ICML 2024

Recent advancements in foundation models have improved autonomous tool usage and reasoning, but their capabilities in map-based reasoning remain underexplored. To address this, we introduce MapEval, a benchmark designed to assess foundation models across three distinct tasks - textual, API-based, and visual reasoning - through 700 multiple-choice questions spanning 180 cities and 54 countries, covering spatial relationships, navigation, travel planning, and real-world map interactions. Unlike prior benchmarks that focus on simple location queries, MapEval requires models to handle long-context reasoning, API interactions, and visual map analysis, making it the most comprehensive evaluation framework for geospatial AI. On evaluation of 30 foundation models, including Claude-3.5-Sonnet, GPT-4o, and Gemini-1.5-Pro, none surpass 67% accuracy, with open-source models performing significantly worse and all models lagging over 20% behind human performance. These results expose critical gaps in spatial inference, as models struggle with distances, directions, route planning, and place-specific reasoning, highlighting the need for better geospatial AI to bridge the gap between foundation models and real-world navigation. All the resources are available at: https://mapeval.github.io/.

This paper addresses the scarcity of large-scale datasets for accurate object-in-hand pose estimation, which is crucial for robotic in-hand manipulation within the ``Perception-Planning-Control"paradigm. Specifically, we introduce VinT-6D, the first extensive multi-modal dataset integrating vision, touch, and proprioception, to enhance robotic manipulation. VinT-6D comprises 2 million VinT-Sim and 0.1 million VinT-Real splits, collected via simulations in MuJoCo and Blender and a custom-designed real-world platform. This dataset is tailored for robotic hands, offering models with whole-hand tactile perception and high-quality, well-aligned data. To the best of our knowledge, the VinT-Real is the largest considering the collection difficulties in the real-world environment so that it can bridge the gap of simulation to real compared to the previous works. Built upon VinT-6D, we present a benchmark method that shows significant improvements in performance by fusing multi-modal information. The project is available at https://VinT-6D.github.io/.

Large language models (LLMs) are empowering decision-making in several applications, including tool or API usage and answering multiple-choice questions (MCQs). However, incorrect outputs pose significant risks in high-stakes domains like healthcare and finance. To quantify LLM uncertainty and thereby mitigate these risks, recent works employ conformal prediction (CP), a model- and distribution-agnostic framework that uses LLM outputs to generate a \emph{prediction set} containing the true answer with high probability. Leveraging CP, we propose \emph{conformal revision of questions} (CROQ), which revises the question by narrowing down the available choices to those in the prediction set and asking the LLM the revised question. We expect LLMs to be more accurate on revised questions with fewer choices. Furthermore, we expect CROQ to be effective when the prediction sets from CP are small. Commonly used logit scores often lead to large sets, diminishing CROQ's effectiveness. To overcome this, we propose CP-OPT, an optimization framework to learn scores that minimize set sizes while maintaining coverage. Our extensive experiments on MMLU, ToolAlpaca, and TruthfulQA datasets with multiple LLMs show that CROQ improves accuracy over the standard inference, with more pronounced gains when paired with CP-OPT.

Despite recent advances in Vision-Language Models (VLMs), they may over-rely on visual language priors existing in their training data rather than true visual reasoning. To investigate this, we introduce ViLP, a benchmark featuring deliberately out-of-distribution images synthesized via image generation models and out-of-distribution Q&A pairs. Each question in ViLP is coupled with three potential answers and three corresponding images: one that can be resolved by text priors alone and two that demand visual reasoning. Although, humans achieve near-perfect accuracy, modern VLMs falter; for instance, GPT-4 achieves only 66.17% on ViLP. To alleviate this, we propose a self-improving framework in which models generate new VQA data, then apply pixel-level and semantic corruptions to form"good-bad"image pairs for self-training. Our training objectives compel VLMs to focus more on the actual visual inputs, and we demonstrate their effectiveness in boosting the performance of open-source VLMs, including LLaVA-v1.5 and Cambrian.

