Deploying language models on resource-constrained devices, such as mobile phones, wearables, and on-device AI assistants, demands compact, efficient models without sacrificing performance. Compressing Small Language Models (SLMs) is particularly suited for these scenarios, yet their compression dynamics remain underexplored compared to Large Language Models (LLMs). We systematically evaluate leading post-training pruning (SparseGPT, Wanda) and quantization (GPTQ, AWQ) methods across six SLMs from 0.5 to 3.8B, seven languages, and seven downstream tasks. Our results show that quantization consistently outperforms pruning in preserving model fidelity, multilingual perplexity, and reasoning accuracy. However, quantization’s advantages diminish on complex knowledge and reasoning tasks like OpenBookQA, highlighting a disconnect between compression fidelity and downstream task performance. Notably, trends observed in LLMs (e.g., Wanda’s competitive performance to SparseGPT) do not generalize to SLMs. For practitioners, we recommend prioritizing quantization (particularly AWQ) for SLM compression and caution against relying on a single metric.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Large language models (LLMs) demonstrate strong task-specific capabilities through fine-tuning, but merging multiple fine-tuned models often leads to degraded performance due to overlapping instruction-following components. Task Arithmetic (TA), which combines task vectors derived from fine-tuning, enables multi-task learning and task forgetting but struggles to isolate task-specific knowledge from general instruction-following behavior. To address this, we propose Layer-Aware Task Arithmetic (LATA), a novel approach that assigns layer-specific weights to task vectors based on their alignment with instruction-following or task-specific components. By amplifying task-relevant layers and attenuating instruction-following layers, LATA improves task learning and forgetting performance while preserving overall model utility. Experiments on multiple benchmarks, including WikiText-2, GSM8K, and HumanEval, demonstrate that LATA outperforms existing methods in both multi-task learning and selective task forgetting, achieving higher task accuracy and alignment with minimal degradation in output quality. Our findings highlight the importance of layer-wise analysis in disentangling task-specific and general-purpose knowledge, offering a robust framework for efficient model merging and editing.
Curr-ReFT: Overcoming Training Bottlenecks in Small-scale Vision-Language Models via Curriculum Reinforcement Finetuning
PDF ↗State-of-the-art vision-language models (VLMs) require massive scaling that limits practical deployment. Small-scale VLMs offer a practical alternative but face out-of-domain (OOD) collapse when trained with traditional supervised fine-tuning (SFT). Through GeneralPoints experiments, we identify that OOD collapse is due to SFT’s tendency to induce visual hallucinations under distribution shifts, whereas Reinforcement Learning’s (RL) bidirectional reward-driven mechanism with iterative error correction refines visual perception. Although RL-based post-training effectively mitigates OOD degradation, it faces a critical sparse reward dilemma in complex visual reasoning tasks. To this end, we propose Curriculum Reinforcement Finetuning (Curr-ReFT), comprising two sequential stages: (1) Structured Curriculum Reinforcement Learning, which progressively evolves task formats and reward functions to match models’ growing capabilities; and (2) Rejected Sampling-based Self-improvement, which maintains the fundamental capabilities of VLMs through selective learning from high-quality examples. Extensive experiments demonstrate that Curr-ReFT achieves state-of-the-art performance across various visual tasks in both in- and out-of-domain settings and benchmarks.
Aligning small language models with human preferences is challenging, as weak policies struggle to generate informative on-policy samples and suffer from unstable gradients when trained on off-policy signals from stronger models. In this work, we propose ReAlign, a training framework that combines the stability of on-policy learning with the guidance of reviser-assisted supervision. In the ReAlign, we first train a lightweight reviser to improve policy-generated responses using preference-based supervision, conditioned on both the prompt and the initial output. And then, the policy is optimized using a combination of standard on-policy preference pairs and reviser-enhanced pairs constructed as a structured revision task, where the latter provide richer, more learnable feedback. Experimental results on AlpacaEval-2 and Arena-Hard demonstrate that ReAlign significantly boosts alignment performance for weak policies, outperforming strong preference optimization baselines.
