Peer Revenue Past citations are no guarantee of future returns

Papers

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

Computer Science Computation and Language

cs.CL Watch Computation and Language

Wednesday 4 February

Computation and Language · Cross-listed from cs.LG

Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning

Yu-Ang Lee, Ching-Yun Ko, Pin-Yu Chen et al. (+1 more)

Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. Building on this paradigm, recent studies have proposed alternative initialization strategies, architectural modifications, and optimization adjustments, reporting substantial improvements over vanilla LoRA. However, these gains are often demonstrated under fixed or narrowly tuned hyperparameter settings, despite the known sensitivity of neural networks to training configurations. In this work, we systematically re-evaluate nine representative LoRA variants alongside vanilla LoRA through extensive hyperparameter searches over learning rate, batch size, rank, and training duration. Across tasks spanning mathematical reasoning, commonsense reasoning, code generation, and instruction following at diverse model scales, we find that different LoRA methods favor distinct learning rate ranges. Crucially, once learning rates are properly tuned, all methods achieve similar peak performance (within 1-2%), with only subtle rank-dependent behaviors. These results suggest that vanilla LoRA remains a competitive baseline and that improvements reported under a single training configuration may not reflect consistent methodological advantages. Finally, a second-order analysis attributes the differing optimal learning rate ranges to variations in the largest Hessian eigenvalue, aligning with classical learning theories.
Read the full abstract

All authors Yu-Ang Lee, Ching-Yun Ko, Pin-Yu Chen, Mi-Yen Yeh

Computation and Language

BioACE: An Automated Framework for Biomedical Answer and Citation Evaluations

Deepak Gupta, Davis Bartels and Dina Demner-Fushman

With the increasing use of large language models (LLMs) for generating answers to biomedical questions, it is crucial to evaluate the quality of the generated answers and the references provided to support the facts in the generated answers. Evaluation of text generated by LLMs remains a challenge for question answering, retrieval-augmented generation (RAG), summarization, and many other natural language processing tasks in the biomedical domain, due to the requirements of expert assessment to verify consistency with the scientific literature and complex medical terminology. In this work, we propose BioACE, an automated framework for evaluating biomedical answers and citations against the facts stated in the answers. The proposed BioACE framework considers multiple aspects, including completeness, correctness, precision, and recall, in relation to the ground-truth nuggets for answer evaluation. We developed automated approaches to evaluate each of the aforementioned aspects and performed extensive experiments to assess and analyze their correlation with human evaluations. In addition, we considered multiple existing approaches, such as natural language inference (NLI) and pre-trained language models and LLMs, to evaluate the quality of evidence provided to support the generated answers in the form of citations into biomedical literature. With the detailed experiments and analysis, we provide the best approaches for biomedical answer and citation evaluation as a part of BioACE (https://github.com/deepaknlp/BioACE) evaluation package.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

Linear Model Merging Unlocks Simple and Scalable Multimodal Data Mixture Optimization

Davide Berasi, Matteo Farina, Massimiliano Mancini et al. (+1 more)

Selecting the best data mixture is critical for successful Supervised Fine-Tuning (SFT) of Multimodal Large Language Models. However, determining the optimal mixture weights across multiple domain-specific datasets remains a significant bottleneck due to the combinatorial search space and the high cost associated with even a single training run. This is the so-called Data Mixture Optimization (DMO) problem. On the other hand, model merging unifies domain-specific experts through parameter interpolation. This strategy is efficient, as it only requires a single training run per domain, yet oftentimes leads to suboptimal models. In this work, we take the best of both worlds, studying model merging as an efficient strategy for estimating the performance of different data mixtures. We train domain-specific multimodal experts and evaluate their weighted parameter-space combinations to estimate the efficacy of corresponding data mixtures. We conduct extensive experiments on 14 multimodal benchmarks, and empirically demonstrate that the merged proxy models exhibit a high rank correlation with models trained on actual data mixtures. This decouples the search for optimal mixtures from the resource-intensive training process, thereby providing a scalable and efficient strategy for navigating the complex landscape of mixture weights. Code is publicly available at https://github.com/BerasiDavide/mLLMs_merging_4_DMO.
Read the full abstract

All authors Davide Berasi, Matteo Farina, Massimiliano Mancini, Elisa Ricci

Computation and Language · Cross-listed from cs.DB

AutoPrunedRetriever: Persistent Reasoning Graphs for Efficient and Extensible Retrieval-Augmented Generation

Ning Wang, Kuanyan Zhu, Daniel Yuehwoon Yee et al. (+4 more)

Retrieval-augmented generation (RAG) has become a standard approach for knowledge-intensive LLM tasks. Complex reasoning remains challenging because queries often require composing dispersed evidence through multiple dependent reasoning steps. Yet existing RAG systems process queries independently, repeatedly retrieving overlapping evidence and reconstructing established reasoning. Preserving such reasoning makes previously retrieved evidence and intermediate results reusable for subsequent queries. However, inference-time reuse becomes increasingly costly as the history grows, motivating a compact representation that preserves useful information without substantial loss in quality. To address these challenges, we present AutoPrunedRetriever (APR), a graph-based RAG system that encodes corpus facts and query-derived reasoning in a shared symbolic codebook. The codebook serves both as the retrieval index and as persistent memory, representing facts, questions, and answers as compact edge sequences, so that the previously established structure can be directly retrieved and reused. At inference time, APR retrieves query-specific reasoning paths and constructs compact prompts, while incremental consolidation keeps the growing memory compact. For deployments where the retriever cannot be replaced, we introduce AutoPruned Layer (APL), which leaves host retrieval unchanged but replaces the retrieved context with the corresponding entries in the shared symbolic codebook, writing each question and answer back as a compact structure. Across five complex-reasoning benchmarks, APR improves accuracy over the strongest baseline by 1.8-10.8 points while using up to 96% fewer input tokens. APL reduces generator input by 60-96% across host systems and preserves answer quality on document-style contexts.
Read the full abstract

All authors Ning Wang, Kuanyan Zhu, Daniel Yuehwoon Yee, Yitang Gao, Shiying Huang, Zirun Xu, Sainyam Galhotra

Computation and Language · Cross-listed from cs.LG

Internalizing LLM Reasoning via Discovery and Replay of Latent Actions

Zhenning Shi, Yijia Zhu, Junhan Shi et al. (+3 more)

