Computation and Language
Progressive Training for Explainable Citation-Grounded Dialogue: Reducing Hallucination to Zero in English-Hindi LLMs
Vedant Pandya
Knowledge-grounded dialogue systems aim to generate informative, contextually relevant responses by conditioning on external knowledge sources. However, most existing approaches focus exclusively on English, lack explicit citation mechanisms for verifying factual claims, and offer limited transparency into model decision-making. We present XKD-Dial, a progressive four-stage training pipeline for explainable, knowledge-grounded dialogue generation in a bilingual (English-Hindi) setting, comprising: (1)
multilingual adaptation, (2) English dialogue SFT with citation grounding, (3) bilingual dialogue SFT, and (4) GRPO
alignment with citation-aware rewards. We evaluate six models spanning encoder-decoder (250M-3B) and decoder-only (1B-7B) architectures at every pipeline stage. Our key contributions are: (i) three post-hoc explainability analyses - cross-
attention alignment, Integrated Gradients attribution, and occlusion-based causal grounding - applied systematically across the training trajectory to reveal how citation behaviour is learned, not only whether it is learned; (ii) citation-grounded SFT reduces
hallucination to 0.0% for encoder-decoder models from Stage 2 onward; (iii) the progressive pipeline prevents catastrophic forgetting while improving Hindi capabilities; (iv) smaller models match larger models on English after SFT; and (v) GRPO provides marginal improvement over well-designed SFT for structured citation tasks. We evaluate across six automatic metrics (BLEU, ROUGE, BERTScore, FactScore, Citation-F1, and
hallucination rate).
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Reasoning over mathematical objects: on-policy reward modeling and test time aggregation
Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim et al. (+18 more)
The ability to precisely derive mathematical objects is a core requirement for downstream STEM applications, including mathematics, physics, and chemistry, where
reasoning must culminate in formally structured expressions. Yet, current LM evaluations of mathematical and scientific
reasoning rely heavily on simplified answer formats such as numerical values or multiple choice options due to the convenience of automated assessment. In this paper we provide three contributions for improving
reasoning over mathematical objects: (i) we build and release training data and
benchmarks for deriving mathematical objects, the Principia suite; (ii) we provide training recipes with strong LLM-judges and verifiers, where we show that on-policy judge training boosts performance; (iii) we show how on-policy training can also be used to scale test-time compute via aggregation. We find that strong LMs such as Qwen3-235B and o3 struggle on Principia, while our training recipes can bring significant improvements over different LLM backbones, while simultaneously improving results on existing numerical and MCQA tasks, demonstrating cross-format generalization of
reasoning abilities.
Read the full abstract ▾
All authors Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim, Ilia Kulikov, Jack Lanchantin, Xian Li, Tianjian Li, Bo Liu, Graham Neubig, Anaelia Ovalle, Swarnadeep Saha, Sainbayar Sukhbaatar, Sean Welleck, Jason Weston, Chenxi Whitehouse, Adina Williams, Jing Xu, Ping Yu, Weizhe Yuan, Jingyu Zhang, Wenting Zhao
Computation and Language
A Human-in/on-the-Loop Framework for Accessible Text Generation
Lourdes Moreno and Paloma Martínez
Plain
Language and Easy-to-Read formats in text simplification are essential for cognitive accessibility. Yet current automatic simplification and evaluation pipelines remain largely automated, metric-driven, and fail to reflect user comprehension or normative standards. This paper introduces a hybrid framework that explicitly integrates human participation into LLM-based accessible text generation. Human-in-the-Loop (HiTL) contributions guide adjustments during generation, while Human-on-the-Loop (HoTL) supervision ensures systematic post-generation review. Empirical evidence from user studies and annotated resources is operationalized into (i) checklists aligned with standards, (ii) Event-Condition-Action trigger rules for activating expert oversight, and (iii) accessibility Key Performance Indicators (KPIs). The framework shows how human-centered mechanisms can be encoded for evaluation and reused to provide structured feedback that improves model adaptation. By
embedding the human role in both generation and supervision, it establishes a traceable, reproducible, and auditable process for creating and evaluating accessible texts. In doing so, it integrates explainability and ethical accountability as core design principles, contributing to more transparent and inclusive NLP systems.
Read the full abstract ▾
Computation and Language
Evaluating LLM-Generated Lessons from the Language Learning Students' Perspective: A Short Case Study on Duolingo
Carlos Rafael Catalan, Patricia Nicole Monderin, Lheane Marie Dizon et al. (+3 more)
Popular
language learning applications such as Duolingo use large
language models (LLMs) to generate lessons for its users. Most lessons focus on general real-world scenarios such as greetings, ordering food, or asking directions, with limited support for profession-specific contexts. This gap can hinder learners from achieving professional-level fluency, which we define as the ability to communicate comfortably various work-related and domain-specific information in the target
language. We surveyed five employees from a multinational company in the Philippines on their experiences with Duolingo. Results show that respondents encountered general scenarios more frequently than work-related ones, and that the former are relatable and effective in building foundational grammar, vocabulary, and cultural knowledge. The latter helps bridge the gap toward professional fluency as it contains domain-specific vocabulary. Each participant suggested lesson scenarios that diverge in contexts when analyzed in aggregate. With this understanding, we propose that
language learning applications should generate lessons that adapt to an individual's needs through personalized, domain specific lesson scenarios while maintaining foundational support through general, relatable lesson scenarios.
