Tool-use large language model (LLM) agents are increasingly deployed to support sensitive workflows, relying on tool calls for retrieval, external API access, and session memory management. While prior research has examined various threats, the risk of systematic data exfiltration by backdoored agents remains underexplored. In this work, we present a backdoor attack framework that embeds semantic triggers into fine-tuned LLM agents. When triggered, the backdoored agent invokes memory-access tool calls to retrieve stored user context and exfiltrates it via disguised retrieval tool calls. We further demonstrate that multi-turn interaction amplifies the impact of data exfiltration, as attacker-controlled retrieval responses can subtly steer subsequent agent behavior and user interactions, enabling sustained and cumulative information leakage over time. Our experimental results expose a critical vulnerability in LLM agents with tool access and highlight the need for defenses against exfiltration-oriented backdoors.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Peer review relies on substantive, evidence-based questions, yet current LLMs generate surface-level queries that perform worse than human reviewer questions in expert evaluation. To address this gap, we curate a high-quality dataset of reviewer questions from OpenReview and conduct a human preference study where expert annotators evaluate question-paper pairs across three dimensions: effort, evidence, and grounding. From these annotations, we train IntelliReward, a reward model built from a frozen autoregressive LLM with trainable multi-head transformers. Validated against expert judgments, IntelliReward predicts reviewer-question quality better than API-based SFT baselines and provides scalable evaluation. We apply Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) with IntelliReward to train IntelliAsk, a question-generation model aligned with human standards of effortful, evidence-based critique. Human evaluations show IntelliAsk generates more grounded, substantive and effortful questions than strong baselines and reduces reliance on first-page content. We also find improvements on reasoning and writing benchmarks, suggesting reviewer-question quality correlates with broader capabilities. Compared to Qwen3-32B, IntelliAsk improves MuSR (68.3 vs 64.7 Acc) and WritingBench (8.31 vs 8.07). We release our code, filtered review dataset, expert annotations, IntelliAsk and IntelliReward to support automatic evaluation of grounding, effort, and evidence in LLM-generated review questions.
Reliable financial reasoning requires knowing not only how to answer, but also when an answer cannot be justified. In real financial practice, problems often rely on implicit assumptions that are taken for granted rather than stated explicitly, causing problems to appear solvable while lacking enough information for a definite answer. We introduce RealFin, a bilingual benchmark that evaluates financial reasoning by systematically removing essential premises from exam-style questions while keeping them linguistically plausible. Based on this, we evaluate models under three formulations that test answering, recognizing missing information, and rejecting unjustified options, and find consistent performance drops when key conditions are absent. General-purpose models tend to over-commit and guess, while most finance-specialized models fail to clearly identify missing premises. These results highlight a critical gap in current evaluations and show that reliable financial models must know when a question should not be answered. The dataset and code are available athttps://github.com/insait-institute/RealFin.
STELLA: A Multimodal LLM for Protein Functional Annotation via Unified Sequence-Structure Encoding
PDF ↗Understanding the intricate interplay among sequence, structure, and function remains a fundamental challenge in proteomics. The sequence-structure-function paradigm posits that biological roles are governed by the tertiary geometric conformations encoded within primary sequences; consequently, integrating these multi-modal descriptors is imperative for accurate functional annotation. While protein language models (pLMs) have achieved significant progress via representation learning on massive sequence data, they often lack the capacity to incorporate high-resolution structural information and the rich textual context that characterizes protein roles. In this work, we present STELLA, a multimodal LLM that synergistically aligns bimodal (sequence-structure) representations with the textual modality to advance protein functional annotation. By leveraging ESM3 for unified bimodal encoding and Llama-3.1-8B-Instruct for natural language modeling, STELLA achieves state-of-the-art performance in two critical tasks: Functional Description Prediction and Enzyme-catalyzed Reaction Prediction. This study demonstrates that multimodal LLMs represent a paradigm shift beyond pure pLMs, offering a new frontier for protein biology and biomedical discovery. The codes can be accessed via https://github.com/ocx-lab/STELLA.