The internalization of chain-of-thought processes into hidden states has emerged as a highly efficient paradigm for scaling test-time compute. However, existing activation steering methods rely on static control vectors that fail to adapt to the non-stationary evolution of complex reasoning tasks. To address this limitation, we propose STIR (Self-Distilled Tools for Internal Reasoning), a framework that reformulates reasoning enhancement as a dynamic latent trajectory control problem. STIR introduces a synergistic three-stage pipeline: (1) differential intrinsic action induction harvests latent reasoning successes to crystallize steering primitives; (2) sparse control basis construction curates a compact, geometrically diverse tool library; and (3) value-modulated trajectory intervention dynamically injects context-specific impulses via anchor-based gating. Extensive experiments on six arithmetic and logical benchmarks across four representative models demonstrate that STIR improves average accuracy by 1.9% to 7.5% while reducing average token consumption by up to 35% compared to vanilla decoding. These findings demonstrate that the benefits of explicit chain-of-thought can be realized through dynamic latent trajectory control, internalizing the reasoning process to bypass the explicit generation while achieving superior fidelity. Our code is available at https://github.com/sznnzs/LLM-Latent-Action.
Read the full abstract

All authors Zhenning Shi, Yijia Zhu, Junhan Shi, Xun Zhang, Lei Wang, Congcong Miao

Computation and Language · Cross-listed from cs.LG

Simulated Adoption: Decoupling Magnitude and Direction in LLM In-Context Conflict Resolution

Long Zhang and Fangwei Lin

Large Language Models (LLMs) frequently prioritize conflicting in-context information over pre-existing parametric memory, a phenomenon often termed sycophancy or compliance. However, the mechanistic realization of this behavior remains obscure, specifically how the model resolves these knowledge conflicts through compliance, and whether this suppression arises from signal magnitude dilution or directional geometric alteration within the residual stream. To resolve this, we conducted a layer-wise geometric analysis across Qwen-3-4B, Llama-3.1-8B, and GLM-4-9B, decomposing the residual stream updates induced by counter-factual contexts into radial (norm-based) and angular (cosine-based) components. Our empirical results reject the universality of the "Manifold Dilution" hypothesis, as two of the three architectures maintained stable residual norms despite exhibiting significant performance degradation on factual queries. Instead, we observed that compliance is consistently characterized by "Orthogonal Interference," where the conflicting context injects a steering vector that is quasi-orthogonal to the ground-truth direction, effectively rotating the hidden state representation. This suggests that models do not "unlearn" or suppress the magnitude of internal truths but rather employ a mechanism of geometric displacement to bypass the correct unembedding vector, effectively simulating adoption while preserving the original structural magnitude. These findings challenge scalar confidence metrics for detecting hallucinations and underscore the necessity of vectorial monitoring to distinguish between genuine knowledge integration and superficial in-context mimicry.
Read the full abstract

Computation and Language · Cross-listed from q-bio.QM

AFD-INSTRUCTION: A Comprehensive Antibody Instruction Dataset with Functional Annotations for LLM-Based Understanding and Design

Ling Luo, Wenbin Jiang, Hongyuan Chang et al. (+5 more)

Large language models (LLMs) have significantly advanced protein representation learning. However, their capacity to interpret and design antibodies through natural language remains limited. To address this challenge, we present AFD-Instruction, the first large-scale instruction dataset with functional annotations tailored to antibodies. This dataset encompasses two key components: antibody understanding, which infers functional attributes directly from sequences, and antibody design, which enables de novo sequence generation under functional constraints. These components provide explicit sequence-function alignment and support antibody design guided by natural language instructions. Extensive instruction-tuning experiments on general-purpose LLMs demonstrate that AFD-Instruction consistently improves performance across diverse antibody-related tasks. By linking antibody sequences with textual descriptions of function, AFD-Instruction establishes a new foundation for advancing antibody modeling and accelerating therapeutic discovery.
Read the full abstract

All authors Ling Luo, Wenbin Jiang, Hongyuan Chang, Xinkang Wang, Xushi Zhang, Yueting Xiong, Mengsha Tong, Rongshan Yu

Computation and Language · Cross-listed from cs.IR

Atomic Information Flow: A Network Flow Model for Tool Attributions in RAG Systems

James Gao, Josh Zhou, Qi Sun et al. (+2 more)

Many tool-based Retrieval Augmented Generation (RAG) systems lack precise mechanisms for tracing final responses back to specific tool components – a critical gap as systems scale to complex multi-agent architectures. We present Atomic Information Flow (AIF), a graph-based network flow model that decomposes tool outputs and LLM calls into atoms: indivisible, self-contained units of information. By modeling LLM orchestration as a directed flow of atoms from tool and LLM nodes to a response super-sink, AIF enables granular attribution metrics for AI explainability. Motivated by the max-flow min-cut theorem in network flow theory, we train a lightweight Gemma3 (4B parameter) language model as a context compressor to approximate the minimum cut of tool atoms using flow signals computed offline by AIF. We note that the base Gemma3-4B model struggles to identify critical information with 54.7% accuracy on HotpotQA, barely outperforming lexical baselines (BM25). However, post-training on AIF signals boosts accuracy to 82.71% (+28.01 points) while achieving 87.52% (+1.85%) context token compression – bridging the gap with the Gemma3-27B variant, a model nearly $7\times$ larger.
Read the full abstract

All authors James Gao, Josh Zhou, Qi Sun, Ryan Huang, Steven Yoo

Computation and Language

Reinforced Attention Learning

Bangzheng Li, Jianmo Ni, Chen Qu et al. (+5 more)

Post-training with Reinforcement Learning (RL) has substantially improved reasoning in Large Language Models (LLMs) via test-time scaling. However, extending this paradigm to Multimodal LLMs (MLLMs) through verbose rationales yields limited gains for perception and can even degrade performance. We propose Reinforced Attention Learning (RAL), a policy-gradient framework that directly optimizes internal attention distributions rather than output token sequences. By shifting optimization from what to generate to where to attend, RAL promotes effective information allocation and improved grounding in complex multimodal inputs. Experiments across diverse image and video benchmarks show consistent gains over GRPO and other baselines. We further introduce On-Policy Attention Distillation, demonstrating that transferring latent attention behaviors yields stronger cross-modal alignment than standard knowledge distillation. Our results position attention policies as a principled and general alternative for multimodal post-training.
Read the full abstract

