Multimodal Large Language Models (MLLMs) have been increasingly used as automatic evaluators—a paradigm known as *MLLM-as-a-Judge*. However, their reliability and vulnerabilities to biases remain underexplored. We find that many MLLM judges fail to reliably integrate key visual or textual cues, yielding unreliable evaluations when evidence is missing or mismatched, and exhibiting instability under semantically irrelevant perturbations. To address this, we systematically define *Compositional Bias* in MLLM-as-a-Judge systems and introduce **MM-JudgeBias**, a benchmark for evaluating it. MM-JudgeBias introduces controlled perturbations across Query, Image, and Response, and evaluates model behavior via two complementary metrics: *Bias-Deviation (BD)* for sensitivity and *Bias-Conformity (BC)* for stability. Our dataset of over 1,800 curated and refined multimodal samples, drawn from 29 source benchmarks, enables a fine-grained diagnosis of nine bias types across diverse tasks and domains. Experiments on 26 state-of-the-art MLLMs reveal systematic modality neglect and asymmetric evaluation tendencies, underscoring the need for more reliable judges.
论文检索
输入标题、作者或关键词，从 12,319 篇学术成果中精准定位
Bargaining is often regarded as a logical arena rather than an art or a matter of intuition, yet Large Language Models (LLMs) still struggle to navigate it due to limited strategic depth and difficulty adapting to complex human factors. Current benchmarks rarely capture this limitation. To bridge this gap, we present a utility feedback centric framework. Our contributions are: (i) AgoraBench, a new benchmark spanning nine challenging settings (e.g., deception, monopoly) that supports diverse strategy modeling; (ii) human-aligned, economically grounded metrics derived from utility theory. This is operationalized via agent utility, negotiation power, and acquisition ratio that implicitly measure how well the negotiation aligns with human preference and (iii) a human preference grounded dataset with learning pipeline that strengthens LLMs’ bargaining ability through both prompting and finetuning. Empirical results indicate that baseline LLM strategies often diverge from human preferences, while our mechanism substantially improves negotiation performance, yielding deeper strategic behavior and stronger opponent awareness.
Incorporating external knowledge is crucial for knowledge-intensive tasks, such as question answering and fact checking. However, language models (LMs) may ignore relevant information that contradicts outdated parametric memory or be distracted by irrelevant contexts. While many context utilisation manipulation techniques (CMTs) have recently been proposed to alleviate these issues, few have seen systematic comparison. In this paper, we develop CUB (Context Utilisation Benchmark) - the first comprehensive benchmark designed to help diagnose CMTs under diverse noisy context conditions within retrieval-augmented generation (RAG). With this benchmark, we conduct the most extensive evaluation to date of seven state-of-the-art methods, representative of the main categories of CMTs, across three diverse datasets and tasks, applied to 11 LMs. Our findings expose critical gaps in current CMT evaluation practices, demonstrating the need for holistic testing. We reveal that most existing CMTs struggle to handle the full spectrum of context types encountered in real-world RAG scenarios. We also find that many CMTs display inflated performance on simple synthesised datasets, compared to more realistic datasets with naturally occurring samples.
GeoLaux: A Benchmark for Evaluating MLLMs’ Geometry Performance on Long-Step Problems Requiring Auxiliary Lines
PDF ↗Geometry problem solving (GPS) poses significant challenges for Multimodal Large Language Models (MLLMs) in diagram comprehension, knowledge application, long-step reasoning, and auxiliary line construction. However, current benchmarks lack fine-grained evaluation for long-step problems necessitating auxiliary construction. To address these limitations, we present GeoLaux, a fine-grained annotated dataset comprising 2186 calculation and proof problems. It features long-step reasoning (with an average solution length of 6.51 steps, maximum of 24 steps) and auxiliary line construction (required in 41.8% of problems). Building on the dataset, we conduct a comprehensive five-dimensional evaluation of 23 leading MLLMs. The evaluation yields three pivotal findings: First, models perform significantly worse on long-step problems compared to short-step ones, with 18 models exhibiting a performance drop of over 50%. Second, it is crucial to enhance models’ understanding, awareness, and proficiency in auxiliary line construction, which is vital for overall geometric reasoning. Third, limited answer hints effectively improve process correctness, whereas explicit answers lead models to neglect intermediate reasoning steps. These findings position GeoLaux both to benchmark MLLMs geometry reasoning abilities and to guide their improvement. Data and code are available at https://github.com/Candice-yu/GeoLaux
Lingua-Graph: A Unified Representation of Cross-Task Common Substructures for Analytic Language Processing
PDF ↗Structural understanding of natural language requires explicit recovery of internal meaning structures (entities, facts, nested relations), yet current structural-analytic tasks are fragmented by inconsistent task requirements across datasets. We investigate the problem of robust cross-task structural understanding under heterogeneous requirements across structural-analytic tasks and outline a perspective called Analytic NLP in which tasks can be reformulated into a representation-then-decision paradigm. In this paper, we suggest a solution for the representation layer, called Lingua-Graph, which explicitly captures entities, facts, and relations. By representing predictions as explicit graphs with labeled nodes and edges, Lingua-Graph also improves interpretability, enabling transparent inspection and error analysis of intermediate meaning structures. We construct a labeled Lingua-Graph dataset and train a baseline parser. Experiments show that Lingua-Graph provides substantially higher entity-structure hostability than alternative representations on average, and OpenIE systems based on Lingua-Graph achieve superior performance on three benchmarks, demonstrating that better intermediate structures translate into downstream gains. The data, code and the trained model are publicly released at https://github.com/rudaoshi/Lingua.
