Jesse Thomason

Ishika Singh, Valts Blukis, Arsalan Mousavian et al. · gatech, nvidia

Task planning can require defining myriad domain knowledge about the world in which a robot needs to act. To ameliorate that effort, large language models (LLMs) can be used to score potential next actions during task planning, and even generate action sequences directly, given an instruction in natural language with no additional domain information. However, such methods either require enumerating all possible next steps for scoring, or generate free-form text that may contain actions not possible on a given robot in its current context. We present a programmatic LLM prompt structure that enables plan generation functional across situated environments, robot capabilities, and tasks. Our key insight is to prompt the LLM with program-like specifications of the available actions and objects in an environment, as well as with example programs that can be executed. We make concrete recommendations about prompt structure and generation constraints through ablation experiments, demonstrate state of the art success rates in VirtualHome household tasks, and deploy our method on a physical robot arm for tabletop tasks. Website at progprompt.github.io

Matt Deitke, Dhruv Batra, Yonatan Bisk et al. · allen-ai, cmu

We present a retrospective on the state of Embodied AI research. Our analysis focuses on 13 challenges presented at the Embodied AI Workshop at CVPR. These challenges are grouped into three themes: (1) visual navigation, (2) rearrangement, and (3) embodied vision-and-language. We discuss the dominant datasets within each theme, evaluation metrics for the challenges, and the performance of state-of-the-art models. We highlight commonalities between top approaches to the challenges and identify potential future directions for Embodied AI research.

Ting-Yun Chang, Harvey Yiyun Fu, Deqing Fu et al.

Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.

Vishnu Sashank Dorbala, Gunnar Sigurdsson, Robinson Piramuthu et al. · uw

Household environments are visually diverse. Embodied agents performing Vision-and-Language Navigation (VLN) in the wild must be able to handle this diversity, while also following arbitrary language instructions. Recently, Vision-Language models like CLIP have shown great performance on the task of zero-shot object recognition. In this work, we ask if these models are also capable of zero-shot language grounding. In particular, we utilize CLIP to tackle the novel problem of zero-shot VLN using natural language referring expressions that describe target objects, in contrast to past work that used simple language templates describing object classes. We examine CLIP's capability in making sequential navigational decisions without any dataset-specific finetuning, and study how it influences the path that an agent takes. Our results on the coarse-grained instruction following task of REVERIE demonstrate the navigational capability of CLIP, surpassing the supervised baseline in terms of both success rate (SR) and success weighted by path length (SPL). More importantly, we quantitatively show that our CLIP-based zero-shot approach generalizes better to show consistent performance across environments when compared to SOTA, fully supervised learning approaches when evaluated via Relative Change in Success (RCS).

Allen Chang, Xiaoyuan Zhu, Aarav Monga et al. · uw

Benchmarks for language-guided embodied agents typically assume text-based instructions, but deployed agents will encounter spoken instructions. While Automatic Speech Recognition (ASR) models can bridge the input gap, erroneous ASR transcripts can hurt the agents' ability to complete tasks. In this work, we propose training a multimodal ASR model to reduce errors in transcribing spoken instructions by considering the accompanying visual context. We train our model on a dataset of spoken instructions, synthesized from the ALFRED task completion dataset, where we simulate acoustic noise by systematically masking spoken words. We find that utilizing visual observations facilitates masked word recovery, with multimodal ASR models recovering up to 30% more masked words than unimodal baselines. We also find that a text-trained embodied agent successfully completes tasks more often by following transcribed instructions from multimodal ASR models. github.com/Cylumn/embodied-multimodal-asr

Georgios Chochlakis, Tejas Srinivasan, Jesse Thomason et al. · uw

We propose the Vision-and-Augmented-Language Transformer (VAuLT). VAuLT is an extension of the popular Vision-and-Language Transformer (ViLT), and improves performance on vision-and-language (VL) tasks that involve more complex text inputs than image captions while having minimal impact on training and inference efficiency. ViLT, importantly, enables efficient training and inference in VL tasks, achieved by encoding images using a linear projection of patches instead of an object detector. However, it is pretrained on captioning datasets, where the language input is simple, literal, and descriptive, therefore lacking linguistic diversity. So, when working with multimedia data in the wild, such as multimodal social media data, there is a notable shift from captioning language data, as well as diversity of tasks. We indeed find evidence that the language capacity of ViLT is lacking. The key insight and novelty of VAuLT is to propagate the output representations of a large language model (LM) like BERT to the language input of ViLT. We show that joint training of the LM and ViLT can yield relative improvements up to 20% over ViLT and achieve state-of-the-art or comparable performance on VL tasks involving richer language inputs and affective constructs, such as for Target-Oriented Sentiment Classification in TWITTER-2015 and TWITTER-2017, and Sentiment Classification in MVSA-Single and MVSA-Multiple. Our code is available at https://github.com/gchochla/VAuLT.

Gunnar A. Sigurdsson, Jesse Thomason, Gaurav S. Sukhatme et al. · uw

Household robots operate in the same space for years. Such robots incrementally build dynamic maps that can be used for tasks requiring remote object localization. However, benchmarks in robot learning often test generalization through inference on tasks in unobserved environments. In an observed environment, locating an object is reduced to choosing from among all object proposals in the environment, which may number in the 100,000s. Armed with this intuition, using only a generic vision-language scoring model with minor modifications for 3d encoding and operating in an embodied environment, we demonstrate an absolute performance gain of 9.84% on remote object grounding above state of the art models for REVERIE and of 5.04% on FAO. When allowed to pre-explore an environment, we also exceed the previous state of the art pre-exploration method on REVERIE. Additionally, we demonstrate our model on a real-world TurtleBot platform, highlighting the simplicity and usefulness of the approach. Our analysis outlines a "bag of tricks" essential for accomplishing this task, from utilizing 3d coordinates and context, to generalizing vision-language models to large 3d search spaces.