Sentence embeddings are central to modern natural language processing, powering tasks such as clustering, semantic search, and retrieval-augmented generation. Yet, they remain largely opaque: their internal features are not directly interpretable, and users lack fine-grained control for downstream tasks. To address this issue, we introduce a formal framework to characterize the organization of features in sentence embeddings through information-theoretic means. Building on this foundation, we develop a method to identify interpretable feature directions and show how they can be composed to capture richer semantic structures. Experiments on both synthetic and real-world datasets confirm the presence of this semantic geometry and highlight the utility of our approach for enhancing interpretability and fine-grained control in sentence embeddings.
Large language models are increasingly expected to adapt to individual users, reflecting differences in preferences, values, and communication styles. To evaluate whether models can serve diverse populations, we introduce MTPA, a benchmark that leverages large-scale survey data (WVS, EVS, GSS) to construct real, hyper-granular personas spanning demographics, beliefs, and values. Unlike prior benchmarks that rely on synthetic profiles or narrow trait prediction, MTPA conditions models on real personas and systematically tests their behavior across core alignment tasks. We show that persona conditioning exposes pluralistic misalignment: while aggregate metrics suggest models are truthful and safe, subgroup-specific evaluations reveal hidden pockets of degraded factuality, fairness disparities, and inconsistent value alignment. Alongside the benchmark, we release a dataset, toolkit, and baseline evaluations. MTPA is designed with extensibility and sustainability in mind: as the underlying survey datasets are regularly updated, MTPA supports regular integration of new populations and user traits.
Exploring and Evaluating Multimodal Knowledge Reasoning Consistency of Multimodal Large Language Models
PDF ↗In recent years, multimodal large language models (MLLMs) have achieved significant breakthroughs, enhancing understanding across text and vision. However, current MLLMs still face challenges in effectively integrating knowledge across these modalities during multimodal knowledge reasoning, leading to inconsistencies in reasoning outcomes. To systematically explore this issue, we propose four evaluation tasks and construct a new dataset. We conduct a series of experiments on this dataset to analyze and compare the extent of consistency degradation in multimodal knowledge reasoning within MLLMs. Based on the experimental results, we identify factors contributing to the observed degradation in consistency. Our research provides new insights into the challenges of multimodal knowledge reasoning and offers valuable guidance for future efforts aimed at improving MLLMs.
LongTableBench: Benchmarking Long-Context Table Reasoning across Real-World Formats and Domains
PDF ↗We introduce LongTableBench, a benchmark for evaluating long-context reasoning over semi-structured tables across diverse formats, tasks, and domains. It comprises 5,950 QA instances spanning 7 table formats (e.g., Markdown, HTML, SQL), 18 domains, and input lengths up to 128K tokens, including multi-turn and multi-table settings. To ensure data quality, we combine symbolic supervision, cross-model validation, and human review. Evaluating 52 LLMs—including general-purpose, table-specific, and reasoning-enhanced models—reveals that only the strongest models maintain robust performance under increasing context lengths and format diversity. We further show that end-to-end models outperform compression-based approaches, especially on tasks requiring semantic integration. LongTableBench provides a rigorous, scalable testbed for advancing long-context tabular understanding and highlights key limitations in current LLMs’ structural and reasoning capabilities. The code and data are available at https://github.com/liyaooi/LongTableBench.