Read the full abstract ▾
All authors Carlos Rafael Catalan, Patricia Nicole Monderin, Lheane Marie Dizon, Gap Estrella, Raymund John Sarmimento, Marie Antoinette Patalagsa
Computation and Language
Why Better Cross-Lingual Alignment Fails for Better Cross-Lingual Transfer: Case of Encoders
Yana Veitsman, Yihong Liu and Hinrich Schütze
Cross-lingual
alignment is often assumed to improve cross-lingual transfer by bringing representations of different
languages closer together. However, improvements in representational
alignment do not consistently translate into better downstream performance. We investigate this disconnect using XLM-R models explicitly aligned across four
language pairs with
token-level, sentence-level, and masked-
language-modeling objectives. We evaluate their zero-shot transfer on a
token-level task (part-of-
speech tagging) and a sentence-level task (sentence classification), and analyze both representational changes and the gradients induced by the
alignment and downstream objectives. We find that
embedding-based
alignment metrics do not reliably indicate whether
alignment will improve or degrade downstream performance. Moreover,
alignment and downstream-task gradients are often nearly orthogonal, particularly when the
alignment objective and downstream task operate at different representational levels. These findings suggest that representation
alignment alone is insufficient for assessing cross-lingual transfer, and that the compatibility between
alignment and downstream objectives should be considered when designing/evaluating
alignment methods.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
RewardFlow: Topology-Aware Reward Propagation on State Graphs for Agentic RL with Large Language Models
Xiao Feng, Bo Han, Zhanke Zhou et al. (+5 more)
Reinforcement learning (RL) shows promise for enhancing LLM agentic
reasoning, yet sparse terminal rewards hinder fine-grained optimization. Process reward modeling offers an alternative but incurs high computational costs, reward hacking risks, and annotation bottlenecks. We introduce RewardFlow, a lightweight method for estimating state-level rewards in agentic
reasoning. By constructing state graphs that capture the intrinsic topological structure of trajectories, RewardFlow performs
topology-aware propagation to estimate each state's contribution to success, yielding principled, annotation-free dense rewards. Used for RL optimization, RewardFlow substantially outperforms prior baselines across four agentic
benchmarks: +6.2% average success rate on text-based tasks, +29.7% on visual
reasoning over the strongest baseline across three model scales, and +10% accuracy on DeepResearch, with superior robustness and training efficiency. The implementation of RewardFlow is publicly available at https://github.com/tmlr-group/RewardFlow.
Read the full abstract ▾
All authors Xiao Feng, Bo Han, Zhanke Zhou, Jiaqi Fan, Jiangchao Yao, Ka Ho Li, Dahai Yu, Michael Kwok-Po Ng
Computation and Language
Structure of Basic Human Values in Russian Social Media
Maria Milkova and Maksim Rudnev
What is known about the values of national populations rests almost entirely on questionnaires, which
prompt respondents to rate researcher-supplied descriptions of each motivation. Social media instead records values as they are invoked spontaneously, in communication addressed to a real audience. We measure value expression in broadly sampled, heterogeneous Russian-
language social-media data, where value-relevant content is sparse and often ambiguous, and read the result against the survey-based profile of the same population. To address subjectivity in value interpretation, we compare alternative LLM-
prompt configurations against expert judgments and use error-driven calibration to select a theory-based operationalization. Using VKontakte data collected from one million randomly generated user IDs, comprising 7.5 million public text posts, we develop a multi-stage framework combining candidate selection, repeated LLM annotation, and transfer to an XLM-RoBERTa encoder for large-scale prediction (a macro PR-AUC of .60 on the expert
benchmark, with continuous scores positively correlated with graded expert confidence across all ten values). The circular structure proposed by Schwartz is largely recovered (congruence .95 at the post level, .94 at the user level). The aggregate profile agrees with survey findings at the anchors - Benevolence highest and broadly distributed across users, Power lowest - but shows relatively elevated Hedonism and relatively low Universalism. Universalism is further displaced structurally, co-expressed with Security and Tradition and concentrated in the output of highly active users.
Read the full abstract ▾
Computation and Language
Mi:dm K 2.5 Pro
KT Tech innovation Group
The evolving LLM landscape requires capabilities beyond simple text generation, prioritizing multi-step
reasoning, long-context understanding, and agentic workflows. This shift challenges existing models in enterprise environments, especially in Korean-
language and domain-specific scenarios where scaling is insufficient. We introduce Mi:dm K 2.5 Pro, a 32B parameter flagship LLM designed to address enterprise-grade complexity through
reasoning-focused optimization. Our methodology builds a robust data foundation via a quality-centric curation pipeline utilizing abstract syntax tree (AST) analysis for code, gap-filling synthesis for mathematics, and an LLM-based quality evaluator. Pre-training scales the model via layer-predictor-based Depth Upscaling (DuS) and a progressive strategy supporting a 128K
token context window. Post-training introduces a specialized multi-stage pipeline, including
Reasoning SFT, model merging, and asynchronous reinforcement learning (RL), to develop complex problem-solving skills. "Fusion Training" then rebalances these capabilities with conversational fluency, consistent response styling, and reliable tool-use. The evaluations show that Mi:dm K 2.5 Pro achieves competitive performance against leading global and domestic models. In addition, it sets state-of-the-art results on Korean-specific
benchmarks, showcasing deep linguistic and cultural understanding. Finally, Responsible AI evaluations validate safety against attacks, ensuring a secure profile for deployment with a balance of harmlessness and responsiveness.