Jing Gu, Eliana Stefani, Qi Wu et al. · uw

A long-term goal of AI research is to build intelligent agents that can communicate with humans in natural language, perceive the environment, and perform real-world tasks. Vision-and-Language Navigation (VLN) is a fundamental and interdisciplinary research topic towards this goal, and receives increasing attention from natural language processing, computer vision, robotics, and machine learning communities. In this paper, we review contemporary studies in the emerging field of VLN, covering tasks, evaluation metrics, methods, etc. Through structured analysis of current progress and challenges, we highlight the limitations of current VLN and opportunities for future work. This paper serves as a thorough reference for the VLN research community.

Xiangtong Yao, Hongkuan Zhou, Oier Mees et al. · cmu

Language-conditioned robot manipulation is an emerging field aimed at enabling seamless communication and cooperation between humans and robotic agents by teaching robots to comprehend and execute instructions conveyed in natural language. This interdisciplinary area integrates scene understanding, language processing, and policy learning to bridge the gap between human instructions and robot actions. In this comprehensive survey, we systematically explore recent advancements in language-conditioned robot manipulation. We categorize existing methods based on the primary ways language is integrated into the robot system, namely language for state evaluation, language as a policy condition, language for cognitive planning and reasoning, and language in unified vision-language-action models. Specifically, we further analyze state-of-the-art techniques from five axes of action granularity, data and supervision regimes, system cost and latency, environments and evaluations, and cross-modal task specification. Additionally, we highlight the key debates in the field. Finally, we discuss open challenges and future research directions, focusing on potentially enhancing generalization capabilities and addressing safety issues in language-conditioned robot manipulators.

Jeremy Morgan, Prajwal Vijay, Hyeonho Oh et al.

Vision-Language-Action (VLA) models demonstrate promising generalization in robotic manipulation, driven by advances in large-scale vision and language pre-training. This progress can be misleading. Despite the zero-shot perception and language capabilities of VLAs, their overall task performance often degrades under distribution shifts, revealing gaps in how these systems translate high-level understanding into robust behavior. To systematically study this gap, we introduce Colosseum V2, a large-scale simulation benchmark for evaluating VLA generalization in robot learning across diverse conditions. The benchmark comprises 28 tasks spanning 13 task categories and two robot morphologies, covering a wide range of manipulation primitives and long-horizon behaviors. Built on the ManiSkill simulator, Colosseum V2 enables fast, GPU-parallelized evaluation and supports both in-domain and out-of-domain testing at scale. We evaluate state-of-the-art methods, including Action Chunking Transformers (ACT) and Pi0.5, and reveal limitations in both base performance and generalization. We demonstrate strong correlations between simulation and real-world metrics that support the ecological validity of the benchmark. By standardizing tasks, metrics, and evaluation protocols within a unified benchmark, Colosseum V2 enables reproducible and fair comparisons, reduced evaluation overhead, and accelerated progress toward general-purpose robot policies.

Jacob Krantz, Shurjo Banerjee, Wang Zhu et al. · uw

We present Iterative Vision-and-Language Navigation (IVLN), a paradigm for evaluating language-guided agents navigating in a persistent environment over time. Existing Vision-and-Language Navigation (VLN) benchmarks erase the agent's memory at the beginning of every episode, testing the ability to perform cold-start navigation with no prior information. However, deployed robots occupy the same environment for long periods of time. The IVLN paradigm addresses this disparity by training and evaluating VLN agents that maintain memory across tours of scenes that consist of up to 100 ordered instruction-following Room-to-Room (R2R) episodes, each defined by an individual language instruction and a target path. We present discrete and continuous Iterative Room-to-Room (IR2R) benchmarks comprising about 400 tours each in 80 indoor scenes. We find that extending the implicit memory of high-performing transformer VLN agents is not sufficient for IVLN, but agents that build maps can benefit from environment persistence, motivating a renewed focus on map-building agents in VLN.

Tejas Srinivasan, Ting-Yun Chang, Leticia Leonor Pinto Alva et al. · uw

Current state-of-the-art vision-and-language models are evaluated on tasks either individually or in a multi-task setting, overlooking the challenges of continually learning (CL) tasks as they arrive. Existing CL benchmarks have facilitated research on task adaptation and mitigating "catastrophic forgetting", but are limited to vision-only and language-only tasks. We present CLiMB, a benchmark to study the challenge of learning multimodal tasks in a CL setting, and to systematically evaluate how upstream continual learning can rapidly generalize to new multimodal and unimodal tasks. CLiMB includes implementations of several CL algorithms and a modified Vision-Language Transformer (ViLT) model that can be deployed on both multimodal and unimodal tasks. We find that common CL methods can help mitigate forgetting during multimodal task learning, but do not enable cross-task knowledge transfer. We envision that CLiMB will facilitate research on a new class of CL algorithms for this challenging multimodal setting.

Tejas Srinivasan, Xiang Ren, Jesse Thomason · uw

Aligning image and text encoders from scratch using contrastive learning requires large amounts of paired image-text data. We alleviate this need by aligning individually pre-trained language and vision representation models using a much smaller amount of paired data, augmented with a curriculum learning algorithm to learn fine-grained vision-language alignments. TOnICS (Training with Ontology-Informed Contrastive Sampling) initially samples minibatches whose image-text pairs contain a wide variety of objects to learn object-level alignment, and progressively samples minibatches where all image-text pairs contain the same object to learn finer-grained contextual alignment. Aligning pre-trained BERT and VinVL models to each other using TOnICS outperforms CLIP on downstream zero-shot image retrieval while using less than 1% as much training data.