Vision-Language Models (VLMs) often prioritize linguistic fluency over visual fidelity, leading to hallucinations where generated text contradicts the image. Countering this bias typically requires resource-heavy fine-tuning or high-latency verification methods that provide feedback only after the full response is generated. To overcome these limitations, we present a framework for Token-level Inference-Time Alignment (TITA) that steers the decoding process without updating the base model parameters. By training a lightweight reward model to capture visual preferences, TITA extracts implicit guidance through log-probability ratios. This approach functions as an inference-time adaptation of Direct Preference Optimization (DPO), injecting dense feedback to correct the output distribution at every generation step. Across diverse architectures including LLaVA-1.5, Qwen3-VL, and InternVL3.5, TITA consistently improves performance on 13 benchmarks. For example, TITA boosts LLaVA-1.5-7B by 8.6% on MMVet and achieves a 74.0 MMStar score with Qwen3-VL-8B. Specifically, these gains incur negligible overhead (~0.2s per query), offering a superior trade-off between alignment effectiveness and efficiency. Our code is available at: https://github.com/Thecommonirin/TITA.
Recent work has demonstrated the promise of orchestrating large language models (LLMs) within evolutionary and agentic optimization systems. However, the mechanisms driving these optimization gains remain poorly understood. In this work, we present a large-scale study of LLM-guided evolutionary search, collecting optimization trajectories for 15 LLMs across 8 tasks. Although zero-shot problem-solving ability correlates with final optimization outcomes, it explains only part of the variance: models with similar initial capability often induce dramatically different search trajectories and outcomes. By analyzing these trajectories, we find that strong LLM optimizers behave as local refiners, producing frequent incremental improvements while progressively localizing the search in semantic space. Conversely, weaker optimizers exhibit large semantic drift, with sporadic breakthroughs followed by stagnation. Notably, various measures of solution novelty do not predict final performance; novelty is beneficial only when the search remains sufficiently localized around high-performing regions of the solution space. Our results highlight the importance of trajectory analysis for understanding and improving LLM-based optimization systems and provide actionable insights for their design and training.
Code optimization remains a core objective in software development, yet modern compilers struggle to navigate the enormous optimization spaces. While recent research has looked into employing large language models (LLMs) to optimize source code directly, these techniques can introduce semantic errors and miss fine-grained compiler-level optimization opportunities. We present HintPilot, which bridges LLM-based reasoning with traditional compiler infrastructures via synthesizing compiler hints—annotations that steer compiler behavior. HintPilot employs retrieval-augmented synthesis over compiler documentation and applies profiling-guided iterative refinement to synthesize semantics-preserving and effective hints. Upon PolyBench and HumanEval-CPP benchmarks, HintPilot achieves up to 6.88x geometric mean speedup over -Ofast while preserving program correctness.
Agentic workflows, composed of multiple collaborating Large Language Models (LLMs), have become a key paradigm for complex problem-solving. However, their effectiveness is often hindered by three critical challenges: high manual design costs, inefficient agentic search, and poor dynamic adaptability to new tasks and human preferences. To address these limitations, we propose HFlow, an evolutionary framework for generating agentic workflows through human-agent collaboration. HFlow employs an evolutionary algorithm to automate the search for optimal workflows by mutating and crossing over their structures, prompts, and LLM backbones. This process is guided by human preferences to ensure rapid convergence, while a hierarchical experience memory enables the generalization of learned strategies. Extensive experiments on math and code generation benchmarks show HFlow surpasses other automated baselines by up to 27.34%, while achieving comparable performance to o1-preview at only one-fourth of the cost. Our work introduces a new paradigm for workflow design that produces cost-effective and adaptive solutions, better aligning automated agentic systems with dynamic human needs.
Large reasoning models exhibit long chain-of-thought reasoning with complex strategies such as backtracking and self-verification. Yet, these capabilities typically require resource-intensive post-training. We investigate whether such behaviors can be elicited in large models without any gradient updates. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logit arithmetic to transfer these capabilities from a substantially smaller reasoning guider to a large non-reasoning target. We further show that we can boost performance by training the guider to correct the target’s errors using preference optimization over mixed model outputs, a setup we refer to as ThinkLogit-DPO. We evaluate these methods across six reasoning benchmarks spanning math, science, and coding domains using the Qwen2.5-32B guided by R1-Distill-Qwen-1.5B, a model 21x smaller. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement of 21.5% and 24.2%, respectively, over the target model. Moreover, ThinkLogit remains effective even when the guider and target come from different model families.Crucially, our method requires zero training for the large model and would incur minimal inference overhead when logits are computed in parallel, presenting a practical solution for enabling long reasoning at scale.