Composed Image Retrieval (CIR) is an important image retrieval paradigm that enables users to retrieve a target image using a multimodal query that consists of a reference image and modification text. Although research on CIR has made significant progress, prevailing setups still rely simple modification texts that typically cover only a limited range of salient changes, which induces two limitations highly relevant to practical applications, namely Insufficient Entity Coverage and Clause-Entity Misalignment. In order to address these issues and bring CIR closer to real-world use, we construct two instruction-rich multi-modification datasets, M-FashionIQ and M-CIRR. In addition, we propose TEMA, the Text-oriented Entity Mapping Architecture, which is the first CIR framework designed for multi-modification while also accommodating simple modifications. Extensive experiments on four benchmark datasets demonstrate that TEMA’s superiority in both original and multi-modification scenarios, while maintaining an optimal balance between retrieval accuracy and computational efficiency. Our codes and constructed multi-modification dataset (M-FashionIQ and M-CIRR) are available at https://github.com/lee-zixu/ACL26-TEMA/
Visual Language Models (VLMs) have become a robust foundation for document question answering. Processing long documents remains challenging due to limited context windows and computational budgets. Existing page-level retrieval methods offer a practical solution, typically encoding pages and queries into vectors and ranking them via cosine similarity. However, such embedding-based methods (i) lack query–page interaction before similarity scoring and (ii) usually require large-scale datasets to align visual and textual embeddings. In this paper, we observe that the cross-modal attention maps of well-trained VLMs are able to highlight semantically relevant regions. Building on this insight, we present CAPS (Cross-modal Attention as Page Selector), a retrieval framework that utilizes attention mechanisms inside VLMs for page selection. Specifically, CAPS first enhances attention-based retrieval capability with a small amount of contrastive data, then identifies the most effective attention head through expert head selection, and finally employs an adaptive filtering mechanism to obtain an appropriate number of relevant page candidates. Extensive experiments on four long-document benchmarks demonstrate that CAPS outperforms state-of-the-art embedding-based methods in both retrieval precision and downstream DocQA accuracy. Notably, CAPS achieves these gains using less than 10% of the training data required by competing baselines, highlighting the data efficiency of attention-based page retrieval.
While Large Language Models (LLMs) exhibit strong semantic capabilities, their resilience to manipulative linguistic patterns such as logical fallacies remains an underexplored area. Prior work has focused on the ability of LLMs to **identify** or **classify** fallacies, but their robustness against these fallacies in persuasive contexts remains largely unexplored.To address this gap, we introduce **LoFa** (Logical Fallacy), a comprehensive benchmark to evaluate LLM robustness against fallacies. We first construct the **LoFa** dataset via a multi-agent pipeline, pairing factual questions with fallacious arguments. Then, we develop a multi-round debate framework to assess model resilience under sustained attacks.Furthermore, to disentangle robustness from a model’s inherent knowledge limitations, we propose a new metric, LFR@k (Logical Fallacy Resistance), to quantify performance. Our experiments reveal that different LLMs exhibit varied robustness to distinct types of fallacies, highlighting unique vulnerability profiles across models.