Wang Zhu, Alekh Agarwal, Mandar Joshi et al. · uw

Understanding visually situated language requires interpreting complex layouts of textual and visual elements. Pre-processing tools, such as optical character recognition (OCR), can map document image inputs to textual tokens, then large language models (LLMs) can reason over text. However, such methods have high computational and engineering complexity. Can small pretrained image-to-text models accurately understand visual documents through similar recognition and reasoning steps instead? We propose Rationale Distillation (RD), which incorporates the outputs of OCR tools, LLMs, and larger multimodal models as intermediate "rationales", and trains a small student model to predict both rationales and answers. On three visual document understanding benchmarks representing infographics, scanned documents, and figures, our Pix2Struct (282M parameters) student model finetuned with RD outperforms the base model by 4-5% absolute accuracy with only 1% higher computational cost.

Yuliang Cai, Jesse Thomason, Mohammad Rostami · uw

The size and the computational load of fine-tuning large-scale pre-trained neural network are becoming two major obstacles in adopting machine learning in many applications. Continual learning (CL) can serve as a remedy through enabling knowledge-transfer across sequentially arriving tasks which relaxes the need to fine-tune all network weights from scratch. However, existing CL algorithms primarily consider learning unimodal vision-only or language-only tasks. We develop a transformer-based CL architecture for learning bimodal vision-and-language tasks based on increasing the number of the learnable parameters dynamically and using knowledge distillation. The new additional parameters are used to specialize the network for each task. Our approach enables sharing information between the tasks while addressing the challenge of catastrophic forgetting. Our approach is scalable learning to a large number of tasks because it requires little memory and time overhead. Our model reaches state-of-the-art performance on challenging vision-and-language tasks.

Lee Kezar, Elana Pontecorvo, Adele Daniels et al.

Sign language recognition and translation technologies have the potential to increase access and inclusion of deaf signing communities, but research progress is bottlenecked by a lack of representative data. We introduce a new resource for American Sign Language (ASL) modeling, the Sem-Lex Benchmark. The Benchmark is the current largest of its kind, consisting of over 84k videos of isolated sign productions from deaf ASL signers who gave informed consent and received compensation. Human experts aligned these videos with other sign language resources including ASL-LEX, SignBank, and ASL Citizen, enabling useful expansions for sign and phonological feature recognition. We present a suite of experiments which make use of the linguistic information in ASL-LEX, evaluating the practicality and fairness of the Sem-Lex Benchmark for isolated sign recognition (ISR). We use an SL-GCN model to show that the phonological features are recognizable with 85% accuracy, and that they are effective as an auxiliary target to ISR. Learning to recognize phonological features alongside gloss results in a 6% improvement for few-shot ISR accuracy and a 2% improvement for ISR accuracy overall. Instructions for downloading the data can be found at https://github.com/leekezar/SemLex.

Tejas Srinivasan, Furong Jia, Mohammad Rostami et al. · uw

Adapters present a promising solution to the catastrophic forgetting problem in continual learning. However, training independent Adapter modules for every new task misses an opportunity for cross-task knowledge transfer. We propose Improvise to Initialize (I2I), a continual learning algorithm that initializes Adapters for incoming tasks by distilling knowledge from previously-learned tasks' Adapters. We evaluate I2I on CLiMB, a multimodal continual learning benchmark, by conducting experiments on sequences of visual question answering tasks. Adapters trained with I2I consistently achieve better task accuracy than independently-trained Adapters, demonstrating that our algorithm facilitates knowledge transfer between task Adapters. I2I also results in better cross-task knowledge transfer than the state-of-the-art AdapterFusion without incurring the associated parametric cost.

Minglai Yang, Xinyan Velocity Yu, Pengyuan Li et al.

Document parsing and recognition are fundamental capabilities for vision-language models (VLMs) and document processing systems. However, existing Optical Character Recognition (OCR) and document parsing benchmarks are increasingly limited in coverage and difficulty: many focus on common document genres or uniformly sampled pages where modern parsers already perform strongly, while offering limited annotation for expert-domain structures such as chemical formula, music notation, complex tables, and cross-page layouts. We introduce Dr. DocBench, a difficulty-aware benchmark for expert-level document parsing. Built from a large-scale multilingual book corpus, Dr. DocBench spans 52 BISAC subject domains and selects challenging documents through parser-failure-based sampling, targeting cases where multiple state-of-the-art systems struggle. It contains 4,514 annotated pages from long documents averaging around 100 pages, with 65k high-quality page- and block-level annotations for layout, reading order, hierarchical relations, and domain-specific visual contents. Evaluations of pipeline-based parsers and general-purpose VLMs show that strong performance on existing benchmarks does not transfer to our expert-level document parsing. Our analysis reveals substantial failures across subjects, content types, and structural attributes, highlighting Dr. DocBench as a comprehensive testbed for diagnosing and advancing document intelligence.

Wang Zhu, Jesse Thomason, Robin Jia · uw

For vision-and-language reasoning tasks, both fully connectionist, end-to-end methods and hybrid, neuro-symbolic methods have achieved high in-distribution performance. In which out-of-distribution settings does each paradigm excel? We investigate this question on both single-image and multi-image visual question-answering through four types of generalization tests: a novel segment-combine test for multi-image queries, contrast set, compositional generalization, and cross-benchmark transfer. Vision-and-language end-to-end trained systems exhibit sizeable performance drops across all these tests. Neuro-symbolic methods suffer even more on cross-benchmark transfer from GQA to VQA, but they show smaller accuracy drops on the other generalization tests and their performance quickly improves by few-shot training. Overall, our results demonstrate the complementary benefits of these two paradigms, and emphasize the importance of using a diverse suite of generalization tests to fully characterize model robustness to distribution shift.

