Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems

Failure attribution in LLM multi-agent systems-identifying the agent and step responsible for task failures-provides crucial clues for systems debugging but remains underexplored and labor-intensive. In this paper, we propose and formulate a new research area: automated failure attribution for LLM multi-agent systems. To support this initiative, we introduce the Who&When dataset, comprising extensive failure logs from 127 LLM multi-agent systems with fine-grained annotations linking failures to specific agents and decisive error steps. Using the Who&When, we develop and evaluate three automated failure attribution methods, summarizing their corresponding pros and cons. The best method achieves 53.5% accuracy in identifying failure-responsible agents but only 14.2% in pinpointing failure steps, with some methods performing below random. Even SOTA reasoning models, such as OpenAI o1 and DeepSeek R1, fail to achieve practical usability. These results highlight the task's complexity and the need for further research in this area. Code and dataset are available at https://github.com/mingyin1/Agents_Failure_Attribution

Nemotron-Research-Tool-N1: Tool-Using Language Models with Reinforced Reasoning

Enabling large language models with external tools has become a pivotal strategy for extending their functionality beyond text generation tasks. Prior work typically enhances tool-use abilities by either applying supervised fine-tuning (SFT) to enforce tool-call correctness or distilling reasoning traces from stronger models for SFT. However, both approaches fall short, either omitting reasoning entirely or producing imitative reasoning that limits generalization. Inspired by the success of DeepSeek-R1 in eliciting reasoning through rule-based reinforcement learning, we develop the Nemotron-Research-Tool-N1 series of tool-using language models using a similar training paradigm. Instead of restrictively supervising intermediate reasoning traces distilled from stronger models, Nemotron-Research-Tool-N1 is optimized with a binary reward that evaluates only the structural validity and functional correctness of tool invocations. This lightweight supervision allows the model to autonomously internalize reasoning strategies, without the need for annotated reasoning trajectories. Experiments on the BFCL and API-Bank benchmarks show that Nemotron-Research-Tool-N1-7B and Nemotron-Research-Tool-N1-14B, built on Qwen-2.5-7B/14B-Instruct, achieve state-of-the-art results, outperforming GPT-4o on both evaluations.

Discovering Knowledge Deficiencies of Language Models on Massive Knowledge Base

Large language models (LLMs) possess impressive linguistic capabilities but often fail to faithfully retain factual knowledge, leading to hallucinations and unreliable outputs. Understanding LLMs' knowledge deficiencies by exhaustively evaluating against full-scale knowledge bases is computationally prohibitive, especially for closed-weight models. We propose stochastic error ascent (SEA), a scalable and efficient framework for discovering knowledge deficiencies (errors) in closed-weight LLMs under a strict query budget. Rather than naively probing all knowledge candidates, SEA formulates error discovery as a stochastic optimization process: it iteratively retrieves new high-error candidates by leveraging the semantic similarity to previously observed failures. To further enhance search efficiency and coverage, SEA employs hierarchical retrieval across document and paragraph levels, and constructs a relation directed acyclic graph to model error propagation and identify systematic failure modes. Empirically, SEA uncovers 40.7x more knowledge errors than Automated Capability Discovery and 26.7% more than AutoBencher, while reducing the cost-per-error by 599x and 9x, respectively. Human evaluation confirms the high quality of generated questions, while ablation and convergence analyses validate the contribution of each component in SEA. Further analysis on the discovered errors reveals correlated failure patterns across LLM families and recurring deficits, highlighting the need for better data coverage and targeted fine-tuning in future LLM development.