Temporal question answering is an established method for evaluating temporal reasoning in large language models. Expected answers are often numeric (e.g., dates or durations), yet model responses are evaluated like regular text with exact match (EM), unable to distinguish small from large errors. In this investigative work, we frame temporal question answering as a numerical estimation task to assess the shortcomings of EM. We introduce TempAnswerQA, a benchmark distilled from Test of Time and TempTabQA, where all questions require a numerical, temporal answer, allowing us to evaluate models beyond EM. We use the forecasting metrics symmetric mean absolute percentage error (sMAPE) and mean absolute scaled error (MASE). With sMAPE, we find that error size and EM are decoupled. Models with low EM still have low sMAPE (both 20%), and some models have high sMAPE despite high EM. Scaling errors by the deviation of the ground truth data with MASE reshuffles model rankings compared to EM, revealing gaps in models’ understanding of temporal domain knowledge, especially when trained with synthetic data. Lastly, the models’ most frequent error is to deviate by only \pm1 from the ground truth. sMAPE and MASE, unlike EM, adequately weight these errors. Our findings underscore the need for specialised metrics for temporal QA tasks. Our code and data are available on https://github.com/aauss/temporal-answer-qa.
AgentInit: Initializing LLM-based Multi-Agent Systems via Diversity and Expertise Orchestration for Effective and Efficient Collaboration
PDF ↗Proper initialization is crucial for any system, particularly in multi-agent systems (MAS), where it plays a pivotal role in determining both the system’s efficiency and effectiveness. However, existing MAS initialization methods do not fully account for the collaborative needs of the generated agents in subsequent stages. Inspired by the principles of effective team composition, we propose , which aims to optimize the structure of agent teams. Specifically, in addition to multi-round interactions and reflections between agents during agent generation, AgentInit incorporates a Natural Language to Format mechanism to ensure consistency and standardization. Balanced team selection strategies using Pareto principles are subsequently applied to jointly consider agent team diversity and task relevance to promote effective and efficient collaboration and enhance overall system performance. Experiments show that AgentInit consistently outperforms state-of-the-art initialization methods and pre-defined strategies across various frameworks and tasks, achieving an overall performance improvement of up to 1.2 and 1.7, respectively, while also significantly reducing token consumption. Further analysis confirms its strong transferability to similar tasks and verifies the effectiveness of its key components, demonstrating its capability and adaptability as a reliable MAS initialization method. Source code and models are available at https://github.com/1737423697/AgentInit.
Recently, Vision-Language-Action (VLA) models have demonstrated strong performance on a range of robotic tasks. These models rely on multimodal inputs, with language instructions playing a crucial role-not only in predicting actions, but also in robustly interpreting user intent, even when the requests are impossible to fulfill. In this work, we investigate how VLAs can recognize, interpret, and respond to false-premise instructions-natural language commands that reference objects or conditions absent from the environment. We propose — Instruct-Verify-and-Act (IVA) — a unified framework that (i) detects when an instruction cannot be executed due to a false premise, (ii) engages in language-based clarification or correction, and (iii) grounds plausible alternatives in perception and action. Towards this end, we construct a large-scale instruction tuning setup with structured language prompts and train a VLA model capable of handling both accurate and erroneous requests. Our approach leverages a contextually augmented, semi-synthetic dataset containing paired positive and false-premise instructions, enabling robust detection and natural language correction. Our experiments show that IVA can improves false premise detection accuracy by 58.89% over baselines, while increasing successful responses in false-premise scenarios by 27.89%.