Lee Kezar, Jesse Thomason, Zed Sevcikova Sehyr · uw

We use insights from research on American Sign Language (ASL) phonology to train models for isolated sign language recognition (ISLR), a step towards automatic sign language understanding. Our key insight is to explicitly recognize the role of phonology in sign production to achieve more accurate ISLR than existing work which does not consider sign language phonology. We train ISLR models that take in pose estimations of a signer producing a single sign to predict not only the sign but additionally its phonological characteristics, such as the handshape. These auxiliary predictions lead to a nearly 9% absolute gain in sign recognition accuracy on the WLASL benchmark, with consistent improvements in ISLR regardless of the underlying prediction model architecture. This work has the potential to accelerate linguistic research in the domain of signed languages and reduce communication barriers between deaf and hearing people.

Lee Kezar, Riley Carlin, Tejas Srinivasan et al. · uw

Like speech, signs are composed of discrete, recombinable features called phonemes. Prior work shows that models which can recognize phonemes are better at sign recognition, motivating deeper exploration into strategies for modeling sign language phonemes. In this work, we learn graph convolution networks to recognize the sixteen phoneme "types" found in ASL-LEX 2.0. Specifically, we explore how learning strategies like multi-task and curriculum learning can leverage mutually useful information between phoneme types to facilitate better modeling of sign language phonemes. Results on the Sem-Lex Benchmark show that curriculum learning yields an average accuracy of 87% across all phoneme types, outperforming fine-tuning and multi-task strategies for most phoneme types.

Sara Mohammadinejad, Jesse Thomason, Jyotirmoy V. Deshmukh · uw

Natural language is an intuitive way for humans to communicate tasks to a robot. While natural language (NL) is ambiguous, real world tasks and their safety requirements need to be communicated unambiguously. Signal Temporal Logic (STL) is a formal logic that can serve as a versatile, expressive, and unambiguous formal language to describe robotic tasks. On one hand, existing work in using STL for the robotics domain typically requires end-users to express task specifications in STL, a challenge for non-expert users. On the other, translating from NL to STL specifications is currently restricted to specific fragments. In this work, we propose DIALOGUESTL, an interactive approach for learning correct and concise STL formulas from (often) ambiguous NL descriptions. We use a combination of semantic parsing, pre-trained transformer-based language models, and user-in-the-loop clarifications aided by a small number of user demonstrations to predict the best STL formula to encode NL task descriptions. An advantage of mapping NL to STL is that there has been considerable recent work on the use of reinforcement learning (RL) to identify control policies for robots. We show we can use Deep Q-Learning techniques to learn optimal policies from the learned STL specifications. We demonstrate that DIALOGUESTL is efficient, scalable, and robust, and has high accuracy in predicting the correct STL formula with a few number of demonstrations and a few interactions with an oracle user.

Chancharik Mitra, Abrar Anwar, Rodolfo Corona et al.

When connecting objects and their language referents in an embodied 3D environment, it is important to note that: (1) an object can be better characterized by leveraging comparative information between itself and other objects, and (2) an object's appearance can vary with camera position. As such, we present the Multi-view Approach to Grounding in Context (MAGiC), which selects an object referent based on language that distinguishes between two similar objects. By pragmatically reasoning over both objects and across multiple views of those objects, MAGiC improves over the state-of-the-art model on the SNARE object reference task with a relative error reduction of 12.9\% (representing an absolute improvement of 2.7\%). Ablation studies show that reasoning jointly over object referent candidates and multiple views of each object both contribute to improved accuracy. Code: https://github.com/rcorona/magic_snare/

Abhinav Gupta, Toben H. Mintz, Jesse Thomason

While word embeddings derive meaning from co-occurrence patterns, human language understanding is grounded in sensory and motor experience. We present $\text{SENSE}$ $(\textbf{S}\text{ensorimotor }$ $\textbf{E}\text{mbedding }$ $\textbf{N}\text{orm }$ $\textbf{S}\text{coring }$ $\textbf{E}\text{ngine})$, a learned projection model that predicts Lancaster sensorimotor norms from word lexical embeddings. We also conducted a behavioral study where 281 participants selected which among candidate nonce words evoked specific sensorimotor associations, finding statistically significant correlations between human selection rates and $\text{SENSE}$ ratings across 6 of the 11 modalities. Sublexical analysis of these nonce words selection rates revealed systematic phonosthemic patterns for the interoceptive norm, suggesting a path towards computationally proposing candidate phonosthemes from text data.

Wilbert Pumacay, Ishika Singh, Jiafei Duan et al. · uw

To realize effective large-scale, real-world robotic applications, we must evaluate how well our robot policies adapt to changes in environmental conditions. Unfortunately, a majority of studies evaluate robot performance in environments closely resembling or even identical to the training setup. We present THE COLOSSEUM, a novel simulation benchmark, with 20 diverse manipulation tasks, that enables systematical evaluation of models across 14 axes of environmental perturbations. These perturbations include changes in color, texture, and size of objects, table-tops, and backgrounds; we also vary lighting, distractors, physical properties perturbations and camera pose. Using THE COLOSSEUM, we compare 5 state-of-the-art manipulation models to reveal that their success rate degrades between 30-50% across these perturbation factors. When multiple perturbations are applied in unison, the success rate degrades $\geq$75%. We identify that changing the number of distractor objects, target object color, or lighting conditions are the perturbations that reduce model performance the most. To verify the ecological validity of our results, we show that our results in simulation are correlated ($\bar{R}^2 = 0.614$) to similar perturbations in real-world experiments. We open source code for others to use THE COLOSSEUM, and also release code to 3D print the objects used to replicate the real-world perturbations. Ultimately, we hope that THE COLOSSEUM will serve as a benchmark to identify modeling decisions that systematically improve generalization for manipulation. See https://robot-colosseum.github.io/ for more details.

