We introduce FinWorkBench (a.k.a. Finch) for evaluating AI agents on real-world, enterprise-grade finance and accounting workflows that interleave data entry, structuring, formatting, web search, cross-file retrieval, calculation, modeling, validation, translation, visualization, and reporting. Finch is sourced from authentic enterprise workspaces from Enron (15,000 files and 500,000 emails) and other financial institutions, covering the period 2000–2025 and preserving the in-the-wild messiness of multimodal artifacts such as tables and charts across diverse domains including budgeting, trading, asset management, and operational management.We propose a workflow construction process that combines LLM-assisted mining of workflows from authentic enterprise environments with expert annotation: (1) LLM-assisted, expert-verified derivation of workflows from real-world email threads and spreadsheet version histories, and (2) meticulous annotation requiring over 700 hours of expert effort. This yields 172 composite workflows with 384 tasks, involving 1,710 spreadsheets with 27 million cells, along with PDFs and other artifacts, capturing the intrinsically messy, long-horizon, knowledge-intensive, and collaborative nature of real-world enterprise work.We conduct both human and automated evaluations of frontier AI systems, including GPT 5.1, Claude Sonnet/Opus 4.5, Gemini 3 Pro, Grok 4, and Qwen 3 Max. Under human evaluation, GPT 5.1 Pro spends an average of 16.8 minutes per workflow yet passes only 38.4% of workflows. Comprehensive case studies further surface the challenges that real-world enterprise workflows pose for AI agents.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
In the era of noisy intermediate-scale quantum (NISQ) computing, Quantum Neural Networks (QNNs) have emerged as a promising approach for various applications, yet their training is often hindered by barren plateaus (BPs), where gradient variance vanishes exponentially as the qubit size increases. Most initialization-based mitigation strategies rely heavily on pre-designed static parameter distributions, thereby lacking adaptability to diverse model sizes or data conditions. To address these limitations, we propose AdaInit, a foundational framework that leverages large language models with the submartingale property to iteratively synthesize initial parameters for QNNs that yield non-negligible gradient variance, thereby mitigating BPs. Unlike conventional one-shot initialization methods, AdaInit adaptively explores the parameter space by incorporating dataset characteristics and gradient feedback, with theoretical guarantees of convergence to finding a set of effective initial parameters for QNNs. We provide rigorous theoretical analyses of the submartingale-based process and empirically validate that AdaInit consistently outperforms existing initialization methods in maintaining higher gradient variance across various QNN scales. We believe this work may initiate a new avenue to mitigate BPs.
Applying large language models (LLMs) to assist in psycho-counseling is an emerging and meaningful approach, driven by the significant gap between patient needs and the availability of mental health support. However, current LLMs struggle to consistently provide effective responses to client speeches, largely due to the lack of supervision from high-quality real psycho-counseling data, whose content is typically inaccessible due to client privacy concerns. Furthermore, the quality of therapists’ responses in available sessions can vary significantly based on their professional training and experience. Assessing the quality of therapists’ responses remains an open challenge. We address these challenges by first proposing a set of professional and comprehensive principles to evaluate therapists’ responses to client speeches. Using these principles, we create a **Psy**cho-**Co**unseling **Pref**erence dataset, **PsyCoPref**, which contains 36k high-quality preference comparison pairs. This dataset aligns with the preferences of professional psychotherapists, providing a robust foundation for evaluating and improving LLMs in psycho-counseling. Experiments on reward modeling and preference learning demonstrate that PsyCoPref is an excellent resource for LLMs to acquire essential skills for responding to clients in a counseling session. Our best-aligned model achieves an impressive win rate of 87% against GPT-4o.
While most reading comprehension benchmarks for LLMs focus on factual information that can be answered by localizing specific textual evidence, many real-world tasks require understanding distributional information, such as population-level trends and preferences expressed across collections of text. We introduce Text2DistBench, a reading comprehension benchmark for evaluating LLMs’ ability to infer distributional knowledge from natural language. Built from real-world YouTube comments about movie and music entities, the benchmark provides models with entity metadata and associated comments, and requires them to answer distributional questions, such as estimating the proportions of positive and negative comments, or identifying the most and second most frequent topics discussed among viewers. To support reliable and long-term evaluation, the construction pipeline of Text2DistBench is fully automated and continuously updated to incorporate newly emerging entities over time. Experiments across multiple LLMs show that while models substantially outperform random baselines, performance varies widely across different distribution types and characteristics. These findings highlight both the capabilities and limitations of current LLMs in distributional reading comprehension and demonstrate the value of Text2DistBench as a practical and scalable testbed for future research.