Vision-language models are increasingly deployed as computer-use agents (CUAs) that operate desktops and browsers. Top-performing CUAs are framework-based systems that decompose planning and execution, while end-to-end screenshot-to-action policies are easier to deploy but lag behind on benchmarks such as OSWorld-Verified. GUI datasets like OSWorld pose two bottlenecks: they expose only a few hundred interactive, verifiable tasks and environments, and expert trajectories must be gathered by interacting with these environments, making such data hard to scale. We therefore ask how reinforcement learning from verifiable rewards (RLVR) can best exploit a small pool of exist expert trajectories to train end-to-end policies. Na"ively mixing these off-policy traces into on-policy RLVR is brittle: even after format conversion, expert trajectories exhibit structural mismatch and distribution shift from the learner. We propose BEPA (Bi-Level Expert-to-Policy Assimilation), which turns static expert traces into policy-aligned guidance via self-rolled reachable trajectories under the base policy (LEVEL-1) and a per-task, dynamically updated cache used in RLVR (LEVEL-2). On OSWorld-Verified, BEPA improves UITARS1.5-7B success from 22.87% to 32.13% and raises a held-out split from 5.74% to 10.30%, with consistent gains on MMBench-GUI and Online-Mind2Web. Our code and data are available at an anonymous repository: https://anonymous.4open.science/r/ACL_BEPA.
SAIR-Comb : A Structure-Aware Iterative Refinement Framework for Combinatorics Autoformalization
PDF ↗Autoformalization aims to bridge the gap between human mathematical intuition and formal proof by automating the translation of informal reasoning into machine-verifiable languages. Despite significant breakthroughs catalyzed by Large Language Models (LLMs), autoformalizing Combinatorics remains a formidable challenge due to its intricate structural dependencies and the severe scarcity of high-quality formal datasets. To address these challenges, we propose SAIR-Comb, a Structure-Aware Iterative Refinement framework for Combinatorics powered by Lean 4 and LLMs. SAIR-Comb employs a multi-stage pipeline: first, it performs data augmentation and refinement by rectifying syntactic, semantic, and structural errors, guided by a curated manual combinatorics dataset. The model then undergoes a two-stage training regime: expert iteration with syntactic grounding, followed by reinforcement learning (RL) to align formal reasoning trajectories. Furthermore, we introduce Structural Consistency—a rigorous new metric designed to expose formalizing failures that elude traditional semantic-only evaluations. Experiments demonstrate that SAIR-Comb achieves strong performance on the specialized CombiBench while remaining highly competitive on general-domain benchmarks, including PutnamBench and ProverBench.
In high-precision scenarios, vision language models suffer from Linguistic Priors Hallucination. When processing familiar text, models tend to over-rely on internal parametric knowledge, effectively "reciting" the content rather than "reading" the image. In this paper, we first systematically investigate this phenomenon by constructing the GlitchText Probing Dataset. We discover that the model’s reliance on visual grounding diminishes significantly as the generation length increases. To mitigate this, we propose PAR (Positional Perturbation and Attention Recycling), a training-free, inference-time intervention framework. PAR consists of two parts: (1) Positional Perturbation (PP) injects structured phase noise into the rotary positional embeddings; (2) Foveal Attention Recycling (FAR) detects over-confident linguistic priors and dynamically redistributes attention mass back to important visual regions. Extensive experiments across state-of-the-art models, demonstrate that PAR significantly reduces hallucination rates (reducing CER by 12%), particularly in long-context scenarios, while maintaining robust generalization on standard benchmarks.