Ting-Yun Chang, Jesse Thomason, Robin Jia

The concept of localization in LLMs is often mentioned in prior work; however, methods for localization have never been systematically and directly evaluated. We propose two complementary benchmarks that evaluate the ability of localization methods to pinpoint LLM components responsible for memorized data. In our INJ benchmark, we actively inject a piece of new information into a small subset of LLM weights, enabling us to directly evaluate whether localization methods can identify these "ground truth" weights. In our DEL benchmark, we evaluate localization by measuring how much dropping out identified neurons deletes a memorized pretrained sequence. Despite their different perspectives, our two benchmarks yield consistent rankings of five localization methods. Methods adapted from network pruning perform well on both benchmarks, and all evaluated methods show promising localization ability. On the other hand, even successful methods identify neurons that are not specific to a single memorized sequence.

Tejas Srinivasan, Jack Hessel, Tanmay Gupta et al. · allen-ai

Selective prediction minimizes incorrect predictions from vision-language models (VLMs) by allowing them to abstain from answering when uncertain. However, when deploying a vision-language system with low tolerance for inaccurate predictions, selective prediction may be over-cautious and abstain too frequently, even on many correct predictions. We introduce ReCoVERR, an inference-time algorithm to reduce the over-abstention of a selective vision-language system without increasing the error rate of the system's predictions. When the VLM makes a low-confidence prediction, instead of abstaining ReCoVERR tries to find relevant clues in the image that provide additional evidence for the prediction. ReCoVERR uses an LLM to pose related questions to the VLM, collects high-confidence evidences, and if enough evidence confirms the prediction the system makes a prediction instead of abstaining. ReCoVERR enables three VLMs (BLIP2, InstructBLIP, and LLaVA-1.5) to answer up to 20% more questions on the VQAv2 and A-OKVQA tasks without decreasing system accuracy, thus improving overall system reliability. Our code is available at https://github.com/tejas1995/ReCoVERR.

Wang Zhu, Ishika Singh, Yuan Huang et al.

Data augmentation via back-translation is common when pretraining Vision-and-Language Navigation (VLN) models, even though the generated instructions are noisy. But: does that noise matter? We find that nonsensical or irrelevant language instructions during pretraining can have little effect on downstream performance for both HAMT and VLN-BERT on R2R, and is still better than only using clean, human data. To underscore these results, we concoct an efficient augmentation method, Unigram + Object, which generates nonsensical instructions that nonetheless improve downstream performance. Our findings suggest that what matters for VLN R2R pretraining is the quantity of visual trajectories, not the quality of instructions.

Zain Merchant, Abrar Anwar, Emily Wang et al.

Navigating unfamiliar environments presents significant challenges for blind and low-vision (BLV) individuals. In this work, we construct a dataset of images and goals across different scenarios such as searching through kitchens or navigating outdoors. We then investigate how grounded instruction generation methods can provide contextually-relevant navigational guidance to users in these instances. Through a sighted user study, we demonstrate that large pretrained language models can produce correct and useful instructions perceived as beneficial for BLV users. We also conduct a survey and interview with 4 BLV users and observe useful insights on preferences for different instructions based on the scenario.

Wang Bill Zhu, Qiutong Tony Yi, Robin Jia et al.

Large language models (LLMs) perform substantially below human level on existing theory-of-mind (ToM) benchmarks, even when augmented with chain-of-thought prompting or probabilistic belief updates. We argue that these failures primarily arise from unreliable implicit state tracking rather than limitations in high-level reasoning. We introduce PDDL-Mind, a neuro-symbolic framework that decouples environment state evolution from belief inference. By translating narrative descriptions into explicit states and actions expressed in Planning Domain Definition Language (PDDL), and by verifying action-induced state transitions against a predefined domain, PDDL-Mind provides LLMs with a logically consistent and explicit representation of world states for ToM tasks. Experiments on MMToM-QA, MuMA and FanToM show that PDDL-Mind achieves over 5% absolute accuracy gain over the best existing state-of-the-art method on ToM benchmark questions.

Wei Yang, Shixuan Li, Heng Ping et al.

Multi-agent systems (MAS) can substantially extend the reasoning capacity of large language models (LLMs), yet most frameworks still aggregate agent outputs with majority voting. This heuristic discards the evidential structure of reasoning traces and is brittle under the confabulation consensus, where agents share correlated biases and converge on the same incorrect rationale. We introduce AgentAuditor, which replaces voting with a path search over a Reasoning Tree that explicitly represents agreements and divergences among agent traces. AgentAuditor resolves conflicts by comparing reasoning branches at critical divergence points, turning global adjudication into efficient, localized verification. We further propose Anti-Consensus Preference Optimization (ACPO), which trains the adjudicator on majority-failure cases and rewards evidence-based minority selections over popular errors. AgentAuditor is agnostic to MAS setting, and we find across 5 popular settings that it yields up to 5% absolute accuracy improvement over a majority vote, and up to 3% over using LLM-as-Judge.

Parsa Hejabi, Elnaz Rahmati, Alireza S. Ziabari et al.

Large Language Models (LLMs) have shown impressive capabilities in complex tasks and interactive environments, yet their creativity remains underexplored. This paper introduces a simulation framework utilizing the game Balderdash to evaluate both the creativity and logical reasoning of LLMs. In Balderdash, players generate fictitious definitions for obscure terms to deceive others while identifying correct definitions. Our framework enables multiple LLM agents to participate in this game, assessing their ability to produce plausible definitions and strategize based on game rules and history. We implemented a centralized game engine featuring various LLMs as participants and a judge LLM to evaluate semantic equivalence. Through a series of experiments, we analyzed the performance of different LLMs, examining metrics such as True Definition Ratio, Deception Ratio, and Correct Guess Ratio. The results provide insights into the creative and deceptive capabilities of LLMs, highlighting their strengths and areas for improvement. Specifically, the study reveals that infrequent vocabulary in LLMs' input leads to poor reasoning on game rules and historical context (https://github.com/ParsaHejabi/Simulation-Framework-for-Multi-Agent-Balderdash).