Despite the rapid advancements in Multimodal Large Language Models (MLLMs), a critical question regarding their visual grounding mechanism remains unanswered: do these models genuinely read text embedded in images, or do they merely rely on parametric shortcuts in the text prompt? In this work, we diagnose this issue by introducing the Visualized-Question (VQ) setting, where text queries are rendered directly onto images to structurally mandate visual engagement. Our diagnostic experiments on Qwen2.5-VL reveal a startling capability-utilization gap: despite possessing strong OCR capabilities, models suffer a performance degradation of up to 12.7% in the VQ setting, exposing a deep-seated modality laziness. To bridge this gap, we propose SimpleOCR, a plug-and-play training strategy that imposes a structural constraint on the learning process. By transforming training samples into the VQ format with randomized styles, SimpleOCR effectively invalidates text-based shortcuts, compelling the model to activate and optimize its visual text extraction pathways. Empirically, SimpleOCR yields robust gains without architectural modifications. On four representative OOD benchmarks, it surpasses the base model by 5.4% and GRPO based on original images by 2.7%, while exhibiting extreme data efficiency, achieving superior performance with 30x fewer samples (8.5K) than recent RL-based methods. Furthermore, its plug-and-play nature allows seamless integration with advanced RL strategies like NoisyRollout to yield complementary improvements. Code is available at https://github.com/aiming-lab/SimpleOCR.
Large language models (LLMs) have achieved impressive proficiency in basic arithmetic, rivaling human-level performance on standard numerical tasks. However, little attention has been given to how these models perform when numerical expressions deviate from the prevailing conventions present in their training corpora. In this work, we investigate numerical reasoning across a wide range of numeral scripts and formats. We show that LLM accuracy drops substantially when numerical inputs are rendered in underrepresented scripts or formats, despite the underlying mathematical reasoning being identical. We further demonstrate that targeted prompting strategies, such as few-shot prompting and explicit numeral mapping, can greatly narrow this gap. Our findings highlight an overlooked challenge in multilingual numerical reasoning and provide actionable insights for working with LLMs to reliably interpret, manipulate, and generate numbers across diverse numeral scripts and formatting styles.
Multimodal Large Language Models (MLLMs) show strong potential for cross-modal understanding by integrating powerful language models with multimodal encoders. However, extending MLLMs to handle a diverse range of modalities introduces two critical and intertwined challenges: (1) the reliance on fully paired multimodal data, often scarce or costly to acquire across all modalities, and (2) the computational inefficiency from processing numerous modality tokens and requiring substantial model updates for each new modality. To address these challenges, we enable MLLMs to handle missing modalities by generating representations for absent inputs. Furthermore, recognizing that an increasing number of modalities leads to linearly scaling token counts and that lengthy generated sequences can hinder performance, we employ a dual-stage compression mechanism. It first reduces the number of tokens per modality and then condenses information from multiple modalities into a single, compact token sequence. This culminates in Flex-M^3, a novel MLLM framework designed for flexible and efficient learning across arbitrary combinations of modalities. Experiments across diverse multimodal benchmarks and backbones demonstrate that Flex-M^3 robustly handles varied modality inputs and scales efficiently. Notably, Flex-M outperforms its counterpart trained on only full-modality data, with consistent improvements of 2.29%, 3.15%, 11.01% on multimodal reasoning tasks NExT-QA, MUSIC-AVQA, SQA3D. Moreover, Flex-M^3 demonstrates superior robustness during inference, even when a high proportion of modalities are missing from the input samples, showcasing its capacity for complex, data-scarce multimodal applications.