Read the full abstract ▾
Computation and Language
Implicit Grading Bias in Large Language Models: How Writing Style Affects Automated Assessment Across Math, Programming, and Essay Tasks
Rudra Jadhav, Janhavi Danve and Sonalika Shaw
As large
language models (LLMs) are increasingly deployed as automated graders in educational settings, concerns about
fairness and bias in their evaluations have become critical. This study investigates whether LLMs exhibit implicit grading bias based on writing style when the underlying content correctness remains constant. We constructed a controlled
dataset of 180 student responses across three subjects (Mathematics, Programming, and Essay/Writing), each with three surface-level perturbation types: grammar errors, informal
language, and non-native phrasing. Two state-of-the-art open-source LLMs – LLaMA 3.3 70B (Meta) and Qwen 2.5 72B (Alibaba) – were prompted to grade responses on a 1-10 scale with explicit instructions to evaluate content correctness only and to disregard writing style. Our results reveal statistically significant grading bias in Essay/Writing tasks across both models and all perturbation types (p < 0.05), with effect sizes ranging from medium (Cohen's d = 0.64) to very large (d = 4.25). Informal
language received the heaviest penalty, with LLaMA deducting an average of 1.90 points and Qwen deducting 1.20 points on a 10-point scale – penalties comparable to the difference between a B+ and C+ letter grade. Non-native phrasing was penalized 1.35 and 0.90 points respectively. In sharp contrast, Mathematics and Programming tasks showed minimal bias, with most conditions failing to reach statistical significance. These findings demonstrate that LLM grading bias is subject-dependent, style-sensitive, and persists despite explicit counter-bias instructions in the grading
prompt. We discuss implications for equitable deployment of LLM-based grading systems and recommend bias
auditing protocols before institutional adoption.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Are complicated loss functions necessary for teaching LLMs to reason?
Gabriele Carrino, Andrea Sassella, Nicolo Brunello et al. (+2 more)
Recent advances in large
language models (LLMs) highlight the importance of post training techniques for improving
reasoning and mathematical ability. Group Relative Policy Optimization (GRPO) has shown promise in this domain by combining group relative advantage estimation, PPO style clipping, and KL regularization. However, its complexity raises the question of whether all components are necessary for fostering
reasoning behaviors. We conduct a systematic analysis of GRPO and identify two key findings: (1) incorporating negative feedback is essential training solely on actions above a baseline limits learning; and (2) PPO style constraints, such as policy ratio clipping, are not required to improve mathematical
reasoning or performance. Building on these insights, we propose REINFORCE with Group Relative Advantage (RGRA), a simplified variant that retains group relative advantage estimation but removes PPO style clipping and policy ratio terms. Experiments across standard mathematical
benchmarks indicate that RGRA has the potential to achieve stronger performance than GRPO. Our results suggest that simpler REINFORCE based approaches can effectively enhance
reasoning in LLMs, offering a more transparent and efficient alternative to GRPO.
Read the full abstract ▾
All authors Gabriele Carrino, Andrea Sassella, Nicolo Brunello, Federico Toschi, Mark James Carman
Computation and Language
Automatic detection of Gen-AI texts: A comparative framework of neural models
Cristian Buttaro and Irene Amerini
The rapid proliferation of Large
Language Models has significantly increased the difficulty of distinguishing between human-written and AI generated texts, raising critical issues across academic, editorial, and social domains. This paper investigates the problem of AI generated text detection through the design, implementation, and comparative evaluation of multiple machine learning based detectors. Four neural architectures are developed and analyzed: a Multilayer Perceptron, a one-dimensional Convolutional Neural Network, a MobileNet-based CNN, and a
Transformer model. The proposed models are benchmarked against widely used online detectors, including ZeroGPT, GPTZero, QuillBot, Originality.AI, Sapling, IsGen, Rephrase, and Writer. Experiments are conducted on the COLING
Multilingual Dataset, considering both English and Italian configurations, as well as on an original thematic
dataset focused on Art and Mental Health. Results show that supervised detectors achieve more stable and robust performance than commercial tools across different
languages and domains, highlighting key strengths and limitations of current detection strategies.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Memento-Skills: Let Agents Design Agents
Huichi Zhou, Siyuan Guo, Anjie Liu et al. (+14 more)
We introduce Memento-Skills, a generalist, continually-learnable LLM
agent system that functions as an
agent-designing
agent: it autonomously constructs, adapts, and improves task-specific
agents through experience. The system is built on a
memory-based reinforcement learning framework with stateful
prompts, where reusable skills (stored as structured markdown files) serve as persistent, evolving
memory. These skills encode both behaviour and context, enabling the
agent to carry forward knowledge across interactions. Starting from simple elementary skills (like Web search and terminal operations), the
agent continually improves via the Read–Write Reflective Learning mechanism introduced in Memento~2~\cite{wang2025memento2}. In the read phase, a behaviour-trainable skill router selects the most relevant skill conditioned on the current stateful
prompt; in the write phase, the
agent updates and expands its skill library based on new experience. This closed-loop design enables continual learning without updating LLM parameters, as all adaptation is realised through the
evolution of externalised skills and
prompts. Unlike prior approaches that rely on human-designed
agents, Memento-Skills enables a generalist
agent to design
agents end-to-end for new tasks. Through iterative skill generation and refinement, the system progressively improves its own capabilities. Experiments on the General AI Assistants
benchmark and Humanity's Last Exam demonstrate sustained gains, achieving 26.2% and 116.2% relative improvements in overall accuracy, respectively. Code is available at https://github.com/Memento-Teams/Memento-Skills.