Heng Ping, Arijit Bhattacharjee, Peiyu Zhang et al.

Automation of Register Transfer Level (RTL) design can help developers meet increasing computational demands. Large Language Models (LLMs) show promise for Hardware Description Language (HDL) generation, but face challenges due to limited parametric knowledge and domain-specific constraints. While prompt engineering and fine-tuning have limitations in knowledge coverage and training costs, multi-agent architectures offer a training-free paradigm to enhance reasoning through collaborative generation. However, current multi-agent approaches suffer from two critical deficiencies: susceptibility to noise propagation and constrained reasoning space exploration. We propose VeriMoA, a training-free mixture-of-agents (MoA) framework with two synergistic innovations. First, a quality-guided caching mechanism to maintain all intermediate HDL outputs and enables quality-based ranking and selection across the entire generation process, encouraging knowledge accumulation over layers of reasoning. Second, a multi-path generation strategy that leverages C++ and Python as intermediate representations, decomposing specification-to-HDL translation into two-stage processes that exploit LLM fluency in high-resource languages while promoting solution diversity. Comprehensive experiments on VerilogEval 2.0 and RTLLM 2.0 benchmarks demonstrate that VeriMoA achieves 15--30% improvements in Pass@1 across diverse LLM backbones, especially enabling smaller models to match larger models and fine-tuned alternatives without requiring costly training.

Yiming Tang, Abrar Anwar, Jesse Thomason

Understanding social signals in multi-party conversations is important for human-robot interaction and artificial social intelligence. Social signals include body pose, head pose, speech, and context-specific activities like acquiring and taking bites of food when dining. Past work in multi-party interaction tends to build task-specific models for predicting social signals. In this work, we address the challenge of predicting multimodal social signals in multi-party settings in a single model. We introduce M3PT, a causal transformer architecture with modality and temporal blockwise attention masking to simultaneously process multiple social cues across multiple participants and their temporal interactions. We train and evaluate M3PT on the Human-Human Commensality Dataset (HHCD), and demonstrate that using multiple modalities improves bite timing and speaking status prediction. Source code: https://github.com/AbrarAnwar/masked-social-signals/.

Zizhao Hu, Mohammad Rostami, Jesse Thomason

Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered tasks require high-level human alignment. Prior works provide mixed opinions on their utility: some report performance gains when using expert personas for certain domains and their contribution to data diversity in synthetic data creation, while others find near-zero or negative impact on general utility. To fully leverage the benefits of the LLM persona and avoid its harmfulness, a more comprehensive investigation of the mechanism is crucial. In this work, we study how model optimization, task type, prompt length, and placement can impact expert persona effectiveness across instruction-tuned and reasoning LLMs, and provide insight into conditions under which expert personas fail and succeed. Based on our findings, we developed a pipeline to fully leverage the benefits of an expert persona, named PRISM (Persona Routing via Intent-based Self-Modeling), which self-distills an intent-conditioned expert persona into a gated LoRA adapter through a bootstrapping process that requires no external data, models, or knowledge. PRISM enhances human preference and safety alignment on generative tasks while maintaining accuracy on discriminative tasks across all models, with minimal memory and computing overhead.

Rohan Gupta, Trevor Asbery, Zain Merchant et al.

Coordinating heterogeneous robot fleets to achieve multiple goals is challenging in multi-robot systems. We introduce an open-source and extensible framework for centralized multi-robot task planning and scheduling that leverages LLMs to enable fleets of heterogeneous robots to accomplish multiple tasks. RobotFleet provides abstractions for planning, scheduling, and execution across robots deployed as containerized services to simplify fleet scaling and management. The framework maintains a shared declarative world state and two-way communication for task execution and replanning. By modularizing each layer of the autonomy stack and using LLMs for open-world reasoning, RobotFleet lowers the barrier to building scalable multi-robot systems. The code can be found here: https://github.com/therohangupta/robot-fleet.

Wei Yang, Jiacheng Pang, Shixuan Li et al.

Multi-agent systems (MAS) built on Large Language Models (LLMs) are being used to approach complex problems and can surpass single model inference. However, their success hinges on navigating a fundamental cognitive tension: the need to balance broad, divergent exploration of the solution space with a principled, convergent synthesis to the optimal solution. Existing paradigms often struggle to manage this duality, leading to premature consensus, error propagation, and a critical credit assignment problem that fails to distinguish between genuine reasoning and superficially plausible arguments. To resolve this core challenge, we propose the Multi-Agent Exploration-Synthesis framework Through Role Orchestration (Maestro), a principled paradigm for collaboration that structurally decouples these cognitive modes. Maestro uses a collective of parallel Execution Agents for diverse exploration and a specialized Central Agent for convergent, evaluative synthesis. To operationalize this critical synthesis phase, we introduce Conditional Listwise Policy Optimization (CLPO), a reinforcement learning objective that disentangles signals for strategic decisions and tactical rationales. By combining decision-focused policy gradients with a list-wise ranking loss over justifications, CLPO achieves clean credit assignment and stronger comparative supervision. Experiments on mathematical reasoning and general problem-solving benchmarks demonstrate that Maestro, coupled with CLPO, consistently outperforms existing state-of-the-art multi-agent approaches, delivering absolute accuracy gains of 6% on average and up to 10% at best.