Generate Any Scene: Evaluating and Improving Text-to-Vision Generation with Scene Graph Programming

DALL-E and Sora have gained attention by producing implausible images, such as "astronauts riding a horse in space." Despite the proliferation of text-to-vision models that have inundated the internet with synthetic visuals, from images to 3D assets, current benchmarks predominantly evaluate these models on real-world scenes paired with captions. We introduce Generate Any Scene, a framework that systematically enumerates scene graphs representing a vast array of visual scenes, spanning realistic to imaginative compositions. Generate Any Scene leverages 'scene graph programming', a method for dynamically constructing scene graphs of varying complexity from a structured taxonomy of visual elements. This taxonomy includes numerous objects, attributes, and relations, enabling the synthesis of an almost infinite variety of scene graphs. Using these structured representations, Generate Any Scene translates each scene graph into a caption, enabling scalable evaluation of text-to-vision models through standard metrics. We conduct extensive evaluations across multiple text-to-image, text-to-video, and text-to-3D models, presenting key findings on model performance. We find that DiT-backbone text-to-image models align more closely with input captions than UNet-backbone models. Text-to-video models struggle with balancing dynamics and consistency, while both text-to-video and text-to-3D models show notable gaps in human preference alignment. We demonstrate the effectiveness of Generate Any Scene by conducting three practical applications leveraging captions generated by Generate Any Scene: 1) a self-improving framework where models iteratively enhance their performance using generated data, 2) a distillation process to transfer specific strengths from proprietary models to open-source counterparts, and 3) improvements in content moderation by identifying and generating challenging synthetic data.

Template Matters: Understanding the Role of Instruction Templates in Multimodal Language Model Evaluation and Training

Current multimodal language models (MLMs) evaluation and training approaches overlook the influence of instruction format, presenting an elephant-in-the-room problem. Previous research deals with this problem by manually crafting instructions, failing to yield significant insights due to limitations in diversity and scalability. In this work, we propose a programmatic instruction template generator capable of producing over 39B unique template combinations by filling randomly sampled positional synonyms into weighted sampled meta templates, enabling us to comprehensively examine the MLM's performance across diverse instruction templates. Our experiments across eight common MLMs on five benchmark datasets reveal that MLMs have high template sensitivities with at most 29% performance gaps between different templates. We further augment the instruction tuning dataset of LLaVA-1.5 with our template generator and perform instruction tuning on LLaVA-1.5-7B and LLaVA-1.5-13B. Models tuned on our augmented dataset achieve the best overall performance when compared with the same scale MLMs tuned on at most 75 times the scale of our augmented dataset, highlighting the importance of instruction templates in MLM training. The code is available at https://github.com/shijian2001/TemplateMatters .

TACO: Learning Multi-modal Action Models with Synthetic Chains-of-Thought-and-Action

While open-source multi-modal language models perform well on simple question answering tasks, they often fail on complex questions that require multiple capabilities, such as fine-grained recognition, visual grounding, and reasoning, and that demand multi-step solutions. We present TACO, a family of multi-modal large action models designed to improve performance on such complex, multi-step, and multi-modal tasks. During inference, TACO produces chains-of-thought-and-action (CoTA), executes intermediate steps by invoking external tools such as OCR, depth estimation and calculator, then integrates both the thoughts and action outputs to produce coherent responses. To train TACO, we create a large dataset of over 1M synthetic CoTA traces generated with GPT-4o and Python programs. We then experiment with various data filtering and mixing techniques and obtain a final subset of 293K high-quality CoTA examples. This dataset enables TACO to learn complex reasoning and action paths, surpassing existing models trained on instruction tuning data with only direct answers. Our model TACO outperforms the instruction-tuned baseline across 8 benchmarks, achieving a 3.6% improvement on average, with gains of up to 15% in MMVet tasks involving OCR, mathematical reasoning, and spatial reasoning. Training on high-quality CoTA traces sets a new standard for complex multi-modal reasoning, highlighting the need for structured, multi-step instruction tuning in advancing open-source mutli-modal models' capabilities.