Read the full abstract ▾
All authors Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou, Na Wang, Chunyang Sun, Jun Wang
Computation and Language · Cross-listed from cs.LG
CausalRM: Causal-Theoretic Reward Modeling for RLHF from Observational User Feedbacks
Hao Wang, Licheng Pan, Zhichao Chen et al. (+7 more)
Despite the success of reinforcement learning from human feedback (RLHF) in aligning
language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly conditions. In this work, we introduce observational reward modeling – learning reward models with observational user feedback (e.g., clicks, copies, and upvotes) – as a scalable and cost-effective alternative. We identify two fundamental challenges in this setting: (1) observational feedback is noisy due to annotation errors, which deviates it from true user preference; (2) observational feedback is biased by user preference, where users preferentially provide feedback on responses they feel strongly about, which creats a distribution shift between training and inference data. To address these challenges, we propose CausalRM, a causal-theoretic reward modeling framework that aims to learn unbiased reward models from observational feedback. To tackle challenge (1), CausalRM introduces a noise-aware surrogate loss term that is provably equivalent to the primal loss under noise-free conditions by explicitly modeling the annotation error generation process. To tackle challenge (2), CausalRM uses propensity scores – the probability of a user providing feedback for a given response – to reweight training samples, yielding a loss function that eliminates user preference bias. Extensive experiments across diverse LLM backbones and
benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on downstream RLHF tasks – including a 49.2% gain on WildGuardMix and a 32.7% improvement on HarmBench. Code is available on our project website.
Read the full abstract ▾
All authors Hao Wang, Licheng Pan, Zhichao Chen, Chunyuan Zheng, Zhixuan Chu, Xiaoxi Li, Yuan Lu, Xinggao Liu, Haoxuan Li, Zhouchen Lin
Computation and Language · Cross-listed from cs.LG
STEP: Scientific Time-Series Encoder Pretraining via Cross-Domain Distillation
Chen Zhang, Liwei Liu, Jun Tao et al. (+6 more)
Scientific time series are central to scientific AI but are typically sparse, highly heterogeneous, and limited in scale, making unified representation learning particularly challenging. Meanwhile, foundation models pretrained on relevant time series domains such as audio, general time series, and
brain signals contain rich knowledge, but their applicability to scientific signals remains underexplored. In this paper, we investigate the transferability and complementarity of foundation models from relevant time series domains, and study how to effectively leverage them to build a unified encoder for scientific time series. We first systematically evaluate relevant foundation models, showing the effectiveness of knowledge transfer to scientific tasks and their complementary strengths. Based on this observation, we propose STEP, a Scientific Time Series Encoder Pretraining framework via cross domain
distillation. STEP introduces adaptive patching to handle extreme-length sequences and a statistics compensation scheme to accommodate diverse numerical scales. It further leverages cross-domain
distillation to integrate knowledge from multiple foundation models into a unified encoder. By combining complementary representations across different domains, STEP learns general-purpose and transferable features tailored for scientific signals. Experiments on seven scientific time series tasks demonstrate that STEP provides both an effective structure and an effective pretraining paradigm, taking a STEP toward scientific time series representation learning.