Zizhao Hu, Ameya Godbole, Johnny Tian-Zheng Wei et al.

Machine unlearning for large language models (LLMs) aims to selectively remove memorized content such as private data, copyrighted text, or hazardous knowledge, without costly full retraining. Most existing methods require a retain set of curated examples to prevent catastrophic degradation of general model utility, creating an extra data dependency that complicates deployment. We propose SHRED (Self-distillation via High-surprisal-only Retain-set-free Entropy Demotion), a retain-set-free unlearning method built on a key insight: not all tokens within a forget set instance carry memorized information equally. High-information tokens concentrate the model's memorized knowledge, while low-information tokens reflect general language competence. SHRED operates in two stages. (1) Selection: We perform a forward pass on a forget set instance, collect per-token autoregressive probabilities, and select the bottom (lowest probability, highest Shannon information) as forget positions; the remaining positions are retained as benign anchors. (2) Training: We construct modified KL targets that demote the memorized token's logit at forget positions while preserving the original distribution at benign positions. The model is then trained via a single top KL self-distillation objective that simultaneously drives forgetting and utility preservation. We evaluate SHRED across four standard unlearning benchmarks and demonstrate that it establishes a new Pareto-optimal trade-off between forget efficacy and model utility, outperforming retain-set-dependent methods. Our analysis shows that SHRED is robust against relearning attacks and membership-inference attacks, and it maintains stable utility even after many sequential unlearning runs.

David Bai, Ishika Singh, David Traum et al.

Classical planning formulations like the Planning Domain Definition Language (PDDL) admit action sequences guaranteed to achieve a goal state given an initial state if any are possible. However, reasoning problems defined in PDDL do not capture temporal aspects of action taking, such as concurrent actions between two agents when there are no conflicting conditions, without significant modification and definition to existing PDDL domains. A human expert aware of such constraints can decompose a goal into subgoals, each reachable through single agent planning, to take advantage of simultaneous actions. In contrast to classical planning, large language models (LLMs) directly used for inferring plan steps rarely guarantee execution success, but are capable of leveraging commonsense reasoning to assemble action sequences. We combine the strengths of both classical planning and LLMs by approximating human intuitions for multi-agent planning goal decomposition. We demonstrate that LLM-based goal decomposition leads to faster planning times than solving multi-agent PDDL problems directly while simultaneously achieving fewer plan execution steps than a single agent plan alone, as well as most multiagent plans, while guaranteeing execution success. Additionally, we find that LLM-based approximations of subgoals result in similar multi-agent execution lengths to those specified by human experts. Website and resources at https://glamor-usc.github.io/twostep

Anthony Liang, Jesse Thomason, Erdem Bıyık · uw

Training robots to perform complex control tasks from high-dimensional pixel input using reinforcement learning (RL) is sample-inefficient, because image observations are comprised primarily of task-irrelevant information. By contrast, humans are able to visually attend to task-relevant objects and areas. Based on this insight, we introduce Visual Saliency-Guided Reinforcement Learning (ViSaRL). Using ViSaRL to learn visual representations significantly improves the success rate, sample efficiency, and generalization of an RL agent on diverse tasks including DeepMind Control benchmark, robot manipulation in simulation and on a real robot. We present approaches for incorporating saliency into both CNN and Transformer-based encoders. We show that visual representations learned using ViSaRL are robust to various sources of visual perturbations including perceptual noise and scene variations. ViSaRL nearly doubles success rate on the real-robot tasks compared to the baseline which does not use saliency.

Tejas Srinivasan, Jesse Thomason

Trust biases how users rely on AI recommendations in AI-assisted decision-making tasks, with low and high levels of trust resulting in increased under- and over-reliance, respectively. We propose that AI assistants should adapt their behavior through trust-adaptive interventions to mitigate such inappropriate reliance. For instance, when user trust is low, providing an explanation can elicit more careful consideration of the assistant's advice by the user. In two decision-making scenarios -- laypeople answering science questions and doctors making medical diagnoses -- we find that providing supporting and counter-explanations during moments of low and high trust, respectively, yields up to 38% reduction in inappropriate reliance and 20% improvement in decision accuracy. We are similarly able to reduce over-reliance by adaptively inserting forced pauses to promote deliberation. Our results highlight how AI adaptation to user trust facilitates appropriate reliance, presenting exciting avenues for improving human-AI collaboration.

Wei Yang, Jesse Thomason

Multi-agent systems of large language models (LLMs) show promise for complex reasoning, but their effectiveness is often limited by fixed collaboration protocols. These frameworks typically focus on macro-level orchestration while overlooking agents' internal deliberative capabilities. This critical meta-cognitive blindspot treats agents as passive executors unable to adapt their strategy based on internal cognitive states like uncertainty or confidence. We introduce the Meta-Policy Deliberation Framework (MPDF), where agents learn a decentralized policy over a set of high-level meta-cognitive actions: Persist, Refine, and Concede. To overcome the instability of traditional policy gradients in this setting, we develop SoftRankPO, a novel reinforcement learning algorithm. SoftRankPO stabilizes training by shaping advantages based on the rank of rewards mapped through smooth normal quantiles, making the learning process robust to reward variance. Experiments show that MPDF with SoftRankPO achieves a a 4-5% absolute gain in average accuracy across five mathematical and general reasoning benchmarks compared to six state-of-the-art heuristic and learning-based multi-agent reasoning algorithms. Our work presents a paradigm for learning adaptive, meta-cognitive policies for multi-agent LLM systems, shifting the focus from designing fixed protocols to learning dynamic, deliberative strategies.

Siddartha Devic, Tejas Srinivasan, Jesse Thomason et al.