Existing research of video understanding still struggles to achieve in-depth comprehension and reasoning in complex videos, primarily due to the under-exploration of two key bottlenecks: fine-grained spatial-temporal perceptive understanding and cognitive-level video scene comprehension. This paper bridges the gap by presenting a novel solution. We first introduce a novel video Multimodal Large Language Model (MLLM), MotionEpic, which achieves fine-grained pixel-level spatial-temporal video grounding by integrating video spatial-temporal scene graph (STSG) representation. Building upon MotionEpic, we then develop a Video-of-Thought (VoT) reasoning framework. VoT inherits the Chain-of-Thought (CoT) core, breaking down a complex task into simpler and manageable sub-problems, and addressing them step-by-step from a low-level pixel perception to high-level cognitive interpretation. Extensive experiments across various complex video QA benchmarks demonstrate that our overall framework strikingly boosts existing state-of-the-art. To our knowledge, this is the first attempt at successfully implementing the CoT technique for achieving human-level video reasoning, where we show great potential in extending it to a wider range of video understanding scenarios. Project is open at https://haofei.vip/VoT

Large Multimodal Models (LMMs) extend Large Language Models (LLMs) by handling diverse inputs such as images, audio, and video, but at the cost of adding a multimodal encoding stage that increases both computational and memory overhead. This step negatively affects key Service Level Objectives (SLOs), such as time to first token (TTFT) and time per output token (TPOT). We introduce Encode-Prefill-Decode (EPD) Disaggregation, a novel framework that separates the encoding, prefill, and decode stages onto dedicated resources. Unlike current systems, which bundle encoding and prefill together, our approach decouples these steps, unlocking new opportunities and optimizations. These include a mechanism to cache multimedia tokens for efficient transfer, a novel way to parallelize the encoding load within a request, a module for optimal resource allocation for disaggregated serving, and a novel role-switching method to handle changing workload characteristics. Experimental evaluations with popular LMMs show substantial gains in memory efficiency (up to 15x lower peak memory utilization), batch sizes (up to 22x larger), 10x more images per request, and 2.2x larger KV caches. Furthermore, it leads to significant improvements in SLO attainment (up to 90-100% improvement) and TTFT (up to 71% reduction), compared to systems that do not disaggregate. The code is available at https://github.com/vbdi/epdserve.

Large language models (LLMs) have been routinely used to solve various tasks using step-by-step reasoning. However, the structure of intermediate reasoning steps, or thoughts, is rigid and unidirectional, such as chains, trees, or acyclic-directed graphs. Consequently, the resulting inflexible and forward-only reasoning may not address challenging tasks and fail when the LLM frequently gives false responses, i.e., ``hallucinations''. This paper proposes a new reasoning framework, called Thought Rollback (TR), allowing LLMs to adaptively build thought structure while maintaining effective reasoning toward problem-solving under ``hallucinations''. The core mechanism of TR is rolling back thoughts, which allows LLMs to perform error analysis on thoughts, and thus roll back to any previously mistaken thought for revision. Subsequently, by including such trial-and-error in the prompt to guide the LLM, each rollback leads to one more reliable reasoning path. Therefore, starting with a simple prompt without human annotations, LLM with TR adaptively and gradually explores thoughts for a correct solution. Comprehensive experiments on mathematical problems and multi-task reasoning demonstrate the state-of-the-art performance of TR in terms of problem-solving rate and interaction cost. For instance, the solving rate of GPT-4 with TR outperforms the current best by $9\%$ on the MATH dataset.

Conformal prediction is a valuable tool for quantifying predictive uncertainty of machine learning models. However, its applicability relies on the assumption of data exchangeability, a condition which is often not met in real-world scenarios. In this paper, we consider the problem of adaptive conformal inference without any assumptions about the data generating process. Existing approaches for adaptive conformal inference are based on optimizing the pinball loss using variants of online gradient descent. A notable shortcoming of such approaches is in their explicit dependence on and sensitivity to the choice of the learning rates. In this paper, we propose a different approach for adaptive conformal inference that leverages parameter-free online convex optimization techniques. We prove that our method controls long-term miscoverage frequency at a nominal level and demonstrate its convincing empirical performance without any need of performing cumbersome parameter tuning.