All authors Bangzheng Li, Jianmo Ni, Chen Qu, Ian Miao, Liu Yang, Xingyu Fu, Muhao Chen, Derek Zhiyuan Cheng

Computation and Language · Cross-listed from cs.LG

Rethinking the Trust Region in LLM Reinforcement Learning

Penghui Qi, Xiangxin Zhou, Zichen Liu et al. (+4 more)

Reinforcement learning (RL) has become a cornerstone for fine-tuning Large Language Models (LLMs), with Proximal Policy Optimization (PPO) serving as the de facto standard algorithm. Despite its ubiquity, we argue that the core ratio clipping mechanism in PPO is structurally ill-suited for the large vocabularies inherent to LLMs. PPO constrains policy updates based on the probability ratio of sampled tokens, which serves as a noisy single-sample Monte Carlo estimate of the true policy divergence. This creates a sub-optimal learning dynamic: updates to low-probability tokens are aggressively over-penalized, while potentially catastrophic shifts in high-probability tokens are under-constrained, leading to training inefficiency and instability. To address this, we propose Divergence Proximal Policy Optimization (DPPO), which substitutes heuristic clipping with a more principled constraint based on a direct estimate of policy divergence (e.g., Total Variation or KL). To avoid huge memory footprint, we introduce the efficient Binary and Top-K approximations to capture the essential divergence with negligible overhead. Extensive empirical evaluations demonstrate that DPPO achieves superior training stability and efficiency compared to existing methods, offering a more robust foundation for RL-based LLM fine-tuning. Our code is available at https://github.com/sail-sg/Stable-RL.
Read the full abstract

All authors Penghui Qi, Xiangxin Zhou, Zichen Liu, Tianyu Pang, Chao Du, Min Lin, Wee Sun Lee

Computation and Language · Cross-listed from cs.LG

Subliminal Effects in Your Data: A General Mechanism via Log-Linearity

Ishaq Aden-Ali, Noah Golowich, Allen Liu et al. (+3 more)

Training modern large language models (LLMs) has become a veritable smorgasbord of algorithms and datasets designed to elicit particular behaviors, making it critical to develop techniques to understand the effects of datasets on the model's properties. This is exacerbated by recent experiments that show datasets can transmit signals that are not directly observable from individual datapoints, posing a conceptual challenge for dataset-centric understandings of LLM training and suggesting a missing fundamental account of such phenomena. Towards understanding such effects, inspired by recent work on the linear structure of LLMs, we uncover a general mechanism through which hidden subtexts can arise in generic datasets. We introduce Logit-Linear-Selection (LLS), a method that prescribes how to select subsets of a generic preference dataset to elicit a wide range of hidden effects. We apply LLS to discover subsets of real-world datasets so that models trained on them exhibit behaviors ranging from having specific preferences, to responding to prompts in a different language not present in the dataset, to taking on a different persona. Crucially, the effect persists for the selected subset, across models with varying architectures, supporting its generality and universality.
Read the full abstract

All authors Ishaq Aden-Ali, Noah Golowich, Allen Liu, Abhishek Shetty, Ankur Moitra, Nika Haghtalab

Computation and Language

CoT is Not the Chain of Truth: An Empirical Internal Analysis of Reasoning LLMs for Fake News Generation

Zhao Tong, Chunlin Gong, Yiping Zhang et al. (+5 more)

From generating headlines to fabricating news, the Large Language Models (LLMs) are typically assessed by their final outputs, under the safety assumption that a refusal response signifies safe reasoning throughout the entire process. Challenging this assumption, our study reveals that during fake news generation, even when a model rejects a harmful request, its Chain-of-Thought (CoT) reasoning may still internally contain and propagate unsafe narratives. To analyze this phenomenon, we introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates the role of individual attention heads through Jacobian-based spectral metrics. Within this framework, we introduce three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond or embed deceptive reasoning patterns. Extensive experiments on multiple reasoning-oriented LLMs show that the generation risk rises significantly when the thinking mode is activated, where the critical routing decisions are concentrated in only a few contiguous mid-depth layers. By precisely identifying the attention heads responsible for this divergence, our work challenges the assumption that refusal implies safety and provides a new understanding perspective for mitigating latent reasoning risks.
Read the full abstract

All authors Zhao Tong, Chunlin Gong, Yiping Zhang, Haichao Shi, Qiang Liu, Xingcheng Xu, Shu Wu, Xiao-Yu Zhang

Computation and Language

Decomposed Prompting Does Not Fix Knowledge Gaps, But Helps Models Say "I Don't Know"

Dhruv Madhwal, Lyuxin David Zhang, Dan Roth et al. (+2 more)

Large language models often struggle to recognize their knowledge limits in closed-book question answering, leading to confident hallucinations. While decomposed prompting is typically used to improve accuracy, we investigate its impact on reliability. We evaluate three task-equivalent prompting regimes: Direct, Assistive, and Incremental, across different model scales and multi-hop QA benchmarks. We find that although accuracy gains from decomposition diminish in frontier models, disagreements between prompting regimes remain highly indicative of potential errors. Because factual knowledge is typically stable while hallucinations are stochastic, cross-regime agreement provides a precise signal of internal uncertainty. We leverage this signal to implement a training-free abstention policy that requires no retrieval or fine-tuning. Our results show that disagreement-based abstention outperforms standard uncertainty baselines as an error detector, improving both F1 and AUROC across settings. This demonstrates that decomposition-based prompting can serve as a practical diagnostic probe for model reliability in closed-book QA.
Read the full abstract

All authors Dhruv Madhwal, Lyuxin David Zhang, Dan Roth, Tomer Wolfson, Vivek Gupta