Read the full abstract ▾
All authors Chen Zhang, Liwei Liu, Jun Tao, Xiaoyu Yang, Xuenan Xu, Kai Chen, Bowen Zhou, Wen Wu, Chao Zhang
Computation and Language · Cross-listed from cs.LG
HISR: Hindsight Information Modulated Segmental Process Rewards For Multi-turn Agentic Reinforcement Learning
Zhicong Lu, Zichuan Lin, Wei Jia et al. (+7 more)
While large
language models excel in diverse domains, their performance on complex longhorizon agentic decision-making tasks remains limited. Most existing methods concentrate on designing effective reward models (RMs) to advance performance via multi-turn reinforcement learning. However, they suffer from delayed propagation in sparse outcome rewards and unreliable credit assignment with potentially overly fine-grained and unfocused turnlevel process rewards. In this paper, we propose (HISR) exploiting Hindsight Information to modulate Segmental process Rewards, which closely aligns rewards with sub-goals and underscores significant segments to enhance the reliability of credit assignment. Specifically, a segment-level process RM is presented to assign rewards for each sub-goal in the task, avoiding excessively granular allocation to turns. To emphasize significant segments in the trajectory, a hindsight model is devised to reflect the preference of performing a certain action after knowing the trajectory outcome. With this characteristic, we design the ratios of sequence likelihoods between hindsight and policy model to measure action importance. The ratios are subsequently employed to aggregate segment importance scores, which in turn modulate segmental process rewards, enhancing credit assignment reliability. Extensive experimental results on three publicly
benchmarks demonstrate the validity of our method.
Read the full abstract ▾
All authors Zhicong Lu, Zichuan Lin, Wei Jia, Changyuan Tian, Deheng Ye, Peiguang Li, Li Jin, Nayu Liu, Guangluan Xu, Wei Feng
Computation and Language · Cross-listed from cs.SD
Words at Play: Benchmarking Audio Pun Understanding in Large Audio-Language Models
Yuchen Su, Shaoxin Zhong, Yonghua Zhu et al. (+6 more)
Puns represent a typical linguistic phenomenon that exploits polysemy and phonetic ambiguity to generate humour, posing unique challenges for natural
language understanding. Within pun research, audio plays a central role in human communication except text and images, while
datasets and systematic resources for spoken puns remain scarce, leaving this crucial modality largely underexplored. In this paper, we present APUN-Bench, the first
benchmark dedicated to evaluating large audio
language models (LALMs) on audio pun understanding. Our
benchmark contains 4,434 audio samples annotated across three stages: pun recognition, pun word location and pun meaning inference. We conduct a deep analysis of APUN-Bench by systematically evaluating 10 state-of-the-art LALMs, uncovering substantial performance gaps in recognizing, localizing, and interpreting audio puns. This analysis reveals key challenges, such as positional biases in audio pun location and error cases in meaning inference, offering actionable insights for advancing humour-aware audio intelligence.
Read the full abstract ▾
All authors Yuchen Su, Shaoxin Zhong, Yonghua Zhu, Ruofan Wang, Zijian Huang, Qiqi Wang, Na Zhao, Diana Benavides-Prado, Michael Witbrock
Computation and Language
A Comparative Empirical Study of Catastrophic Forgetting Mitigation in Sequential Task Adaptation for Continual Natural Language Processing Systems
Aram Abrahamyan and Sachin Kumar
Neural
language models deployed in real-world applications must continually adapt to new tasks and domains without forgetting previously acquired knowledge. This work presents a comparative empirical study of catastrophic forgetting mitigation in continual intent classification. Using the CLINC150
dataset, we construct a 10-task label-disjoint scenario and evaluate three backbone architectures: a feed-forward Artificial Neural Network (ANN), a Gated Recurrent Unit (GRU), and a
Transformer encoder, under a range of continual learning (CL) strategies. We consider one representative method from each major CL family: replay-based Maximally Interfered
Retrieval (MIR), regularization-based Learning without Forgetting (LwF), and parameter-isolation via Hard
Attention to Task (HAT), both individually and in all pairwise and triple combinations. Performance is assessed with average accuracy, macro F1, and backward transfer, capturing the stability-plasticity trade-off across the task sequence. Our results show that naive sequential fine-tuning suffers from severe forgetting for all architectures and that no single CL method fully prevents it. Replay emerges as a key ingredient: MIR is the most reliable individual strategy, and combinations that include replay (MIR+HAT, MIR+LwF, MIR+LwF+HAT) consistently achieve high final performance with near-zero or mildly positive backward transfer. The optimal configuration is architecture-dependent. MIR+HAT yields the best result for ANN and
Transformer, MIR+LwF+HAT, on the other hand, works the best for GRU, and in several cases CL methods even surpass joint training, indicating a regularization effect. These findings highlight the importance of jointly selecting backbone architecture and CL mechanism when designing continual intent-classification systems.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CR
MOSAIC: Multi-Objective Slice-Aware Iterative Curation for Alignment
Yipu Dou and Wang Yang
We study how to allocate a fixed supervised fine-tuning budget when three objectives must be balanced at once: multi-turn safety
alignment, low over-refusal on benign boundary queries, and instruction following under verifiable constraints. We propose MOSAIC (Multi-Objective Slice-Aware Iterative Curation for
Alignment), a multi-objective framework for closed-loop data mixture search built on a unified L1-L3 evaluation interface. MOSAIC turns slice-level failure profiles into executable data actions, including
dataset-level mixture ratios, bucket-level weights, and focus criteria. Under a fixed 1M-
token budget and five rounds of independent fine-tuning from the same base model, MOSAIC improves internal XGuard from 2.76 to 4.67 while keeping OrBench at 4.41 and IFEval at 3.65. The final Pareto solution also generalizes better than a random static LoRA baseline on independent attack, over-refusal, and capability tests, suggesting that structured failure diagnosis can serve as a practical control signal for budgeted data construction. Code is available at https://github.com/douyipu/mosaic.