FormalScience: Scalable Human-in-the-Loop Autoformalisation of Science with Agentic Code Generation in Lean
PDF ↗Formalising informal mathematical reasoning into formally verifiable code is a significant challenge for large language models. In scientific fields such as physics, domain-specific machinery (e.g. Dirac notation, vector calculus) imposes additional formalisation challenges that modern LLMs and agentic approaches have yet to tackle. To aid autoformalisation in scientific domains, we present FormalScience; a domain-agnostic human-in-the-loop agentic pipeline that enables a single domain expert (without deep formal language experience) to produce syntactically correct and semantically aligned formal proofs of informal reasoning for low economic cost. Applying FormalScience to physics, we construct FormalPhysics, a dataset of 200 university-level (LaTeX) physics problems and solutions (primarily quantum mechanics and electromagnetism), along with their Lean4 formal representations. Compared to existing formal math benchmarks, FormalPhysics achieves perfect formal validity and exhibits greater statement complexity. We evaluate open-source models and proprietary systems on a statement autoformalisation task on our dataset via zero-shot prompting, self-refinement with error feedback, and a novel multi-stage agentic approach, and explore autoformalisation limitations in modern LLM-based approaches. We provide the first systematic characterisation of semantic drift in physics autoformalisation in terms of concepts such as notational collapse and abstraction elevation which reveals what formal language verifies when full semantic preservation is unattainable. We release the codebase together with an interactive UI-based FormalScience system which facilitates autoformalisation and theorem proving in scientific domains beyond physics.[<https://github.com/jmeadows17/formal-science>]
Although camouflaged object segmentation has advanced rapidly in recent years, existing methods are still confined to visual mask prediction under fixed task assumptions. They cannot interactively respond to user requests, nor can they proactively understand and reason about the user’s intent. Our work tackles this issue by proposing a novel task, Language-Guided Reasoning Camouflaged Object Segmentation (LRCOS). Given a camouflaged image and an implicit query text instruction that requires reasoning, LRCOS aims to output intent-consistent segmentation mask. To establish a benchmark for this task, we build CamoQuery, comprising 12,437 image–mask samples and 25971 implicit query text instructions. To better reflect real-world camouflaged scenarios, we additionally collect MCD, a multi-instance camouflage dataset where multiple camouflaged targets co-exist within the same scene, increasing the need for reasoning. Building on CamoQuery, we further propose COSA, a vision–language segmentation assistant that segments the intended camouflaged object from implicit queries and produces a reasoning explanation. Experiments on CamoQuery demonstrate that COSA has strong reasoning segmentation capability in camouflaged scenes and exhibits zero-shot capability.
Decoder-only large language models (LLMs) have been increasingly adopted to build embedding models for diverse tasks. To overcome the inherent limitations of causal attention in representation learning, many existing methods modify the attention mechanism to be bidirectional, potentially undermining LLMs’ ability to extract semantic information acquired during pre-training. Meanwhile, leading unidirectional approaches often rely on extra input text to generate contextualized embeddings, inevitably increasing computational costs. In this work, we propose Causal2Vec, a general-purpose embedding model tailored to enhance the performance of decoder-only LLMs without altering their original architectures or introducing significant computational overhead. Specifically, we first employ a lightweight BERT-style model to pre-encode the input text into a single Contextual token, which is then prepended to the LLM’s input sequence, allowing each token to capture contextualized information even without attending to future tokens. Furthermore, to mitigate the recency bias introduced by last-token pooling, we concatenate the last hidden states of Contextual and EOS tokens as the final text embedding. In practice, Causal2Vec achieves a new state-of-the-art performance on the MTEB benchmark among models trained solely on publicly available retrieval datasets.
InsLogicBench: An Argumentation Logic Grounded Benchmark for Complex Insurance Claims Adjudication
PDF ↗Insurance claims adjudication demands not only accurate decisions but also interpretable reasoning grounded in policy clauses. However, existing benchmarks are limited to information retrieval or simple multiple-choice setups, which fail to require step-by-step inferences from facts to conclusions. To address this gap, we introduce InsLogicBench, a benchmark providing complete reasoning traces that link factual inputs, relevant policy clauses, and final verdicts. We construct the dataset using a controllable synthesis framework based on the Nested Toulmin Model. By capturing the defeasible logic of insurance policies through hierarchical truth assignment and enforcing validity via consistency verification, we ensure interpretability and logical rigor across generated examples. We evaluate eight Large Language Models (LLMs) on InsLogicBench. Results show significant difficulties in handling exception clauses and verifying missing conditions. Notably, models often produce correct final decisions but fail to provide precise justifications, highlighting a critical discrepancy between their decision accuracy and logical reasoning capabilities.