Computation and Language · Cross-listed from cs.OS

Horizon-LM: A RAM-Centric Architecture for LLM Training

Zhengqing Yuan, Lichao Sun and Yanfang Ye

The rapid growth of large language models (LLMs) has outpaced the evolution of single-GPU hardware, making model scale increasingly constrained by memory capacity rather than computation. While modern training systems extend GPU memory through distributed parallelism and offloading across CPU and storage tiers, they fundamentally retain a GPU-centric execution paradigm in which GPUs host persistent model replicas and full autograd graphs. As a result, scaling large models remains tightly coupled to multi-GPU clusters, complex distributed runtimes, and unpredictable host memory consumption, creating substantial barriers for node-scale post-training workloads such as instruction tuning, alignment, and domain adaptation. We present Horizon-LM, a memory-centric training system that redefines the roles of CPU and GPU for large-model optimization. Horizon-LM treats host memory as the authoritative parameter store and uses GPUs solely as transient compute engines through a CPU-master, GPU-template execution model. By eliminating persistent GPU-resident modules and autograd graphs, employing explicit recomputation with manual gradient propagation, and introducing a pipelined double-buffered execution engine, Horizon-LM decouples model scale from GPU count and bounds memory usage to the theoretical parameter footprint. On a single H200 GPU with 1.5\,TB host RAM, Horizon-LM reliably trains models up to 120B parameters. On a standard single A100 machine, Horizon-LM achieves up to 12.2$\times$ higher training throughput than DeepSpeed ZeRO-3 with CPU offloading while preserving numerical correctness. Across platforms and scales, Horizon-LM sustains high device utilization and predictable memory growth, demonstrating that host memory, not GPU memory, defines the true feasibility boundary for node-scale large-model training.
Read the full abstract

2602.04816 · 4 Feb · This paper contained an error in the throughput computation used in the experimental evaluation. Specifically, the TFLOPS calculation omitted the 12HL term in the training FLOPs formula, which led to systematic underestimation of the reported throughput numbers in the experimental results. We are withdrawing this version to correct the evaluation and avoid confusion for readers

Abstract PDF $MEM ×9$ALIGN ×1$EVO ×1$LANG ×1

Computation and Language

SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization

Jiarui Yuan, Tailin Jin, Weize Chen et al. (+1 more)

True self-evolution requires agents to act as lifelong learners that internalize novel experiences to solve future problems. However, rigorously measuring this foundational capability is hindered by two obstacles: the entanglement of prior knowledge, where “new” knowledge may appear in pre-training data, and the entanglement of reasoning complexity, where failures may stem from problem difficulty rather than an inability to recall learned knowledge. We introduce SE-Bench, a diagnostic environment that obfuscates the NumPy library and its API doc into a pseudo-novel package with randomized identifiers. Agents are trained to internalize this package and evaluated on simple coding tasks without access to documentation, yielding a clean setting where tasks are trivial with the new API doc but impossible for base models without it. Our investigation reveals three insights: (1) the Open-Book Paradox, where training with reference documentation inhibits retention, requiring "Closed-Book Training" to force knowledge compression into weights; (2) the RL Gap, where standard RL fails to internalize new knowledge completely due to PPO clipping and negative gradients; and (3) the viability of Self-Play for internalization, proving models can learn from self-generated, noisy tasks when coupled with SFT, but not RL. Overall, SE-Bench establishes a rigorous diagnostic platform for self-evolution with knowledge internalization. Our code and dataset can be found at https://github.com/thunlp/SE-Bench.
Read the full abstract

All authors Jiarui Yuan, Tailin Jin, Weize Chen, Zeyuan Liu

Computation and Language

OmniSIFT: Modality-Asymmetric Token Compression for Efficient Omni-modal Large Language Models

Yue Ding, Yiyan Ji, Jungang Li et al. (+12 more)

Omni-modal Large Language Models (Omni-LLMs) have demonstrated strong capabilities in audio-video understanding tasks. However, their reliance on long multimodal token sequences leads to substantial computational overhead. Despite this challenge, token compression methods designed for Omni-LLMs remain limited. To bridge this gap, we propose OmniSIFT (Omni-modal Spatio-temporal Informed Fine-grained Token compression), a modality-asymmetric token compression framework tailored for Omni-LLMs. Specifically, OmniSIFT adopts a two-stage compression strategy: (i) a spatio-temporal video pruning module that removes video redundancy arising from both intra-frame structure and inter-frame overlap, and (ii) a vision-guided audio selection module that filters audio tokens. The entire framework is optimized end-to-end via a differentiable straight-through estimator. Extensive experiments on five representative benchmarks demonstrate the efficacy and robustness of OmniSIFT. Notably, for Qwen2.5-Omni-7B, OmniSIFT introduces only 4.85M parameters while maintaining lower latency than training-free baselines such as OmniZip. With merely 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the performance of the full-token model on several tasks.
Read the full abstract

All authors Yue Ding, Yiyan Ji, Jungang Li, Xuyang Liu, Xinlong Chen, Junfei Wu, Bozhou Li, Bohan Zeng, Yang Shi, Yushuo Guan, Yuanxing Zhang, Jiaheng Liu, Qiang Liu, Pengfei Wan, Liang Wang

Computation and Language · Cross-listed from cs.SD

Speaker-Aware Simulation Improves Conversational Speech Recognition

Máté Gedeon and Péter Mihajlik

Automatic speech recognition (ASR) for conversational speech remains challenging due to the limited availability of large-scale, well-annotated multi-speaker dialogue data and the complex temporal dynamics of natural interactions. Speaker-aware simulated conversations (SASC) offer an effective data augmentation strategy by transforming single-speaker recordings into realistic multi-speaker dialogues. However, prior work has primarily focused on English data, leaving questions about the applicability to lower-resource languages. In this paper, we adapt and implement the SASC framework for Hungarian conversational ASR. We further propose C-SASC, an extended variant that incorporates pause modeling conditioned on utterance duration, enabling a more faithful representation of local temporal dependencies observed in human conversation while retaining the simplicity and efficiency of the original approach. We generate synthetic Hungarian dialogues from the BEA-Large corpus and combine them with real conversational data for ASR training. Both SASC and C-SASC are evaluated extensively under a wide range of simulation configurations, using conversational statistics derived from CallHome, BEA-Dialogue, and GRASS corpora. Experimental results show that speaker-aware conversational simulation consistently improves recognition performance over naive concatenation-based augmentation. While the additional duration conditioning in C-SASC yields modest but systematic gains–most notably in character-level error rates–its effectiveness depends on the match between source conversational statistics and the target domain. Overall, our findings confirm the robustness of speaker-aware conversational simulation for Hungarian ASR and highlight the benefits and limitations of increasingly detailed temporal modeling in synthetic dialogue generation.
Read the full abstract

Computation and Language

Beyond Many-Shot Translation: Scaling In-Context Demonstrations For Low-Resource Machine Translation

Luis Frentzen Salim, Esteban Carlin, Alexandre Morinvil et al. (+2 more)