Read the full abstract ▾
Computation and Language
Learning to Self-Evolve
Xiaoyin Chen, Canwen Xu, Yite Wang et al. (+3 more)
We introduce Learning to Self-Evolve (LSE), a reinforcement learning framework that trains large
language models (LLMs) to improve their own contexts at test time. We situate LSE in the setting of test-time self-
evolution, where a model iteratively refines its context from feedback on seen problems to perform better on new ones. Existing approaches rely entirely on the inherent
reasoning ability of the model and never explicitly train it for this task. LSE reduces the multi-step
evolution problem to a single-step RL objective, where each context edit is rewarded by the improvement in downstream performance. We pair this objective with a tree-guided
evolution loop. On Text-to-SQL generation (BIRD) and general question answering (MMLU-Redux), a 4B-parameter model trained with LSE outperforms self-evolving policies powered by GPT-5 and Claude Sonnet 4.5, as well as
prompt optimization methods including GEPA and TextGrad, and transfers to guide other models without additional training. Our results highlight the effectiveness of treating self-
evolution as a learnable skill.
Read the full abstract ▾
All authors Xiaoyin Chen, Canwen Xu, Yite Wang, Boyi Liu, Zhewei Yao, Yuxiong He
Computation and Language
DiscoPhon: Benchmarking the Unsupervised Discovery of Phoneme Inventories With Discrete Speech Units
Maxime Poli, Manel Khentout, Angelo Ortiz Tandazo et al. (+3 more)
We introduce DiscoPhon, a
multilingual benchmark for evaluating unsupervised phoneme discovery from discrete
speech units. DiscoPhon covers 6 dev and 6 test
languages, chosen to span a wide range of phonemic contrasts. Given only 10 hours of
speech in a previously unseen
language, systems must produce discrete units that are mapped to a predefined phoneme inventory, through either a many-to-one or a one-to-one assignment. The resulting sequences are evaluated for unit quality, recognition and
segmentation. We provide four pretrained
multilingual HuBERT and SpidR baselines, and show that phonemic information is available enough in current models for derived units to correlate well with phonemes, though with variations across
languages.
Read the full abstract ▾
All authors Maxime Poli, Manel Khentout, Angelo Ortiz Tandazo, Ewan Dunbar, Emmanuel Chemla, Emmanuel Dupoux
Computation and Language
Cross-Modal Rationale Transfer for Explainable Humanitarian Classification on Social Media
Thi Huyen Nguyen, Koustav Rudra and Wolfgang Nejdl
Advances in social media data dissemination enable the provision of real-time information during a crisis. The information comes from different classes, such as infrastructure damages, persons missing or stranded in the affected zone, etc. Existing methods attempted to classify text and images into various humanitarian categories, but their decision-making process remains largely opaque, which affects their deployment in real-life applications. Recent work has sought to improve transparency by extracting textual rationales from tweets to explain predicted classes. However, such explainable classification methods have mostly focused on text, rather than crisis-related images. In this paper, we propose an interpretable-by-design
multimodal classification framework. Our method first learns the joint representation of text and image using a visual
language transformer model and extracts text rationales. Next, it extracts the image rationales via the mapping with text rationales. Our approach demonstrates how to learn rationales in one modality from another through cross-modal rationale transfer, which saves annotation effort. Finally, tweets are classified based on extracted rationales. Experiments are conducted over CrisisMMD
benchmark dataset, and results show that our proposed method boosts the classification Macro-F1 by 2-35% while extracting accurate text
tokens and image patches as rationales. Human evaluation also supports the claim that our proposed method is able to retrieve better image rationale patches (12%) that help to identify humanitarian classes. Our method adapts well to new, unseen
datasets in zero-shot mode, achieving an accuracy of 80%.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
myMNIST: Benchmark of PETNN, KAN, and Classical Deep Learning Models for Burmese Handwritten Digit Recognition
Ye Kyaw Thu, Thazin Myint Oo and Thepchai Supnithi
We present the first systematic
benchmark on a standardized iteration of the publicly available Burmese Handwritten Digit
Dataset (BHDD), which we have designated as myMNIST Benchmarking. While BHDD serves as a foundational resource for Myanmar NLP/AI, it lacks a comprehensive, reproducible performance baseline across modern architectures. We evaluate eleven architectures spanning classical deep learning models (Multi-Layer Perceptron, Convolutional Neural Network, Long Short-Term
Memory, Gated Recurrent Unit,
Transformer), recent alternatives (FastKAN, EfficientKAN), an energy-based model (JEM), and physics-inspired PETNN variants (Sigmoid, GELU, SiLU). Using Precision, Recall, F1-Score, and Accuracy as evaluation metrics, our results show that the CNN remains a strong baseline, achieving the best overall scores (F1 = 0.9959, Accuracy = 0.9970). The PETNN (GELU) model closely follows (F1 = 0.9955, Accuracy = 0.9966), outperforming LSTM, GRU,
Transformer, and KAN variants. JEM, representing energy-based modeling, performs competitively (F1 = 0.9944, Accuracy = 0.9958). KAN-based models (FastKAN, EfficientKAN) trail the top performers but provide a meaningful alternative baseline (Accuracy ~0.992). These findings (i) establish reproducible baselines for BHDD across diverse modeling paradigms, (ii) highlight PETNN's strong performance relative to classical and
Transformer-based models, and (iii) quantify the gap between energy-inspired PETNNs and a true energy-based model (JEM). We release this
benchmark to facilitate future research on Myanmar digit recognition and to encourage broader evaluation of emerging architectures on regional scripts.