As large language models (LLMs) are increasingly integrated into real-world applications, ensuring their safety and robustness is critical. Automated red-teaming methods generate adversarial attacks to identify vulnerabilities, but existing approaches often face challenges like slow performance, limited categorical diversity, and high resource demands. We propose Ferret, a novel method that enhances the baseline, Rainbow Teaming by generating multiple adversarial prompt mutations per iteration and ranking them using scoring functions such as reward models, Llama Guard, and LLM-as-a-judge. Ferret achieves a 95% attack success rate (ASR), a 46% improvement over baseline, and reduces time to a 90% ASR by 15.2%. Additionally, it generates transferable adversarial prompts effective on larger LLMs. Our code is available at https://github.com/declare-lab/ferret
Cardiovascular diseases are a leading cause of death and disability worldwide. Electrocardiogram (ECG) is critical for diagnosing and monitoring cardiac health, but obtaining large-scale annotated ECG datasets is labor-intensive and time-consuming. Recent ECG Self-Supervised Learning (eSSL) methods mitigate this by learning features without extensive labels but fail to capture fine-grained clinical semantics and require extensive task-specific fine-tuning. To address these challenges, we propose \textbf{SuPreME}, a \textbf{Su}pervised \textbf{Pre}-training framework for \textbf{M}ultimodal \textbf{E}CG representation learning. SuPreME is pre-trained using structured diagnostic labels derived from ECG report entities through a one-time offline extraction with Large Language Models (LLMs), which help denoise, standardize cardiac concepts, and improve clinical representation learning. By fusing ECG signals with textual cardiac queries instead of fixed labels, SuPreME enables zero-shot classification of unseen conditions without further fine-tuning. We evaluate SuPreME on six downstream datasets covering 106 cardiac conditions, achieving superior zero-shot AUC performance of 77.20%, surpassing state-of-the-art eSSLs by 4.98%. Results demonstrate SuPreME’s effectiveness in leveraging structured, clinically relevant knowledge for high-quality ECG representations.
The primary goal of traditional federated learning is to protect data privacy by enabling distributed edge devices to collaboratively train a shared global model while keeping raw data decentralized at local clients. The rise of large language models (LLMs) has introduced new challenges in distributed systems, as their substantial computational requirements and the need for specialized expertise raise critical concerns about protecting intellectual property (IP). This highlights the need for a federated learning approach that can safeguard both sensitive data and proprietary models. To tackle this challenge, we propose FedQSN, a federated learning approach that leverages random masking to obscure a subnetwork of model parameters and applies quantization to the remaining parameters. Consequently, the server transmits only a privacy-preserving proxy of the global model to clients during each communication round, thus enhancing the model’s confidentiality. Experimental results across various models and tasks demonstrate that our approach not only maintains strong model performance in federated learning settings but also achieves enhanced protection of model parameters compared to baseline methods.
Research is a fundamental process driving the advancement of human civilization, yet it demands substantial time and effort from researchers. In recent years, the rapid development of artificial intelligence (AI) technologies has inspired researchers to explore how AI can accelerate and enhance research. To monitor relevant advancements, this paper presents a systematic review of the progress in this domain. Specifically, we organize the relevant studies into three main categories: hypothesis formulation, hypothesis validation, and manuscript publication. Hypothesis formulation involves knowledge synthesis and hypothesis generation. Hypothesis validation includes the verification of scientific claims, theorem proving, and experiment validation. Manuscript publication encompasses manuscript writing and the peer review process. Furthermore, we identify and discuss the current challenges faced in these areas, as well as potential future directions for research. Finally, we also offer a comprehensive overview of existing benchmarks and tools across various domains that support the integration of AI into the research process. We hope this paper serves as an introduction for beginners and fosters future research.
Current LLMs are trained to refuse potentially harmful input queries regardless of whether users actually had harmful intents, causing a tradeoff between safety and user experience. Through a study of 480 participants evaluating 3,840 query-response pairs, we examine how different refusal strategies affect user perceptions across varying motivations. Our findings reveal that response strategy largely shapes user experience, while actual user motivation has negligible impact. Partial compliance—providing general information without actionable details—emerges as the optimal strategy, reducing negative user perceptions by over 50% to flat-out refusals. Complementing this, we analyze response patterns of 9 state-of-the-art LLMs and evaluate how 6 reward models score different refusal strategies, demonstrating that models rarely deploy partial compliance naturally and reward models currently undervalue it. This work demonstrates that effective guardrails require focusing on crafting thoughtful refusals rather than detecting intent, offering a path toward AI safety mechanisms that ensure both safety and sustained user engagement.