NSF-CoT: Neuro-Symbolic Formal Verification of Chain-of-Thought Faithfulness in Contextual Question Answering
PDF ↗Chain-of-thought (CoT) prompting makes language models write step-by-step explanations, but these steps may not match what the model actually used to choose its answer. Existing faithfulness checks often only test whether changing the written chain changes the answer, without verifying whether the steps are truly supported by the given evidence, or they require special prompts that do not generalize well. We present NSF-CoT, a neuro-symbolic formal verification method that checks CoT faithfulness step by step for contextual question answering. NSF-CoT (1) converts the provided context facts and each reasoning step into simple logical statements, (2) uses counterfactual attribution to estimate which context facts the model relied on while generating each step, and (3) verifies each step using a hybrid checker that combines an SMT solver with an LLM-based entailment judge. For every step, we score groundedness (supported by the full context), validity (supported by the facts the model relied on), and utility (helps reach the final answer), and combine them into a faithfulness score. Across OpenBookQA, QASC, and HotpotQA, NSF-CoT consistently outperforms causal mediation, perturbation probes, and behavioral monitoring, and it identifies reasoning steps that are not only unfaithful but also harmful to the model’s final decision.
Large language models increasingly power AI agents for tasks requiring iterative refinement: document editing demands targeted revisions while preserving cross-references, code refactoring requires tracking function dependencies, and knowledge base updates cascade through related entities. Iterative editing with AI agents faces a fundamental efficiency-consistency tradeoff: maintaining consistency requires full-context awareness of dependencies, but processing entire documents for each edit incurs prohibitive token costs and latency. Isolated edits improve efficiency but risk breaking cross-references and violating semantic constraints. We introduce LEDGER (scaLing Agentic document editing with Dependency-aware Graph rEtRieval), a framework that constructs lightweight dependency graphs capturing semantic relationships and structural hierarchies across document elements. For each edit, graph traversal identifies affected elements and retrieves only necessary context. Experiments across 1,900 test cases spanning six state-of-the-art models show LEDGER achieves 76 consistency versus 56 baseline while reducing token usage by 85 . Critically, LEDGER with low reasoning effort matches baseline performance at high reasoning effort using 70 fewer tokens, suggesting explicit dependency representations can substitute for expensive internal reasoning with implications for agentic systems operating on structured data.
Do Language Models Think Consistently? A Study of Value Preferences Across Varying Response Lengths
PDF ↗Evaluations of LLMs’ ethical risks and value inclinations often rely on short-form surveys and psychometric tests, yet real-world use involves long-form, open-ended responses—leaving value-related risks and preferences in practical settings largely underexplored. In this work, we ask: Do value preferences inferred from short-form tests align with those expressed in long-form outputs? To address this question, we compare value preferences elicited from short-form reactions and long-form responses, varying the number of arguments in the latter to capture users’ differing verbosity preferences. Analyzing five LLMs (llama3-8b, gemma2-9b, mistral-7b, qwen2-7b, and olmo-7b), we find (1) a weak correlation between value preferences inferred from short-form and long-form responses across varying argument counts, and (2) similarly weak correlation between preferences derived from any two distinct long-form generation settings. (3 Alignment yields only modest gains in the consistency of value expression. Further, we examine how long-form generation attributes relate to value preferences, finding that argument specificity negatively correlates with preference strength, while representation across scenarios shows a positive correlation. Our findings underscore the need for more robust methods to ensure consistent value expression across diverse applications.
Proxy optimization, where AI systems exploit evaluator weaknesses rather than improve intended objectives, threatens both reinforcement learning (reward hacking) and LLM alignment (evaluator gaming). We introduce the Evaluator Stress Test (EST), an invariance-based framework that detects proxy gaming by separating exploitable sensitivity (e.g., format, physics bugs) from content-driven improvements using controlled perturbations with semantic validity audits. We validate EST across both domains. In RL, across 15 environments and 5 algorithms (2,156 expert-annotated episodes), EST achieves 78.4% precision and 81.7% recall. In LLM alignment, across 4 tasks, 2 model scales, 2 training methods, and 2 judges (1,200 human-annotated instances), EST achieves 74.2% precision and 78.6% recall with early warning signals preceding quality decline. Cross-domain analysis reveals that proxy-true correlation tracking transfers directly between domains, while perturbation design requires domain adaptation. Closed-loop mitigation improves human win-rate by 8.3 points (LLM) and reduces hacking by 54.6% (RL). We release benchmarks for both domains: 2,156 RL episodes and 1,200 LLM instances.