ProVision: Programmatically Scaling Vision-centric Instruction Data for Multimodal Language Models

With the rise of multimodal applications, instruction data has become critical for training multimodal language models capable of understanding complex image-based queries. Existing practices rely on powerful but costly large language models (LLMs) or multimodal language models (MLMs) to produce instruction data. These are often prone to hallucinations, licensing issues and the generation process is often hard to scale and interpret. In this work, we present a programmatic approach that employs scene graphs as symbolic representations of images and human-written programs to systematically synthesize vision-centric instruction data. Our approach ensures the interpretability and controllability of the data generation process and scales efficiently while maintaining factual accuracy. By implementing a suite of 24 single-image, 14 multi-image instruction generators, and a scene graph generation pipeline, we build a scalable, cost-effective system: ProVision which produces diverse question-answer pairs concerning objects, attributes, relations, depth, etc., for any given image. Applied to Visual Genome and DataComp datasets, we generate over 10 million instruction data points, ProVision-10M, and leverage them in both pretraining and instruction tuning stages of MLMs. When adopted in the instruction tuning stage, our single-image instruction data yields up to a 7% improvement on the 2D split and 8% on the 3D split of CVBench, along with a 3% increase in performance on QBench2, RealWorldQA, and MMMU. Our multi-image instruction data leads to an 8% improvement on Mantis-Eval. Incorporation of our data in both pre-training and fine-tuning stages of xGen-MM-4B leads to an averaged improvement of 1.6% across 11 benchmarks.

EcoAct: Economic Agent Determines When to Register What Action

Recent advancements have enabled Large Language Models (LLMs) to function as agents that can perform actions using external tools. This requires registering, i.e., integrating tool information into the LLM context prior to taking actions. Current methods indiscriminately incorporate all candidate tools into the agent's context and retain them across multiple reasoning steps. This process remains opaque to LLM agents and is not integrated into their reasoning procedures, leading to inefficiencies due to increased context length from irrelevant tools. To address this, we introduce EcoAct, a tool using algorithm that allows LLMs to selectively register tools as needed, optimizing context use. By integrating the tool registration process into the reasoning procedure, EcoAct reduces computational costs by over 50% in multiple steps reasoning tasks while maintaining performance, as demonstrated through extensive experiments. Moreover, it can be plugged into any reasoning pipeline with only minor modifications to the prompt, making it applicable to LLM agents now and future.

Language Model Preference Evaluation with Multiple Weak Evaluators

Despite the remarkable success of Large Language Models (LLMs), evaluating their outputs' quality regarding preference remains a critical challenge. Existing works usually leverage a powerful LLM (e.g., GPT4) as the judge for comparing LLMs' output pairwisely, yet such model-based evaluator is vulnerable to conflicting preference, i.e., output A is better than B, B than C, but C than A, causing contradictory evaluation results. To improve model-based preference evaluation, we introduce GED (Preference Graph Ensemble and Denoise), a novel approach that leverages multiple model-based evaluators to construct preference graphs, and then ensemble and denoise these graphs for better, non-contradictory evaluation results. In particular, our method consists of two primary stages: aggregating evaluations into a unified graph and applying a denoising process to eliminate cyclic inconsistencies, ensuring a directed acyclic graph (DAG) structure. We provide theoretical guarantees for our framework, demonstrating its efficacy in recovering the ground truth preference structure. Extensive experiments across ten benchmark datasets show that GED outperforms baseline methods in model ranking, response selection, and model alignment tasks. Notably, GED combines weaker evaluators like Llama3-8B, Mistral-7B, and Qwen2-7B to surpass the performance of stronger evaluators like Qwen2-72B, highlighting its ability to enhance evaluation reliability and improve model performance.

xGen-MM (BLIP-3): A Family of Open Large Multimodal Models

This report introduces xGen-MM (also known as BLIP-3), a framework for developing Large Multimodal Models (LMMs). The framework comprises meticulously curated datasets, a training recipe, model architectures, and a resulting suite of LMMs. xGen-MM, short for xGen-MultiModal, expands the Salesforce xGen initiative on foundation AI models. Our models undergo rigorous evaluation across a range of tasks, including both single and multi-image benchmarks. Our pre-trained base model exhibits strong in-context learning capabilities and the instruction-tuned model demonstrates competitive performance among open-source LMMs with similar model sizes. In addition, we introduce a safety-tuned model with DPO, aiming to mitigate harmful behaviors such as hallucinations and improve safety. We open-source our models, curated large-scale datasets, and our fine-tuning codebase to facilitate further advancements in LMM research. Associated resources will be available on our project page above.