Recent years have seen significant progress in Text-To-Audio (TTA) synthesis, enabling users to enrich their creative workflows with synthetic audio generated from natural language prompts. Despite this progress, the effects of data, model architecture, training objective functions, and sampling strategies on target benchmarks are not well understood. With the purpose of providing a holistic understanding of the design space of TTA models, we set up a large-scale empirical experiment focused on diffusion and flow matching models. Our contributions include: 1) AF-Synthetic, a large dataset of high quality synthetic captions obtained from an audio understanding model; 2) a systematic comparison of different architectural, training, and inference design choices for TTA models; 3) an analysis of sampling methods and their Pareto curves with respect to generation quality and inference speed. We leverage the knowledge obtained from this extensive analysis to propose our best model dubbed Elucidated Text-To-Audio (ETTA). When evaluated on AudioCaps and MusicCaps, ETTA provides improvements over the baselines trained on publicly available data, while being competitive with models trained on proprietary data. Finally, we show ETTA's improved ability to generate creative audio following complex and imaginative captions -- a task that is more challenging than current benchmarks.

We define the local complexity of a neural network with continuous piecewise linear activations as a measure of the density of linear regions over an input data distribution. We show theoretically that ReLU networks that learn low-dimensional feature representations have a lower local complexity. This allows us to connect recent empirical observations on feature learning at the level of the weight matrices with concrete properties of the learned functions. In particular, we show that the local complexity serves as an upper bound on the total variation of the function over the input data distribution and thus that feature learning can be related to adversarial robustness. Lastly, we consider how optimization drives ReLU networks towards solutions with lower local complexity. Overall, this work contributes a theoretical framework towards relating geometric properties of ReLU networks to different aspects of learning such as feature learning and representation cost.

Shock waves caused by earthquakes can be devastating. Generating realistic earthquake-caused ground motion waveforms help reducing losses in lives and properties, yet generative models for the task tend to generate subpar waveforms. We present High-fidelity Earthquake Groundmotion Generation System (HEGGS) and demonstrate its superior performance using earthquakes from North American, East Asian, and European regions. HEGGS exploits the intrinsic characteristics of earthquake dataset and learns the waveforms using an end-to-end differentiable generator containing conditional latent diffusion model and hi-fidelity waveform construction model. We show the learning efficiency of HEGGS by training it on a single GPU machine and validate its performance using earthquake databases from North America, East Asia, and Europe, using diverse criteria from waveform generation tasks and seismology. Once trained, HEGGS can generate three dimensional E-N-Z seismic waveforms with accurate P/S phase arrivals, envelope correlation, signal-to-noise ratio, GMPE analysis, frequency content analysis, and section plot analysis.

Self-evolving trainin--where models iteratively learn from their own outputs--has emerged as a key approach for complex reasoning tasks, addressing the scarcity of high-quality chain-of-thought data. However, its effectiveness in multimodal reasoning, a domain more intricate than text-only reasoning, remains underexplored, and the understanding of critical factors in this training paradigm remains limited. Furthermore, a central challenge for this training method is performance saturation, which impedes further improvements and scalability. Inspired by reinforcement learning (RL), in this paper, we reframe self-evolving training for multimodal reasoning through the lens of RL, identifying three pivotal factors: Training Method, Reward Model, and Prompt Variation. Through systematic analysis, we establish relatively optimal design principles that significantly enhance multimodal reasoning capabilities. Moreover, delving deeper into training dynamics, we uncover the roots of saturation and propose a new automatic balancing mechanism to mitigate this limitation. Building on these insights, we propose M-STAR (Multimodal Self-evolving Training for Reasoning), a framework that achieves consistent performance gains across models of varying sizes and diverse benchmarks. All resources are made publicly available at https://mstar-lmm.github.io.