Retrieving visual and textual information from medical literature and hospital records can enhance diagnostic accuracy for clinical image interpretation. However, multimodal retrieval-augmented diagnosis is highly challenging. We explore a lightweight mechanism for enhancing diagnostic performance of retrieval-augmented LVLMs. We train an LVLM-aware multimodal retriever, such that the retriever learns to return images and texts that guide the LVLM toward correct predictions. In our low-resource setting, we perform only lightweight fine-tuning with small amounts of data, and use only general-purpose backbone models, achieving competitive results in clinical classification and VQA tasks compared to medically pre-trained models with extensive training. In a novel analysis, we highlight a previously unexplored class of errors that we term inconsistent retrieval predictions: cases where different top-retrieved images yield different predictions for the same target. We find that these cases are challenging for all models, even for non-retrieval models, and that our retrieval optimization mechanism significantly improves these cases over standard RAG. However, our analysis also sheds light on gaps in the ability of LVLMs to utilize retrieved information for clinical predictions.
Large Language Models (LLMs) have made strong progress in reasoning. To enhance the reasoning performance, a common inference-time approach is tree-based search, which decomposes the reasoning process into multiple steps, expands multiple reasoning paths, and uses reward models to prune and select candidates. However, based on our exploration, the simple decomposition may lead to suboptimal searching efficiency: while planning is generally harder, it is the execution errors that are more likely to propagate to later steps. This indicates that planning and execution play different roles in reasoning and should be treated differently during tree-based search. Given this, to enhance the searching efficiency, we propose a dual-phase test-time scaling framework that separates reasoning into planning and execution, and performs search over each phase independently. To further refine the algorithm, we also introduce a dynamic budget allocation mechanism that adaptively redistributes sampling effort based on reward feedback, allowing early stopping on confident steps and reallocation of computation to more challenging steps. Experiments on both math reasoning and code generation benchmarks demonstrate that our approach consistently improves accuracy while reducing redundant computation.
Multi-modal retrieval-augmented generation (MM-RAG) relies heavily on re-rankers to surface the most relevant evidence for image-question queries. However, standard re-rankers typically process the full query image as a global embedding, making them susceptible to visual distractors (e.g., background clutter) that skew similarity scores.We propose **Region-R1**, a query-side region cropping framework that formulates region selection as a decision-making problem during re-ranking, allowing the system to learn to retain the full image or focus only on a question-relevant region before scoring the retrieved candidates. Region-R1 learns a policy with a novel region-aware group relative policy optimization (r-GRPO) to dynamically crop a discriminative region. Across two challenging benchmarks, E-VQA and InfoSeek, Region-R1 delivers consistent gains, achieving state-of-the-art performances by increasing conditional Recall@1 by up to 20%. These results show the great promise of query-side adaptation as a simple but effective way to strengthen MM-RAG re-ranking.
Multimodal Retrieval-Augmented Generation (RAG) systems retrieve evidence at coarse granularities (entire images or scenes), creating a mismatch with fine-grained user queries and making failures unverifiable. We introduce GranuVistaVQA, a multimodal benchmark featuring real-world landmarks with element-level annotations across multiple viewpoints, capturing the partial observation challenge where individual images contain only subsets of entities. We further propose GranuRAG, a multi-granularity framework that treats visual elements as first-class retrieval units through three stages: element-level detection and classification, multi-granularity cross-modal alignment for evidence retrieval, and attribution-constrained generation. By grounding retrieval at the element level rather than relying on implicit attention, our approach enables transparent error diagnosis. Experiments demonstrate that GranuRAG achieves up to 29.2% improvement over six strong baselines for this task.
PAPERMIND: Benchmarking Agentic Reasoning and Critique over Scientific Papers in Multimodal LLMs
PDF ↗Understanding scientific papers requires more than answering isolated questions or summarizing content. It involves an integrated reasoning process that grounds textual and visual information, interprets experimental evidence, synthesizes information across sources, and critically evaluates scientific claims. However, existing benchmarks typically assess these abilities in isolation, making it difficult to evaluate scientific paper understanding as a unified set of interacting cognitive abilities. In this work, we introduce PaperMind , a benchmark designed to evaluate integrated and agent-oriented scientific reasoning over research papers. PaperMind is constructed from real scientific papers across seven domains, including agriculture, biology, chemistry, computer science, medicine, physics, and economics. It comprises four complementary task families that collectively operationalize distinct cognitive facets of scientific paper reasoning, including multimodal grounding, experimental interpretation, cross-source evidence reasoning, and critical assessment. By analyzing model behavior across multiple tasks, PaperMind enables a diagnostic evaluation of integrated scientific reasoning behaviors that are difficult to assess through isolated task evaluations. Extensive experiments on both open-source and closed-source multimodal LLMs reveal consistent performance gaps across tasks, highlighting persistent challenges in integrated scientific reasoning and critique. Our benchmark and dataset are available at https://github.com/Yanjun-Zhao/PaperMind.