IRIS: Interleaved Reinforcement with Incremental Staged Curriculum for Cross-Lingual Mathematical Reasoning
PDF ↗Curriculum learning helps language models tackle complex reasoning by gradually increasing task difficulty. However, it often fails to generate consistent step-by-step reasoning, especially in multilingual and low-resource settings where cross-lingual transfer from English to Indian languages remains limited.We propose IRIS: Interleaved Reinforcement with Incremental Staged Curriculum, a two-axis framework that combines Supervised Fine-Tuning on progressively harder problems (vertical axis) with Reverse Curriculum Reinforcement Learning to reduce reliance on step-by-step guidance (horizontal axis). We design a composite reward combining correctness, step-wise alignment, continuity, and numeric incentives, optimized via Group Relative Policy Optimization (GRPO). We release CL-Math, a dataset of 29k problems with step-level annotations in English, Hindi, and Marathi.Across standard benchmarks and curated multilingual test sets, IRIS consistently improves performance, with strong results on math reasoning tasks and substantial gains in low-resource and bilingual settings, alongside modest improvements in high-resource languages. Our code and dataset will be publicly available at https://github.com/avinanand/IRIS-Interleaved-Reinforcement-
VLA models have emerged as a powerful paradigm for transferring semantic knowledge from web-scale data to physical robotic control. However, current single-frame architectures suffer from intrinsic limitations: temporal myopia that discards historical dynamics, reasoning gaps between high-level instructions and low-level motor commands, and inference inefficiency due to autoregressive scalar decoding. In this work, we propose MIRTH, a unified framework designed to address these challenges. MIRTH augments a pretrained VLA backbone with three key innovations: (1) dual-scale temporal memory hubs that compress long-term scene evolution and short-term motion trends into compact embeddings; (2) latent reasoning tokens optimized via a mutual-information objective carving out a semantic plan space to align multimodal context with action trajectories; and (3) a parallel action decoding scheme that replaces autoregressive generation with vector-wise prediction to maximize control throughput. Extensive evaluations on the LIBERO simulation benchmark and a real-world LeRobot platform demonstrate that MIRTH achieves state-of-the-art performance and exhibiting emergent error recovery capabilities. We will release our code and collected datasets to facilitate reproducible research in embodied AI upon publication.
Learning high quality sentence embeddings from dialogues has drawn increasing attentions as it is essential to solve a variety of dialogue-oriented tasks with low annotation cost. Annotating and gathering utterance relationships in conversations are difficult, while token-level annotations, , entities, slots and templates, are much easier to obtain. Other sentence embedding methods are usually sentence-level self-supervised frameworks and cannot utilize token-level extra knowledge. We introduce Template-aware Dialogue Sentence Embedding (TaDSE), a novel augmentation method that utilizes template information to learn utterance embeddings via self-supervised contrastive learning framework. We further enhance the effect with a synthetically augmented dataset that diversifies utterance-template association, in which slot-filling is a preliminary step. We evaluate TaDSE performance on five downstream benchmark dialogue datasets. The experiment results show that TaDSE achieves significant improvements over previous SOTA methods for dialogue. We further introduce a novel analytic instrument of semantic compression test, for which we discover a correlation with uniformity and alignment. Our code is available at https://github.com/minsik-ai/Template-Contrastive-Embedding
Audio carries richer information than text, including emotion, speaker traits, and environmental context, while also enabling lower-latency processing compared to speech-to-text pipelines. However, recent multimodal information retrieval research has predominantly focused on images, largely overlooking audio, especially in the setting of interleaved audio-text contextual retrieval. In this work, we introduce the Audio-Text Interleaved contextual Retrieval (ATIR) task, where queries can alternate between audio and text modalities. We construct an ATIR benchmark by integrating several Automatic Speech Recognition (ASR), QA, and retrieval datasets, ultimately unifying four types of contextual retrieval tasks. This benchmark substantially addresses the limitations of existing audio retrieval datasets in semantic retrieval. To study this task, we evaluate several off-the-shelf retrievers and train our ATIR model based on a Multimodal Large Language Model (MLLM). We further propose a novel token compression mechanism, which is orthogonal to existing compression methods, to mitigate the challenge of excessive audio tokens in MLLM-based ATIR models. Experimental results show that our ATIR model achieves significant improvements over strong baselines.
Understanding and inferring causal relationships from texts is a core aspect of human cognition and is essential for advancing large language models (LLMs) towards artificial general intelligence. Existing work evaluating LLM causal reasoning primarily relies on synthetic or simplified texts with explicitly stated causal relationships. These texts typically feature short passages and few causal relations, failing to reflect the complexities of real-world reasoning. In this paper, we investigate whether LLMs are capable of inferring causal relationships from real-world texts. We develop a benchmark drawn from real-world academic literature, which includes diverse texts with respect to length, complexity (different levels of explicitness, number of causal events and relationships), and domain. To the best of our knowledge, our benchmark is the first-ever real-world dataset for this task. Our experiments on this dataset show that LLMs face significant challenges in inferring causal relationships from real-world text, with the best-performing model achieving an average F_1 score of only 0.535. Through systematic analysis across aspects of real-world text (explicitness, number of causal events and relationships, length of text, domain), our benchmark offers targeted insights for further research into advancing LLM causal reasoning. Our code and dataset can be found at https://github.com/Ryan-Saklad/ReCITE.