Large Language Models (LLMs) have demonstrated remarkable potential in scientific domains, yet a fundamental question remains unanswered: Can we simulate human research communities with LLMs? Addressing this question can deepen our understanding of the processes behind idea brainstorming and inspire the automatic discovery of novel scientific insights. In this work, we propose ResearchTown, a multi-agent framework for research community simulation. Within this framework, the human research community is simplified as an agent-data graph, where researchers and papers are represented as agent-type and data-type nodes, respectively, and connected based on their collaboration relationships. We also introduce TextGNN, a text-based inference framework that models various research activities (e.g., paper reading, paper writing, and review writing) as special forms of a unified message-passing process on the agent-data graph. To evaluate the quality of the research community simulation, we present ResearchBench, a benchmark that uses a node-masking prediction task for scalable and objective assessment based on similarity. Our experiments reveal three key findings: (1) ResearchTown can provide a realistic simulation of collaborative research activities, including paper writing and review writing; (2) ResearchTown can maintain robust simulation with multiple researchers and diverse papers; (3) ResearchTown can generate interdisciplinary research ideas that potentially inspire pioneering research directions.

We study partially observable assistance games (POAGs), a model of the human-AI value alignment problem which allows the human and the AI assistant to have partial observations. Motivated by concerns of AI deception, we study a qualitatively new phenomenon made possible by partial observability: would an AI assistant ever have an incentive to interfere with the human's observations? First, we prove that sometimes an optimal assistant must take observation-interfering actions, even when the human is playing optimally, and even when there are otherwise-equivalent actions available that do not interfere with observations. Though this result seems to contradict the classic theorem from single-agent decision making that the value of information is nonnegative, we resolve this seeming contradiction by developing a notion of interference defined on entire policies. This can be viewed as an extension of the classic result that the value of information is nonnegative into the cooperative multiagent setting. Second, we prove that if the human is simply making decisions based on their immediate outcomes, the assistant might need to interfere with observations as a way to query the human's preferences. We show that this incentive for interference goes away if the human is playing optimally, or if we introduce a communication channel for the human to communicate their preferences to the assistant. Third, we show that if the human acts according to the Boltzmann model of irrationality, this can create an incentive for the assistant to interfere with observations. Finally, we use an experimental model to analyze tradeoffs faced by the AI assistant in practice when considering whether or not to take observation-interfering actions.

Foundation models are pretrained on large-scale corpora to learn generalizable patterns across domains and tasks -- such as contours, textures, and edges in images, or tokens and sentences in text. In contrast, discovering such generalities in graph-structured data, especially across heterogeneous graph tasks, remains an open challenge. To address this, we propose a novel approach to cross-task generalization in graphs via task-trees, which serve as unified learning instances aligning node-, edge-, and graph-level tasks. We theoretically analyze the stability, transferability, and generalization properties of task-trees, showing that pretraining a graph neural network (GNN) on diverse task-trees with a reconstruction objective induces transferable knowledge. This enables efficient adaptation to downstream tasks with minimal fine-tuning. To validate our framework, we introduce Graph Generality Identifier on Task-Trees (GIT), a graph foundation model that demonstrates strong performance on over 30 graphs across five domains via fine-tuning, in-context learning, and zero-shot generalization. Code and data are available at https://github.com/Zehong-Wang/GIT.

We study the repeated principal-agent bandit game, where the principal indirectly interacts with the unknown environment by proposing incentives for the agent to play arms. Most existing work assumes the agent has full knowledge of the reward means and always behaves greedily, but in many online marketplaces, the agent needs to learn the unknown environment and sometimes explore. Motivated by such settings, we model a self-interested learning agent with exploration behaviors who iteratively updates reward estimates and either selects an arm that maximizes the estimated reward plus incentive or explores arbitrarily with a certain probability. As a warm-up, we first consider a self-interested learning agent without exploration. We propose algorithms for both i.i.d. and linear reward settings with bandit feedback in a finite horizon $T$, achieving regret bounds of $\widetilde{O}(\sqrt{T})$ and $\widetilde{O}( T^{2/3} )$, respectively. Specifically, these algorithms are established upon a novel elimination framework coupled with newly-developed search algorithms which accommodate the uncertainty arising from the learning behavior of the agent. We then extend the framework to handle the exploratory learning agent and develop an algorithm to achieve a $\widetilde{O}(T^{2/3})$ regret bound in i.i.d. reward setup by enhancing the robustness of our elimination framework to the potential agent exploration. Finally, when reducing our agent behaviors to the one studied in (Dogan et al., 2023a), we propose an algorithm based on our robust framework, which achieves a $\widetilde{O}(\sqrt{T})$ regret bound, significantly improving upon their $\widetilde{O}(T^{11/12})$ bound.