ThinkLinker: From Low-Rank Interaction to Knowledge-Aware Verification for Multimodal Entity Linking
PDF ↗Recent advances in Multimodal Entity Linking (MEL) exploit textual and visual information to disambiguate mentions and align them with entities in a knowledge base. Existing methods typically design separate and complex network modules for each type of interaction among multi-granular and multimodal features, while lacking explicit modeling of the joint dependencies among these features. Moreover, most approaches rely on unidirectional retrieval-based matching and lack knowledge-driven verification, leading to unreliable disambiguation in weak-context scenarios. To address these challenges, we propose a novel two-stage MEL framework termed ThinkLinker. First, we introduce a low-rank fusion mechanism to model the joint dependencies among multi-granular and multimodal features, enabling comprehensive and explicit interactions while learning task-relevant discriminative information for candidate ranking in a lower-dimensional space. Subsequently, we develop a bidirectional retrieval-verification paradigm, where the ranked candidate entities guide an LLM-based multi-turn, dialogue-style verification process to generate mention-specific contextual augmentation. The augmented context is then adaptively fused with the original representation to further refine the linking model. Experimental results on public benchmark datasets demonstrate that the proposed ThinkLinker outperforms all state-of-the-art baselines. The code is publicly available at https://github.com/zhouyuanyu/ThinkLinker.
Sculpting the Vector Space: Towards Efficient Multi-Vector Visual Document Retrieval via Prune-then-Merge Framework
PDF ↗Visual Document Retrieval (VDR), which aims to retrieve relevant pages within vast corpora of visually-rich documents, is of significance in current multimodal retrieval applications. The state-of-the-art multi-vector paradigm excels in performance but suffers from prohibitive overhead, a problem that current efficiency methods like pruning and merging address imperfectly, creating a difficult trade-off between compression rate and feature fidelity. To overcome this dilemma, we introduce **Prune-then-Merge**, a novel two-stage framework that synergizes these complementary approaches. Our method first employs an adaptive pruning stage to filter out low-information patches, creating a refined, high-signal set of embeddings. Subsequently, a hierarchical merging stage compresses this pre-filtered set, effectively summarizing semantic content without the noise-induced feature dilution seen in single-stage methods. **Extensive experiments on 29 VDR datasets demonstrate that our framework consistently outperforms existing methods, significantly extending the near-lossless compression range and providing robust performance at high compression ratios.**
From Atomic to Complex tasks: Cross-Tasking Improves Zero-Shot Argument Generation and Retrieval
PDF ↗Cross-task generalization mimics human intelligence through the ability to perform tasks by recalling foundational skills acquired previously. In this paper, we argue that argument generation and argument retrieval are complex tasks that could leverage cross-tasking atomic argument mining and argument quality assessment tasks, even if there is no supervision. We empirically demonstrate the rationale behind our claim through the \textit{ArgLLM} framework, including a total of 18.9K instruction data using a multi-choice question-answering format, scaling up through multi-tasking and model merging, six natural language argumentation atomic tasks to four complex argument generation and argument retrieval tasks. Our results and analysis, using the backbone Mistral and Llama models, show that cross-tasking in zero-shot settings outperforms base models and is robust to varying strategies, tasks, and model sizes, offering a valuable trade-off between computational cost and task performance.