Read the full abstract ▾
Computation and Language
Language Model Maps for Prompt-Response Distributions via Log-Likelihood Vectors
Momose Oyama, Yusuke Takase and Hidetoshi Shimodaira
We propose a method that represents
language models by log-likelihood vectors over
prompt-response pairs and constructs model maps for comparing their conditional distributions. In this space, squared Euclidean distances between models are approximately proportional to the KL divergence between the corresponding conditional distributions. Experiments on a large collection of publicly available
language models show that the maps capture meaningful global structure, including relationships to model attributes and task performance. The representation also captures systematic shifts induced by
prompt modifications and their approximate additive compositionality; we use the vectors to predict downstream task scores and leverage their additive structure to approximate the effects of composite
prompt operations without directly observing the corresponding log-likelihood vectors. We further introduce PMI vectors to reduce the influence of unconditional distributions; in some cases, PMI-based model maps better reflect training-data-related differences. Overall, the framework supports the analysis and prediction of input-dependent model behavior.
Read the full abstract ▾
Computation and Language
ICE: Intervention-Consistent Explanation Evaluation with Statistical Grounding for LLMs
Abhinaba Basu and Pavan Chakraborty
Evaluating whether explanations faithfully reflect a model's
reasoning remains an open problem. Existing
benchmarks use single interventions without statistical testing, making it impossible to distinguish genuine faithfulness from chance-level performance. We show that faithfulness is not a fixed property but an operator-dependent quantity that changes with the intervention method used to measure it. We introduce ICE (Intervention-Consistent Explanation), a framework that evaluates explanations against random baselines of equal size under multiple operators. Evaluating 7 LLMs across 4 tasks with deletion and
retrieval infill operators, we find that switching operators crosses the positive-evidence threshold in 18% of configurations (5 of 28
attention comparisons), with gaps reaching 44 percentage points. Randomized baselines detect anti-faithfulness (explanations worse than random) in nearly one-third of English deletion configurations, invisible without random comparisons. These patterns persist across 6 non-English
languages and 2 attribution methods. The methodology generalizes to step-level chain-of-thought evaluation, where preliminary results on 3 frontier models suggest that high accuracy does not imply faithful
reasoning.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
SpecForge: A Flexible and Efficient Open-Source Training Framework for Speculative Decoding
Shenggui Li, Chao Wang, Yikai Zhu et al. (+14 more)
Large
language models incur high inference latency due to sequential autoregressive decoding. Speculative decoding alleviates this bottleneck by using a lightweight draft model to propose multiple
tokens for batched verification. However, its adoption has been limited by the lack of high-quality draft models and scalable training infrastructure. We introduce SpecForge, an open-source, production-oriented framework for training speculative decoding models with full support for EAGLE-3. SpecForge incorporates target-draft decoupling, hybrid parallelism, optimized training kernels, and integration with production-grade inference engines, enabling up to 9.9x faster EAGLE-3 training for Qwen3-235B-A22B. In addition, we release SpecBundle, a suite of production-grade EAGLE-3 draft models trained with SpecForge for mainstream open-source LLMs. Through a systematic study of speculative decoding training recipes, SpecBundle addresses the scarcity of high-quality drafts in the community, and our draft models achieve up to 4.48x end-to-end inference speedup on SGLang, establishing SpecForge as a practical foundation for real-world speculative decoding deployment.
Read the full abstract ▾
All authors Shenggui Li, Chao Wang, Yikai Zhu, Yubo Wang, Fan Yin, Shuai Shi, Yefei Chen, Xiaomin Dong, Qiaoling Chen, Jin Pan, Ji Li, Laixin Xie, Yineng Zhang, Lei Yu, Yonggang Wen, Ivor Tsang, Tianwei Zhang
Computation and Language
Cross-Lingual LLM-Judge Transfer via Evaluation Decomposition
Ivaxi Sheth, Zeno Jonke, Amin Mantrach et al. (+1 more)
As large
language models are increasingly deployed across diverse real-world applications, extending automated evaluation beyond English has become a critical challenge. Existing evaluation approaches are predominantly English-focused, and adapting them to other
languages is hindered by the scarcity and cost of human-annotated judgments in most
languages. We introduce a decomposition-based evaluation framework built around a Universal Criteria Set (UCS). UCS consists of a shared,
language-agnostic set of evaluation dimensions, producing an interpretable intermediate representation that supports cross-lingual transfer with minimal supervision. Experiments on multiple faithfulness tasks across
languages and model backbones demonstrate consistent improvements over strong baselines without requiring target-
language annotations.