Building machine translation (MT) systems for low-resource languages is notably difficult due to the scarcity of high-quality data. Although Large Language Models (LLMs) have improved MT system performance, adapting them to lesser-represented languages remains challenging. In-context learning (ICL) may offer novel ways to adapt LLMs for low-resource MT by conditioning models on demonstration at inference time. In this study, we explore scaling low-resource machine translation ICL beyond the few-shot setting to thousands of examples with long-context models. We scale in-context token budget to 1M tokens and compare three types of training corpora used as in-context supervision: monolingual unsupervised data, instruction-style data, and parallel data (English–target and Indonesian–target). Our experiments on Javanese and Sundanese show that gains from additional context saturate quickly and can degrade near the maximum context window, with scaling behavior strongly dependent on corpus type. Notably, some forms of monolingual supervision can be competitive with parallel data, despite the latter offering additional supervision. Overall, our results characterize the effective limits and corpus-type sensitivity of long-context ICL for low-resource MT, highlighting that larger context windows do not necessarily yield proportional quality gains.
Read the full abstract

All authors Luis Frentzen Salim, Esteban Carlin, Alexandre Morinvil, Xi Ai, Lun-Wei Ku

Computation and Language

When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?

Xinyu Zhou, Chang Jin, Carsten Eickhoff et al. (+2 more)

Large language models (LLMs) rarely admit uncertainty, often producing fluent but misleading answers, rather than abstaining (i.e., refusing to answer). This weakness is even evident in temporal question answering, where models frequently ignore time-sensitive evidence and conflate facts across different time-periods. In this paper, we present the first empirical study of training LLMs with an abstention ability while reasoning about temporal QA. Existing approaches such as calibration might be unreliable in capturing uncertainty in complex reasoning. We instead frame abstention as a teachable skill and introduce a pipeline that couples Chain-of-Thought (CoT) supervision with Reinforcement Learning (RL) guided by abstention-aware rewards. Our goal is to systematically analyze how different information types and training techniques affect temporal reasoning with abstention behavior in LLMs. Through extensive experiments studying various methods, we find that RL yields strong empirical gains on reasoning: a model initialized by Qwen2.5-1.5B-Instruct surpasses GPT-4o by $3.46\%$ and $5.80\%$ in Exact Match on TimeQA-Easy and Hard, respectively. Moreover, it improves the True Positive rate on unanswerable questions by $20\%$ over a pure supervised fine-tuned (SFT) variant. Beyond performance, our analysis shows that SFT induces overconfidence and harms reliability, while RL improves prediction accuracy but exhibits similar risks. Finally, by comparing implicit reasoning cues (e.g., original context, temporal sub-context, knowledge graphs) with explicit CoT supervision, we find that implicit information provides limited benefit for reasoning with abstention. Our study provides new insights into how abstention and reasoning can be jointly optimized, providing a foundation for building more reliable LLMs.
Read the full abstract

All authors Xinyu Zhou, Chang Jin, Carsten Eickhoff, Zhijiang Guo, Seyed Ali Bahrainian

Computation and Language

Exploiting contextual information to improve stance detection in informal political discourse with LLMs

Arman Engin Sucu, Yixiang Zhou, Mario A. Nascimento et al. (+1 more)

This study investigates the use of Large Language Models (LLMs) for political stance detection in informal online discourse, where language is often sarcastic, ambiguous, and context-dependent. We explore whether providing contextual information, specifically user profile summaries derived from historical posts, can improve classification accuracy. Using a real-world political forum dataset, we generate structured profiles that summarize users' ideological leaning, recurring topics, and linguistic patterns. We evaluate seven state-of-the-art LLMs across baseline and context-enriched setups through a comprehensive cross-model evaluation. Our findings show that contextual prompts significantly boost accuracy, with improvements ranging from +17.5% to +38.5%, achieving up to 74% accuracy that surpasses previous approaches. We also analyze how profile size and post selection strategies affect performance, showing that strategically chosen political content yields better results than larger, randomly selected contexts. These findings underscore the value of incorporating user-level context to enhance LLM performance in nuanced political classification tasks.
Read the full abstract

All authors Arman Engin Sucu, Yixiang Zhou, Mario A. Nascimento, Tony Mullen

Computation and Language · Cross-listed from cs.CY

Inference-Time Reasoning Selectively Reduces Implicit Social Bias in Large Language Models

Molly Apsel and Michael N. Jones

Drawing on constructs from psychology, prior work has identified a distinction between explicit and implicit bias in large language models (LLMs). While many LLMs undergo post-training alignment and safety procedures to avoid expressions of explicit social bias, they still exhibit significant implicit biases on indirect tasks resembling the Implicit Association Test (IAT). Recent work has further shown that inference-time reasoning can impair LLM performance on tasks that rely on implicit statistical learning. Motivated by a theoretical link between implicit associations and statistical learning in human cognition, we examine how reasoning-enabled inference affects implicit bias in LLMs. We find that enabling reasoning significantly reduces measured implicit bias on an IAT-style evaluation for some model classes across fifteen stereotype topics. This effect appears specific to social bias domains, as we observe no corresponding reduction for non-social implicit associations. As reasoning is increasingly enabled by default in deployed LLMs, these findings suggest that it can meaningfully alter fairness evaluation outcomes in some systems, while also raising questions about how alignment procedures interact with inference-time reasoning to drive variation in bias reduction across model types. More broadly, this work highlights how theory from cognitive science and psychology can complement AI evaluation research by providing methodological and interpretive frameworks that reveal new insights into model behavior.
Read the full abstract

Computation and Language

Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases

Casey Ford, Madison Van Doren and Emily Dix

Multimodal large language models (MLLMs) are increasingly deployed in real-world systems, yet their safety under adversarial prompting remains underexplored. We present a two-phase evaluation of MLLM harmlessness using a fixed benchmark of 726 adversarial prompts authored by 26 professional red teamers. Phase 1 assessed GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus; Phase 2 evaluated their successors (GPT-5, Claude Sonnet 4.5, Pixtral Large, and Qwen Omni) yielding 82,256 human harm ratings. Large, persistent differences emerged across model families: Pixtral models were consistently the most vulnerable, whereas Claude models appeared safest due to high refusal rates. Attack success rates (ASR) showed clear alignment drift: GPT and Claude models exhibited increased ASR across generations, while Pixtral and Qwen showed modest decreases. Modality effects also shifted over time: text-only prompts were more effective in Phase 1, whereas Phase 2 produced model-specific patterns, with GPT-5 and Claude 4.5 showing near-equivalent vulnerability across modalities. These findings demonstrate that MLLM harmlessness is neither uniform nor stable across updates, underscoring the need for longitudinal, multimodal benchmarks to track evolving safety behaviour.
Read the full abstract