High-fidelity audio generation techniques, such as voice conversion and singing voice synthesis, have significantly increased the risk of audio deepfakes. Although existing methods perform well on conversational speech deepfake detection, they fail severely under the speech-to-singing domain shift. To address this limitation, we propose GenuVoice, a unified deepfake detector based on a multi-branch mixture-of-experts architecture that integrates three complementary feature views: Wav2Vec 2.0 representations, log-mel spectrograms, and mel-frequency cepstral coefficients (MFCC). Each expert is trained to remain independently discriminative, while a learned gating network dynamically weights expert contributions. A speech-retentive multi-domain fine-tuning strategy enables adaptation to singing without degrading speech performance. GenuVoice achieves 1.82% Equal Error Rate (EER) on CtrSVDD, compared to 37–62% for existing speech-trained detectors, while preserving strong speech performance (0.38% EER on ASVspoof 2019) and generalizing to unseen generators (8.89% EER on held-out ASVspoof 2021). Extensive ablations confirm the importance of multi-expert fusion and speech retention, establishing GenuVoice as an effective unified detector for speech and singing deepfakes. The implementation code is available at https://github.com/aastha-sharma/genuvoice
Financial question answering (QA) over long corporate filings requires evidence to satisfy strict constraints on entities, financial metrics, fiscal periods, and numeric values. However, existing LLM-based rerankers primarily optimize semantic relevance, leading to unstable rankings and opaque decisions on long documents. We propose FINCARDS, a structured reranking framework that reframes financial evidence selection as constraint satisfaction under a finance-aware schema. FINCARDS represents filing chunks and questions using aligned schema fields (entities, metrics, periods, and numeric spans), enabling deterministic field-level matching. Evidence is selected via a multi-stage tournament reranking with stability-aware aggregation, producing auditable decision traces. Across two corporate filing QA benchmarks, FINCARDS substantially improves early-rank retrieval over both lexical and LLM-based reranking baselines, while reducing ranking variance, without requiring model fine-tuning or unpredictable inference budgets. Our code is available at https://github.com/XanderZhou2022/FINCARDS.
Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification
PDF ↗Recent advances in Deep Research Agents (DRAs) are transforming automated knowledge discovery and problem-solving.While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: test-time self-evolving the agent’s ability by iteratively verifying the policy model’s outputs, guided by meticulously crafted rubrics. This approach gives rise to an inference-time scaling of verification, wherein an agent self-improves at test time by evaluating its generated answers to produce iterative feedback and refinements without any additional training. We derive the rubrics based on an automatically constructed DRA Failure Taxonomy, which systematically classifies agent failures into five major categories and thirteen sub-categories. We present DeepVerifier, a rubrics-based outcome reward verifier that leverages the asymmetry of verification and outperforms vanilla agent-as-judge and LLM judge baselines by 12%–48% in meta-evaluation F1 score. To enable practical test-time self-evolution, DeepVerifier integrates as a plug-and-play module during test-time inference. The verifier produces detailed rubric-based feedback, which is fed back to the agent for iterative bootstrapping—refining responses without additional training. This test-time scaling delivers 8%–11% accuracy gains on challenging subsets of GAIA and XBench-DeepResearch when powered by capable closed-source LLMs. Finally, to support open-source advancement, we release DeepVerifier-4K, a curated supervised fine-tuning dataset of 4,646 high-quality agent steps focused on DRA verification. These examples emphasize reflection and self-critique, enabling open models to develop robust verification capabilities.
Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization
PDF ↗Alignment of large language models (LLMs) typically relies on supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), or more recently direct preference optimization (DPO). However, existing objectives largely ignore the global geometry and topology of the representation space: they operate on local token-level likelihoods or scalar preference scores, and do not explicitly constrain how hidden states move from a user prompt to an answer.We view generation as tracing a semantic trajectory in hidden space, and propose a topology-enhanced alignment framework that regularizes these trajectories using 0-dimensional persistent homology. First, at the SFT stage, we introduce a Trajectory Topology Loss (TTL). For each batch, we treat mean-pooled embeddings of prompts and gold answers as a mixed point cloud, run a Union-Find-based 0D persistent homology algorithm, and extract ”prompt–answer bridge” edges that connect previously disconnected components. TTL encourages the model’s actual update direction from prompt to answer to align with these topologically derived bridges, rather than with arbitrary or per-example directions.Second, at the RLHF/DPO stage, we propose Topological Preference Optimization (TPO). TPO constructs topic-specific semantic preference vectors from an offline pipeline and aligns the semantic improvement direction between rejected and chosen responses with these vectors in an intermediate hidden layer. We further introduce an exponential-moving-average-based dynamic weighting scheme to balance DPO and TPO losses, and also explore a fully topological variant that applies persistent homology on the chosen/rejected embedding cloud.We instantiate our methods on Qwen2.5-7B-Instruct and evaluate on UltraChat and Anthropic HH-RLHF. Across both SFT and DPO training, topology-enhanced objectives consistently outperform strong non-topological baselines (including per-example, nearest-neighbor, and random direction regularizers) on automatic preference metrics and LLM-judge evaluations, while maintaining or slightly improving toxicity. These results suggest that incorporating persistent homology and trajectory geometry is a promising and practical direction for more controllable LLM alignment.
