50 times.
Read the full abstract ▾
All authors Shenglai Zeng, Tianqi Zheng, Chuan Tian, Dante Everaert, Yau-Shian Wang, Yupin Huang, Michael J. Morais, Rohit Patki, Jinjin Tian, Xinnan Dai, Kai Guo, Monica Xiao Cheng, Hui Liu