Computation and Language · Cross-listed from cs.LG

From Data to Behavior: Predicting Unintended Model Behaviors Before Training

Mengru Wang, Zhenqian Xu, Junfeng Fang et al. (+4 more)

Large Language Models (LLMs) can acquire unintended biases from seemingly benign training data even without explicit cues or malicious content. Existing methods struggle to detect such risks before fine-tuning, making post hoc evaluation costly and inefficient. To address this challenge, we introduce Data2Behavior, a new task for predicting unintended model behaviors prior to training. We also propose Manipulating Data Features (MDF), a lightweight approach that summarizes candidate data through their mean representations and injects them into the forward pass of a base model, allowing latent statistical signals in the data to shape model activations and reveal potential biases and safety risks without updating any parameters. MDF achieves reliable prediction while consuming only about 20% of the GPU resources required for fine-tuning. Experiments on Qwen3-14B, Qwen2.5-32B-Instruct, and Gemma-3-12b-it confirm that MDF can anticipate unintended behaviors and provide insight into pre-training vulnerabilities.
Read the full abstract

All authors Mengru Wang, Zhenqian Xu, Junfeng Fang, Yunzhi Yao, Shumin Deng, Huajun Chen, Ningyu Zhang

Computation and Language

Modular Expert Merging for Biomedical Retrieval

Sameh Khattab, Jean-Philippe Corbeil, Osman Alperen Çinar-Koraş et al. (+5 more)

Adapting general-purpose LLMs into domain-specialized dense retrievers typically requires large-scale training on mixed-domain data. We show that merging independently trained domain-specialized experts consistently exceeds this approach across four decoder-only LLM families (0.6B-7B), four merging methods, and twelve medical and general retrieval tasks from MTEB, suggesting that parameter-space composition captures complementary domain strengths that large-scale mixed-domain training averages out. To further maximize expert quality, we introduce Synthesize-Train-Merge (STM), a modular framework that synthesizes hard negatives with a top-tier LLM and fine-tunes domain-specialized experts via LoRA before merging them, without continual pre-training. Synthesized hard negatives yield the largest gains for smaller models, and STM achieves strong performance on biomedical retrieval tasks while maintaining competitive general-domain results across all four backbone families.
Read the full abstract

All authors Sameh Khattab, Jean-Philippe Corbeil, Osman Alperen Çinar-Koraş, Amin Dada, Julian Friedrich, Jiawei He, Douglas Teodoro, Jens Kleesiek

Computation and Language

"Be My Cheese?": Cultural Nuance Benchmarking for Machine Translation in Multilingual LLMs

Madison Van Doren, Casey Ford, Jennifer Barajas et al. (+2 more)

We present a large-scale human evaluation benchmark for assessing cultural localisation in machine translation produced by state-of-the-art multilingual large language models (LLMs). Existing MT benchmarks emphasise token-level and grammatical accuracy, but often overlook the pragmatic and culturally grounded competencies required for real-world localisation. Building on a pilot study of 87 translations across 20 languages, we evaluate 7 multilingual LLMs across 15 target languages with 5 native-speaker raters per language. Each rater scored both full-text translations and segment-level instances of culturally nuanced language (idioms, puns, holidays, and culturally embedded concepts) on an ordinal 0-3 quality scale; segment ratings additionally included an NA option for untranslated segments. Across full-text evaluations, mean overall quality is modest (1.68/3): GPT-5 (2.10/3), Claude Sonnet 4 (1.97/3), and Mistral Medium 3.1 (1.84/3) form the strongest tier with fewer catastrophic failures. Segment-level results show sharp category effects: holidays (2.20/3) and cultural concepts (2.19/3) translate notably better than idioms (1.65/3) and puns (1.45/3), and idioms are most likely to be left untranslated. Inter-rater reliability was assessed using Krippendorff's {\alpha} and Gwet's AC2, indicating moderate agreement overall (Krippendorff's {\alpha} = 0.45) with the lowest agreement for puns. These findings demonstrate a persistent gap between grammatical adequacy and cultural resonance. To our knowledge, this is the first multilingual, human-annotated benchmark focused explicitly on cultural nuance in translation and localisation. The results highlight the need for culturally informed training data, improved cross-lingual pragmatics, and evaluation frameworks that support systematic benchmarking of culturally grounded translation.
Read the full abstract

All authors Madison Van Doren, Casey Ford, Jennifer Barajas, Riley VanMeter, Cory Holland

Computation and Language · Cross-listed from cs.LG

Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Models

Moritz Miller, Florent Draye and Bernhard Schölkopf

A central premise in mechanistic interpretability is that meaningful concepts in language models are represented by linear features in activation space. For such features to support reliable interventions, manipulating one feature should not substantially alter the effects of others. In practice, however, feature entanglement leads to interference such that localized interventions can have unintended downstream effects. Motivated by the _Independent Causal Mechanisms_ principle, we propose to constrain internal features to be almost orthogonal. We argue that this promotes modular representations amenable to causal intervention. We formalize this problem by characterizing the gap between an idealized isolated intervention and its realized effect on model outputs in terms of feature interference. We upper-bound the propagation of feature interference in terms of the self-coherence of the feature dictionary, and relate this discrepancy to an explicit orthogonality regularization on the dictionary itself. Empirically, we show that this regularization enables more isolated interventions on mathematical reasoning concepts while preserving model performance. Our code is available under https://github.com/mrtzmllr/sae-icm.
Read the full abstract

Computation and Language

Linguistically Informed Evaluation of Multilingual ASR for African Languages

Fei-Yueh Chen, Lateef Adeleke and C.M. Downey

Word Error Rate (WER) mischaracterizes ASR models' performance for African languages by combining phonological, tone, and other linguistic errors into a single lexical error. By contrast, Feature Error Rate (FER) has recently attracted attention as a viable metric that reveals linguistically meaningful errors in models' performance. In this paper, we evaluate three speech encoders on two African languages by complementing WER with CER, and FER, and add a tone-aware extension (TER). We show that by computing errors on phonological features, FER and TER reveal linguistically-salient error patterns even when word-level accuracy remains low. Our results reveal that models perform better on segmental features, while tones (especially mid and downstep) remain the most challenging features. Results on Yoruba show a striking differential in metrics, with WER=0.788, CER=0.305, and FER=0.151. Similarly for Uneme (an endangered language absent from pretraining data) a model with near-total WER and 0.461 CER achieves the relatively low FER of 0.267. This indicates model error is often attributable to individual phonetic feature errors, which is obscured by all-or-nothing metrics like WER.
Read the full abstract