Teaching According to Talents! Instruction Tuning LLMs with Competence-Aware Curriculum Learning
PDF ↗Efficient instruction tuning aims to enhance the ultimate performance of large language models (LLMs) trained on a given instruction dataset. Curriculum learning as a typical data organization strategy has shown preliminary effectiveness in instruction tuning. However, current curriculum tuning methods suffer from the curriculum rigidity, since they rely solely on static heuristic difficulty metrics. These methods fail to adapt to the evolving capabilities of models during training, resulting in a fixed and potentially sub-optimal learning trajectory. To address the issue, **C**ompetence-**A**ware **M**ulti-**P**erspective c**U**rriculum in**S**truction tuning framework termed **CAMPUS** is proposed. CAMPUS offers several advantages: (1) Dynamic selection for sub-curriculum. (2) Competency-aware adjustment to the curriculum schedule. (3) Multiple difficulty-based scheduling. Extensive experiments prove the superior performance of CAMPUS, compared to other state-of-the-art baselines for efficient instruction tuning.
Instruction tuning of large language models (LLMs) benefits more from a handful of high-quality examples than from hordes of low-quality ones. Existing selection methods typically rely on static, heuristic quality scores and are executed only once before training. Consequently, they neither adapt to the changing state of the model nor target downstream objectives, leaving substantial room for optimization. We propose RAISE (**R**einforced **A**daptive **I**nstruction **SE**lection), a *dynamic*, *task-driven* framework that integrates selection into every training step. At each step, RAISE estimates the expected contribution of each candidate instruction to task performance and admits only the most helpful. By modeling this process as sequential decision making, we optimize the selector with reinforcement learning, yielding an interpretable policy specialized for the target task. Extensive experiments show that RAISE reaches comparable or better results than full-data training while updating only 1% of the steps, demonstrating both high efficacy and significant computational savings.
Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning, positioning them as promising tools for supporting human problem-solving. However, what happens when their performance is affected by *misinformation*, i.e., incorrect inputs introduced by users due to oversights or gaps in knowledge? Such misinformation is prevalent in real-world interactions with LLMs, yet how it propagates within LLMs’ reasoning process remains underexplored. Focusing on mathematical reasoning, we present a comprehensive analysis of how misinformation affects intermediate reasoning steps and final answers. We also examine how effectively LLMs can correct misinformation when explicitly instructed to do so. Even with explicit instructions, LLMs succeed less than half the time in rectifyingmisinformation, despite possessing correct internal knowledge, leading to significant accuracy drops (10.02% – 72.20%), and the degradation holds with thinking models (4.30% – 19.97%). Further analysis shows that applying factual corrections early in the reasoning process most effectively reduces misinformation propagation, and fine-tuning on synthesized data with early-stage corrections significantly improves reasoning factuality. Our work offers a practical approach to mitigating misinformation propagation.
Hallucination detection remains a fundamental challenge for the safe and reliable deployment of large language models (LLMs), especially in applications requiring factual accuracy. Existing hallucination benchmarks often operate at the sequence level and are limited to English, lacking the fine-grained, multilingual supervision needed for comprehensive evaluation. In this work, we introduce PsiloQA, a large-scale, multilingual dataset annotated with span-level hallucinations across 14 languages. PsiloQA is constructed through an automated three-stage pipeline: generating question–answer pairs from Wikipedia using GPT-4o, eliciting potentially hallucinated answers from diverse LLMs in a no-context setting, and automatically annotating hallucinated spans using GPT-4o by comparing against golden answers and retrieved context. We evaluate a wide range of hallucination detection methods-including uncertainty quantification, LLM-based tagging, and fine-tuned encoder models-and show that encoder-based models achieve the strongest performance across languages. Furthermore, PsiloQA demonstrates effective cross-lingual generalization and supports robust knowledge transfer to other benchmarks, all while being significantly more cost-efficient than human-annotated datasets. Our dataset and results advance the development of scalable, fine-grained hallucination detection in multilingual settings.