Class imbalance is a widespread challenge in NLP tasks, significantly hindering robust performance across diverse domains and applications. We introduce Hardness-Aware Meta-Resample (HAMR), a unified framework that adaptively addresses both class imbalance and data difficulty. HAMR employs bi-level optimizations to dynamically estimate instance-level weights that prioritize genuinely challenging samples and minority classes, while a neighborhood-aware resampling mechanism amplifies training focus on hard examples and their semantically similar neighbors. We validate HAMR on six imbalanced datasets covering multiple tasks and spanning biomedical, disaster response, and sentiment domains. Experimental results show that HAMR achieves substantial improvements for minority classes and consistently outperforms strong baselines. Extensive ablation studies demonstrate that our proposed modules synergistically contribute to performance gains and highlight HAMR as a flexible and generalizable approach for class imbalance adaptation.
Not Safe for Work (NSFW) image classifiers play a critical role in safeguarding text-to-image (T2I) systems. However, a concerning phenomenon has emerged in T2I systems – changes in text prompts that manipulate benign image elements can result in failed detection by NSFW classifiers – dubbed "*context shifts*." For instance, while a NSFW image of "*a nude person in an empty scene*" can be easily blocked by most NSFW classifiers, a stealthier one that depicts "*a nude person blending in a group of dressed people*" may evade detection. We ask: how to systematically reveal NSFW image classifiers’ failure against such context shifts?Towards this end, we present an automated red-teaming framework that leverages a set of generative AI tools. We propose an **exploration-exploitation** approach: **First**, in the *exploration* stage, we synthesize a diverse and massive 36K NSFW image dataset that facilitates our study of context shifts. We find that varying fractions (e.g., 4.1% to 36% nude and sexual content) of the dataset are misclassified by NSFW image classifiers like GPT-4o and Gemini. **Second**, in the *exploitation* stage, we leverage these failure cases to train a specialized LLM that rewrites unseen seed prompts into more evasive versions, increasing the likelihood of detection evasion by up to 6 times. Alarmingly, we show **these failures translate to real-world T2I and even T2V systems** like DALL-E 3, Sora, Nano Banana, and Veo 3 – beyond the open-weight image generators in our main study. For example, querying DALL-E 3 with prompts rewritten by our approach increases the chance of obtaining NSFW images from 0 to over 50%.
Repository-level code completion benefits from retrieval-augmented generation (RAG). However, controlling cross-file evidence is difficult because chunk utility is often interaction-dependent: some snippets help only when paired with complementary context, while others harm decoding when they conflict. We propose RepoShapley, a coalition-aware context filtering framework supervised by Shapley-style marginal contributions. Our offline labeling module, ChunkShapley, estimates signed per-chunk effects via teacher-forced probing, feeds them into a lightweight surrogate game that captures saturation and interference, computes exact Shapley values for small retrieval sets, and selects a decoding-optimal coalition through bounded post-verification with the frozen generator. The verified <KEEP> / <DROP> decisions and retrieval triggers are then distilled into a single model via discrete control tokens. Experiments across benchmarks and backbones show that RepoShapley improves completion quality while reducing harmful context and unnecessary retrieval.
Vision-Language Models Mistake Head Orientation for Gaze Direction: Nonverbal Conversation Cues
PDF ↗Where someone looks is a nonverbal communication cue that children and adults readily use.How well can Vision-Language Models (VLMs) infer gaze targets? To construct evaluation stimuli, we captured 1,360 real-world photos of scenes in which a person gazes at one of several objects on a table. Importantly, we also controlled the gazer’s head orientation: sometimes it was directed toward the gaze target, sometimes toward a distractor object, and sometimes left unconstrained. We found a substantial performance gap between VLMs and humans, ruled out alternative explanations such as resolution and object-naming skills, and identified the main reason for the gap as VLMs inferring gaze direction using head orientation rather than eye appearance.Such a bias is likely due to data rather than architecture, as suggested by a proof-of-concept experiment finetuning a transformer-based vision model.Future work should investigate whether these findings hold broadly across various deep learning methods trained on existing data, and whether better data mitigates this problem for all architectures.Pinpointing the reason sets the stage for technologies that can interpret gaze targets to have more efficient interactions with humans.