Computation and Language

LiteToken: Removing Intermediate Merge Residues From BPE Tokenizers

Yike Sun, Haotong Yang, Zhouchen Lin et al. (+1 more)

Tokenization is fundamental to how language models represent and process text, yet the behavior of widely used BPE tokenizers has received far less study than model architectures and training. In this paper, we investigate intermediate merge residues in BPE vocabularies: tokens that are frequent during merge learning so that retained in the final vocabulary, but are mostly further merged and rarely emitted when tokenizing the corpus during tokenizer usage. Such low-frequency tokens not only waste vocabulary capacity but also increase vulnerability to adversarial or atypical inputs. We present a systematic empirical characterization of this phenomenon across commonly used tokenizers and introduce LiteToken, a simple method for removing residue tokens. Because the affected tokens are rarely used, pretrained models can often accommodate the modified tokenizer without additional fine-tuning. Experiments show that LiteToken reduces token fragmentation, reduces parameters, and improves robustness to noisy or misspelled inputs, while preserving overall performance.
Read the full abstract

All authors Yike Sun, Haotong Yang, Zhouchen Lin, Muhan Zhang

Computation and Language

ERNIE 5.0 Technical Report

Haifeng Wang, Hua Wu, Tian Wu et al. (+435 more)

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.
Read the full abstract

All authors Haifeng Wang, Hua Wu, Tian Wu, Yu Sun, Jing Liu, Dianhai Yu, Yanjun Ma, Jingzhou He, Zhongjun He, Dou Hong, Qiwen Liu, Shuohuan Wang, Junyuan Shang, Zhenyu Zhang, Yuchen Ding, Jinle Zeng, Jiabin Yang, Liang Shen, Ruibiao Chen, Weichong Yin, Siyu Ding, Dai Dai, Shikun Feng, Siqi Bao, Bolei He, Yan Chen, Zhenyu Jiao, Ruiqing Zhang, Zeyu Chen, Qingqing Dang, Kaipeng Deng, Jiajun Jiang, Enlei Gong, Guoxia Wang, Yanlin Sha, Yi Liu, Yehan Zheng, Weijian Xu, Jiaxiang Liu, Zengfeng Zeng, Yingqi Qu, Zhongli Li, Zhengkun Zhang, Xiyang Wang, Zixiang Xu, Xinchao Xu, Zhengjie Huang, Dong Wang, Bingjin Chen, Yue Chang, Xing Yuan, Shiwei Huang, Qiao Zhao, Xinzhe Ding, Shuangshuang Qiao, Baoshan Yang, Bihong Tang, Bin Li, Bingquan Wang, Binhan Tang, Binxiong Zheng, Bo Cui, Bo Ke, Bo Zhang, Bowen Zhang, Boyan Zhang, Boyang Liu, Caiji Zhang, Can Li, Chang Xu, Chao Pang, Chao Zhang, Chaoyi Yuan, Chen Chen, Cheng Cui, Chenlin Yin, Chun Gan, Chunguang Chai, Chuyu Fang, Cuiyun Han, Dan Zhang, Danlei Feng, Danxiang Zhu, Dong Sun, Dongbo Li, Dongdong Li, Dongdong Liu, Dongxue Liu, Fan Ding, Fan Hu, Fan Li, Fan Mo, Feisheng Wu, Fengwei Liu, Gangqiang Hu, Gaofeng Lu, Gaopeng Yong, Gexiao Tian, Guan Wang, Guangchen Ni, Guangshuo Wu, Guanzhong Wang, Guihua Liu, Guishun Li, Haibin Li, Haijian Liang, Haipeng Ming, Haisu Wang, Haiyang Lu, Haiye Lin, Han Zhou, Hangting Lou, Hanwen Du, Hanzhi Zhang, Hao Chen, Hao Du, Hao Liu, Hao Zhou, Haochen Jiang, Haodong Tian, Haoshuang Wang, Haozhe Geng, Heju Yin, Hong Chen, Hongchen Xue, Hongen Liu, Honggeng Zhang, Hongji Xu, Hongwei Chen, Hongyang Zhang, Hongyuan Zhang, Hua Lu, Huan Chen, Huan Wang, Huang He, Hui Liu, Hui Zhong, Huibin Ruan, Jiafeng Lu, Jiage Liang, Jiahao Hu, Jiahao Hu, Jiajie Yang, Jialin Li, Jian Chen, Jian Wu, Jianfeng Yang, Jianguang Jiang, Jianhua Wang, Jianye Chen, Jiaodi Liu, Jiarui Zhou, Jiawei Lv, Jiaxin Zhou, Jiaxuan Liu, Jie Han, Jie Sun, Jiefan Fang, Jihan Liu, Jihua Liu, Jing Hu, Jing Qian, Jing Yan, Jingdong Du, Jingdong Wang, Jingjing Wu, Jingyong Li, Jinheng Wang, Jinjin Li, Jinliang Lu, Jinlin Yu, Jinnan Liu, Jixiang Feng, Jiyi Huang, Jiyuan Zhang, Jun Liang, Jun Xia, Jun Yu, Junda Chen, Junhao Feng, Junhong Xiang, Junliang Li, Kai Liu, Kailun Chen, Kairan Su, Kang Hu, Kangkang Zhou, Ke Chen, Ke Wei, Kui Huang, Kun Wu, Kunbin Chen, Lei Han, Lei Sun, Lei Wen, Linghui Meng, Linhao Yu, Liping Ouyang, Liwen Zhang, Longbin Ji, Longzhi Wang, Meng Sun, Meng Tian, Mengfei Li, Mengqi Zeng, Mengyu Zhang, Ming Hong, Mingcheng Zhou, Mingming Huang, Mingxin Chen, Mingzhu Cai, Naibin Gu, Nemin Qiu, Nian Wang, Peng Qiu, Peng Zhao, Pengyu Zou, Qi Wang, Qi Xin, Qian Wang, Qiang Zhu, Qianhui Luo, Qianwei Yang, Qianyue He, Qifei Wu, Qinrui Li, Qiwen Bao, Quan Zhang, Quanxiang Liu, Qunyi Xie, Rongrui Zhan, Rufeng Dai, Rui Peng, Ruian Liu, Ruihao Xu, Ruijie Wang, Ruixi Zhang, Ruixuan Liu, Runsheng Shi, Ruting Wang, Senbo Kang, Shan Lu, Shaofei Yu, Shaotian Gong, Shenwei Hu, Shifeng Zheng, Shihao Guo, Shilong Fan, Shiqin Liu, Shiwei Gu, Shixi Zhang, Shuai Yao, Shuang Zhang, Shuangqiao Liu, Shuhao Liang, Shuwei He, Shuwen Yang, Sijun He, Siming Dai, Siming Wu, Siyi Long, Songhe Deng, Suhui Dong, Suyin Liang, Teng Hu, Tianchan Xu, Tianliang Lv, Tianmeng Yang, Tianyi Wei, Tiezhu Gao, Ting Sun, Ting Zhang, Tingdan Luo, Wei He, Wei Luan, Wei Yin, Wei Zhang, Wei Zhou, Weibao Gong, Weibin Li, Weicheng Huang, Weichong Dang, Weiguo Zhu, Weilong Zhang, Weiqi Tan, Wen Huang, Wenbin Chang, Wenjing Du, Wenlong Miao, Wenpei Luo, Wenquan Wu, Xi Shi, Xi Zhao, Xiang Gao, Xiangguo Zhang, Xiangrui Yu, Xiangsen Wang, Xiangzhe Wang, Xianlong Luo, Xianying Ma, Xiao Tan, Xiaocong Lin, Xiaofei Wang, Xiaofeng Peng, Xiaofeng Wu, Xiaojian Xu, Xiaolan Yuan, Xiaopeng Cui, Xiaotian Han, Xiaoxiong Liu, Xiaoxu Fei, Xiaoxuan Wu, Xiaoyu Wang, Xiaoyu Zhang, Xin Sun, Xin Wang, Xinhui Huang, Xinming Zhu, Xintong Yu, Xinyi Xu, Xinyu Wang, Xiuxian Li, XuanShi Zhu, Xue Xu, Xueying Lv, Xuhong Li, Xulong Wei, Xuyi Chen, Yabing Shi, Yafeng Wang, Yamei Li, Yan Liu, Yanfu Cheng, Yang Gao, Yang Liang, Yang Wang, Yang Wang, Yang Yang, Yanlong Liu, Yannian Fu, Yanpeng Wang, Yanzheng Lin, Yao Chen, Yaozong Shen, Yaqian Han, Yehua Yang, Yekun Chai, Yesong Wang, Yi Song, Yichen Zhang, Yifei Wang, Yifeng Guo, Yifeng Kou, Yilong Chen, Yilong Guo, Yiming Wang, Ying Chen, Ying Wang, Yingsheng Wu, Yingzhan Lin, Yinqi Yang, Yiran Xing, Yishu Lei, Yixiang Tu, Yiyan Chen, Yong Zhang, Yonghua Li, Yongqiang Ma, Yongxing Dai, Yongyue Zhang, Yu Ran, Yu Sun, Yu-Wen Michael Zhang, Yuang Liu, Yuanle Liu, Yuanyuan Zhou, Yubo Zhang, Yuchen Han, Yucheng Wang, Yude Gao, Yuedong Luo, Yuehu Dong, Yufeng Hu, Yuhui Cao, Yuhui Yun, Yukun Chen, Yukun Gao, Yukun Li, Yumeng Zhang, Yun Fan, Yun Ma, Yunfei Zhang, Yunshen Xie, Yuping Xu, Yuqin Zhang, Yuqing Liu, Yurui Li, Yuwen Wang, Yuxiang Lu, Zefeng Cai, Zelin Zhao, Zelun Zhang, Zenan Lin, Zezhao Dong, Zhaowu Pan, Zhaoyu Liu, Zhe Dong, Zhe Zhang, Zhen Zhang, Zhengfan Wu, Zhengrui Wei, Zhengsheng Ning, Zhenxing Li, Zhenyu Li, Zhenyu Qian, Zhenyun Li, Zhi Li, Zhichao Chen, Zhicheng Dong, Zhida Feng, Zhifan Feng, Zhihao Deng, Zhijin Yu, Zhiyang Chen, Zhonghui Zheng, Zhuangzhuang Guo, Zhujun Zhang, Zhuo Sun, Zichang Liu, Zihan Lin, Zihao Huang, Zihe Zhu, Ziheng Zhao, Ziping Chen, Zixuan Zhu, Ziyang Xu, Ziyi Liang, Ziyuan Gao