Multimodal large language models (MLLMs) have achieved strong performance on challenging visual question answering benchmarks, yet their inference efficiency is severely constrained by the rapidly growing context. This growth stems from two primary sources: the large number of visual tokens required to encode images, and the accumulation of intermediate reasoning traces during autoregressive generation. To address these challenges, we propose LaT (**L**ook **a**nd **T**hink), the first modality-decoupled compression method that enables efficient multimodal inference. LaT structures reasoning into alternating looking and thinking steps, thereby explicitly signaling when visual grounding is required. Building on this design, LaT (1) evicts visual tokens whenever visual grounding is unnecessary, and (2) applies co-learning-guided compression after each completed step, mitigating the two sources of context growth respectively. Experimental results demonstrate that LaT reduces the average context length by up to 57%, while maintaining performance comparable to the standard MLLM baseline. The code will be publicly released.
Large Language Models (LLMs) have made remarkable breakthroughs in reasoning, yet continue to struggle with hallucinations, logical errors, and inability to self-correct during complex multi-step tasks. Current approaches like chain-of-thought prompting offer limited reasoning capabilities that fail when precise step validation is required. We propose Environment Augmented Generation (EAG), a framework that enhances LLM reasoning through: (1) real-time environmental feedback validating each reasoning step, (2) dynamic branch exploration for investigating alternative solution paths when faced with errors, and (3) experience-based learning from successful reasoning trajectories. Unlike existing methods, EAG enables deliberate backtracking and strategic replanning through tight integration of execution feedback with branching exploration. Our a1-32B model achieves state-of-the-art performance among similar-sized models across all benchmarks, matching larger models like o1 on competition mathematics while outperforming comparable models by up to 24.4 percentage points. Analysis reveals EAG’s distinctive scaling pattern: initial token investment in environment interaction yields substantial long-term performance dividends, with advantages amplifying proportionally to task complexity.
Automatic speech recognition (ASR) still covers only a small fraction of the world’s languages, mainly due to supervised data scarcity. In-context learning (ICL) with large language models (LLMs) addresses this problem, but prior work largely focuses on high-resource languages covered during training and text-only settings. This paper investigates whether speech LLMs can learn unseen languages with multimodal ICL (MICL), and how this learning can be used to improve ASR. We conduct experiments with two speech LLMs, Phi-4 and Qwen3-Omni, on three diverse endangered languages. Firstly, we find that MICL is effective for unseen languages, leveraging both speech and text modalities. We further show that cross-lingual transfer learning improves MICL efficiency on target languages without training on them. Moreover, we analyze attention patterns to interpret MICL mechanisms, and we observe layer-dependent preferences between audio and text context, with an overall bias towards text. Finally, we show that prompt-based ASR with speech LLMs performs poorly on unseen languages, motivating a simple ASR system that combines a stronger acoustic model with a speech LLM via MICL-based selection of acoustic hypotheses. Results show that MICL consistently improves ASR performance, and that cross-lingual transfer learning matches or outperforms corpus-trained language models without using target-language data. Our code is publicly available
RATION: Entropy-Driven Task-Adaptive Visual Attention Allocation Framework for Multimodal Reasoning
PDF ↗Multimodal Large Language Models (MLLMs) integrate visual encoders with Large Language Models (LLMs) and enable multimodal reasoning. However, for tasks that heavily rely on visual information, the model’s utilization of visual information remains unstable, which leads to reasoning failures. Prior works mainly strengthen multimodal reasoning by improving representation alignment or increasing computation. However, these methods do not explicitly characterize the differences in visual demands across tasks, making it difficult for the model to decide where and how strongly to attend to visual information. Consequently, visual attention allocation becomes a key factor that affects multimodal reasoning. To address these, we propose RATION, an entropy-driven task-adaptive visual attention allocation framework. First, we use a task routing strategy to infer the task type of each sample and identify the key layers. We use visual attention entropy as a control signal to dynamically allocate attention according to task demands. Experiments show that RATION achieves consistent performance gains across diverse reasoning tasks, datasets, and models, providing a clear direction toward more reliable multimodal reasoning.