Model collapse in synthetic data indicates that iterative training on self-generated data leads to a gradual decline in performance. With the proliferation of AI models, synthetic data will fundamentally reshape the web data ecosystem. Future GPT-$\{n\}$ models will inevitably be trained on a blend of synthetic and human-produced data. In this paper, we focus on two questions: what is the impact of synthetic data on language model training, and how to synthesize data without model collapse? We first pre-train language models across different proportions of synthetic data, revealing a negative correlation between the proportion of synthetic data and model performance. We further conduct statistical analysis on synthetic data to uncover distributional shift phenomenon and over-concentration of n-gram features. Inspired by the above findings, we propose token editing on human-produced data to obtain semi-synthetic data. As a proof of concept, we theoretically demonstrate that token-level editing can prevent model collapse, as the test error is constrained by a finite upper bound. We conduct extensive experiments on pre-training from scratch, continual pre-training, and supervised fine-tuning. The results validate our theoretical proof that token-level editing improves model performance.

Large Language Models excel at code generation yet struggle with complex programming tasks that demand sophisticated reasoning. To bridge this gap, traditional process supervision relies on learned reward models requiring costly training data and suffering from reward misalignment, while outcome supervision fails for complex tasks needing coordinated intermediate steps. We introduce Outcome Refining Process Supervision, which unifies process and outcome supervision by leveraging executable verification: a tree-structured search framework generates strategic alternatives, profiles execution metrics, and scores candidates via self-critique mechanisms that integrate runtime feedback with reasoning. Experiments across 5 models and 3 benchmarks show consistent gains, with 26.9% higher correctness and 42.2% improved code efficiency. The results demonstrate that ORPS enables LLMs to overcome local optima in code generation, suggesting a promising direction for combining verifiable outcomes with structured reasoning to tackle complex challenges. We open-source at: https://github.com/zhuohaoyu/ORPS

We delve into the challenge of semi-supervised node classification on the Contextual Stochastic Block Model (CSBM) dataset. Here, nodes from the two-cluster Stochastic Block Model (SBM) are coupled with feature vectors, which are derived from a Gaussian Mixture Model (GMM) that corresponds to their respective node labels. With only a subset of the CSBM node labels accessible for training, our primary objective becomes the accurate classification of the remaining nodes. Venturing into the transductive learning landscape, we, for the first time, pinpoint the information-theoretical threshold for the exact recovery of all test nodes in CSBM. Concurrently, we design an optimal spectral estimator inspired by Principal Component Analysis (PCA) with the training labels and essential data from both the adjacency matrix and feature vectors. We also evaluate the efficacy of graph ridge regression and Graph Convolutional Networks (GCN) on this synthetic dataset. Our findings underscore that graph ridge regression and GCN possess the ability to achieve the information threshold of exact recovery in a manner akin to the optimal estimator when using the optimal weighted self-loops. This highlights the potential role of feature learning in augmenting the proficiency of GCN, especially in the realm of semi-supervised learning.

Continual Federated Learning (CFL) allows distributed devices to collaboratively learn novel concepts from continuously shifting training data while avoiding knowledge forgetting of previously seen tasks. To tackle this challenge, most current CFL approaches rely on extensive rehearsal of previous data. Despite effectiveness, rehearsal comes at a cost to memory, and it may also violate data privacy. Considering these, we seek to apply regularization techniques to CFL by considering their cost-efficient properties that do not require sample caching or rehearsal. Specifically, we first apply traditional regularization techniques to CFL and observe that existing regularization techniques, especially synaptic intelligence, can achieve promising results under homogeneous data distribution but fail when the data is heterogeneous. Based on this observation, we propose a simple yet effective regularization algorithm for CFL named FedSSI, which tailors the synaptic intelligence for the CFL with heterogeneous data settings. FedSSI can not only reduce computational overhead without rehearsal but also address the data heterogeneity issue. Extensive experiments show that FedSSI achieves superior performance compared to state-of-the-art methods.