Computation and Language

LinGO: A Linguistic Graph Optimization Framework with LLMs for Interpreting Intents of Online Uncivil Discourse

Yuan Zhang and Thales Bertaglia

Detecting uncivil language is crucial for maintaining safe, inclusive, and democratic online spaces. Yet existing classifiers often misinterpret posts containing uncivil cues but expressing civil intents, leading to inflated estimates of harmful incivility online. We introduce LinGO, a linguistic graph optimization framework for large language models (LLMs) that leverages linguistic structures and optimization techniques to classify multi-class intents of incivility that use various direct and indirect expressions. LinGO decomposes language into multi-step linguistic components, identifies targeted steps that cause the most errors, and iteratively optimizes prompt and/or example components for targeted steps. We evaluate it using a dataset collected during the 2022 Brazilian presidential election, encompassing four forms of political incivility: Impoliteness (IMP), Hate Speech and Stereotyping (HSST), Physical Harm and Violent Political Rhetoric (PHAVPR), and Threats to Democratic Institutions and Values (THREAT). Each instance is annotated with six types of civil/uncivil intent. We benchmark LinGO using three cost-efficient LLMs: GPT-5-mini, Gemini 2.5 Flash-Lite, and Claude 3 Haiku, and four optimization techniques: TextGrad, AdalFlow, DSPy, and Retrieval-Augmented Generation (RAG). The results show that, across all models, LinGO consistently improves accuracy and weighted F1 compared with zero-shot, chain-of-thought, direct optimization, and fine-tuning baselines. RAG is the strongest optimization technique and, when paired with Gemini model, achieves the best overall performance. These findings demonstrate that incorporating multi-step linguistic components into LLM instructions and optimize targeted components can help the models explain complex semantic meanings, which can be extended to other complex semantic explanation tasks in the future.
Read the full abstract