Read the full abstract ▾
All authors Ivaxi Sheth, Zeno Jonke, Amin Mantrach, Saab Mansour
Computation and Language · Cross-listed from cs.LG
Balancing the Reasoning Load: Difficulty-Differentiated Policy Optimization with Length Redistribution for Efficient and Robust Reinforcement Learning
Yinan Xia, Haotian Zhang and Huiming Wang
Large
Reasoning Models (LRMs) have shown exceptional
reasoning capabilities, but they also suffer from the issue of overthinking, often generating excessively long and redundant answers. For problems that exceed the model's capabilities, LRMs tend to exhibit the overconfidence phenomenon, generating overly short but incorrect answers, which may contribute to suboptimal performance. To address these issues, we propose Difficulty-Differentiated Policy Optimization (DDPO), an efficient reinforcement learning
algorithm that optimizes simple and complex tasks separately based on the overconfidence phenomenon. Specifically, it reduces the output length for simple tasks without compromising accuracy, while for complex tasks, it expands the exploration space to improve performance. We further derive the theoretical conditions for maximizing expected accuracy, which require the length distribution to closely approximate the optimal length and be as concentrated as possible. Based on these conditions, we propose using the difficulty-level average as a well-founded reference for length optimization. Extensive experiments on both in-domain and out-of-domain
benchmarks validate the superiority and effectiveness of DDPO. Compared to GRPO, DDPO reduces the average answer length by 12% while improving accuracy by 1.85% across multiple
benchmarks, achieving a better trade-off between accuracy and length. The code is available at https://github.com/Yinan-Xia/DDPO.
Read the full abstract ▾
Computation and Language
When Names Change Verdicts: Intervention Consistency Reveals Systematic Bias in LLM Decision-Making
Abhinaba Basu and Pavan Chakraborty
Large
language models (LLMs) are increasingly used for high-stakes decisions, yet their susceptibility to spurious features remains poorly characterized. We introduce ICE-Guard, a framework applying intervention consistency testing to detect three types of spurious feature reliance: demographic (name/race swaps), authority (credential/prestige swaps), and framing (positive/negative restatements). Across 3,000 vignettes spanning 10 high-stakes domains, we evaluate 11 LLMs from 8 families and find that (1) authority bias (mean 5.8%) and framing bias (5.0%) substantially exceed demographic bias (2.2%), challenging the field's narrow focus on demographics; (2) bias concentrates in specific domains – finance shows 22.6% authority bias while criminal justice shows only 2.8%; (3) structured decomposition, where the LLM extracts features and a deterministic rubric decides, reduces flip rates by up to 100% (median 49% across 9 models). We demonstrate an ICE-guided detect-diagnose-mitigate-verify loop achieving cumulative 78% bias reduction via iterative
prompt patching. Validation against real COMPAS recidivism data shows COMPAS-derived flip rates exceed pooled synthetic rates, suggesting our
benchmark provides a conservative estimate of real-world bias. Code and data are publicly available.
Read the full abstract ▾
Computation and Language
Minsoo Cheong, Donghyun Son, Woosang Lim et al. (+1 more)
Diffusion-based large
language models (dLLMs) rely on bidirectional
attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum
entropy of newly decoded
token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded
token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature
volatility of decoded
tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded
tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\times$-$26.4\times$ speedup on standard
benchmarks and $22.4\times$-$24.1\times$ on chain-of-thought
benchmarks, with competitive accuracy and decision overhead accounting for only $0.5\%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.
Read the full abstract ▾
All authors Minsoo Cheong, Donghyun Son, Woosang Lim, Sungjoo Yoo
Computation and Language
The Truncation Blind Spot: How Decoding Strategies Systematically Exclude Human-Like Token Choices
Esteban Garces Arias, Nurzhan Sapargali, Christian Heumann et al. (+1 more)
Why does machine-generated text remain detectable? We investigate a mechanistic explanation at the decoding stage: standard strategies such as top-$k$ and
nucleus sampling restrict generation to high-probability
tokens, while human writers routinely choose contextually appropriate words from deeper in the model's probability distribution. Truncation makes a measurable share of these choices unreachable; we call this the truncation blind spot. Across five open models and three domains, 8–18% of human-selected
tokens fall outside common truncation boundaries. Linguistic analysis further reveals disproportionate exclusion of content-word
tokens. In a
benchmark comprising 1.8 million machine generations, classifiers using only predictability and lexical diversity achieve mean AUC-ROC near 0.97, with substantial variation across decoding settings and strong transfer across generators. Probability-floor samplers substantially narrow the blind spot, demonstrating that the choice of truncation criterion matters for retaining human-used
tokens. Together, these findings characterize a source of human–machine distributional mismatch and motivate decoding methods that preserve contextually appropriate low-probability choices while maintaining generation quality. Code and data are available at https://github.com/EstebanGarces/human_vs_machine.
Read the full abstract ▾
All authors Esteban Garces Arias, Nurzhan Sapargali, Christian Heumann, Matthias Aßenmacher