Large Language Models (LLMs) are increasingly assisting users in the real world, yet their reliability remains a concern. Uncertainty quantification (UQ) has been heralded as a tool to enhance human-LLM collaboration by enabling users to know when to trust LLM predictions. We argue that current practices for uncertainty quantification in LLMs are not optimal for developing useful UQ for human users making decisions in real-world tasks. Through an analysis of 40 LLM UQ methods, we identify three prevalent practices hindering the community's progress toward its goal of benefiting downstream users: 1) evaluating on benchmarks with low ecological validity; 2) considering only epistemic uncertainty; and 3) optimizing metrics that are not necessarily indicative of downstream utility. For each issue, we propose concrete user-centric practices and research directions that LLM UQ researchers should consider. Instead of hill-climbing on unrepresentative tasks using imperfect metrics, we argue that the community should adopt a more human-centered approach to LLM uncertainty quantification.

Mert İnan, Anthony Sicilia, Suvodip Dey et al.

While theories of discourse and cognitive science have long recognized the value of unhurried pacing, recent dialogue research tends to minimize friction in conversational systems. Yet, frictionless dialogue risks fostering uncritical reliance on AI outputs, which can obscure implicit assumptions and lead to unintended consequences. To meet this challenge, we propose integrating positive friction into conversational AI, which promotes user reflection on goals, critical thinking on system response, and subsequent re-conditioning of AI systems. We hypothesize systems can improve goal alignment, modeling of user mental states, and task success by deliberately slowing down conversations in strategic moments to ask questions, reveal assumptions, or pause. We present an ontology of positive friction and collect expert human annotations on multi-domain and embodied goal-oriented corpora. Experiments on these corpora, along with simulated interactions using state-of-the-art systems, suggest incorporating friction not only fosters accountable decision-making, but also enhances machine understanding of user beliefs and goals, and increases task success rates.

Marcus Ma, Georgios Chochlakis, Niyantha Maruthu Pandiyan et al.

Multi-label classification is prevalent in real-world settings, but the behavior of Large Language Models (LLMs) in this setting is understudied. We investigate how autoregressive LLMs perform multi-label classification, focusing on subjective tasks, by analyzing the output distributions of the models at each label generation step. We find that the initial probability distribution for the first label often does not reflect the eventual final output, even in terms of relative order and find LLMs tend to suppress all but one label at each generation step. We further observe that as model scale increases, their token distributions exhibit lower entropy and higher single-label confidence, but the internal relative ranking of the labels improves. Finetuning methods such as supervised finetuning and reinforcement learning amplify this phenomenon. We introduce the task of distribution alignment for multi-label settings: aligning LLM-derived label distributions with empirical distributions estimated from annotator responses in subjective tasks. We propose both zero-shot and supervised methods which improve both alignment and predictive performance over existing approaches. We find one method -- taking the max probability over all label generation distributions instead of just using the initial probability distribution -- improves both distribution alignment and overall F1 classification without adding any additional computation.

Abrar Anwar, Rohan Gupta, Zain Merchant et al.

Evaluating learned robot control policies to determine their physical task-level capabilities costs experimenter time and effort. The growing number of policies and tasks exacerbates this issue. It is impractical to test every policy on every task multiple times; each trial requires a manual environment reset, and each task change involves re-arranging objects or even changing robots. Naively selecting a random subset of tasks and policies to evaluate is a high-cost solution with unreliable, incomplete results. In this work, we formulate robot evaluation as an active testing problem. We propose to model the distribution of robot performance across all tasks and policies as we sequentially execute experiments. Tasks often share similarities that can reveal potential relationships in policy behavior, and we show that natural language is a useful prior in modeling these relationships between tasks. We then leverage this formulation to reduce the experimenter effort by using a cost-aware expected information gain heuristic to efficiently select informative trials. Our framework accommodates both continuous and discrete performance outcomes. We conduct experiments on existing evaluation data from real robots and simulations. By prioritizing informative trials, our framework reduces the cost of calculating evaluation metrics for robot policies across many tasks.

Zizhao Hu, Mohammad Rostami, Jesse Thomason

Recent research has highlighted the risk of generative model collapse, where performance progressively degrades when continually trained on self-generated data. However, existing exploration on model collapse is limited to single, unimodal models, limiting our understanding in more realistic scenarios, such as diverse multi-modal AI agents interacting autonomously through synthetic data and continually evolving. We expand the synthetic data training and model collapse study to multi-modal vision-language generative systems, such as vision-language models (VLMs) and text-to-image diffusion models, as well as recursive generate-train loops with multiple models. We find that model collapse, previously observed in single-modality generative models, exhibits distinct characteristics in the multi-modal context, such as improved vision-language alignment and increased variance in VLM image-captioning task. Additionally, we find that general approaches such as increased decoding budgets, greater model diversity, and relabeling with frozen models can effectively mitigate model collapse. Our findings provide initial insights and practical guidelines for reducing the risk of model collapse in self-improving multi-agent AI systems and curating robust multi-modal synthetic datasets.

Ting-Yun Chang, Muru Zhang, Jesse Thomason et al.

Low-bit weight-only quantization significantly reduces the memory footprint of large language models (LLMs), but disproportionately affects certain examples. We analyze diverse 3-4 bit methods on LLMs ranging from 7B-70B in size and find that the quantization errors of 50 pairs of methods are strongly correlated (avg. 0.82) on FineWeb examples. Moreover, the residual stream magnitudes of full-precision models are indicative of future quantization errors. We further establish a hypothesis that relates the residual stream magnitudes to error amplification and accumulation over layers. Using LLM localization techniques, early exiting, and activation patching, we show that examples with large errors rely on precise residual activations in the late layers, and that the outputs of MLP gates play a crucial role in maintaining the perplexity. Our work reveals why certain examples result in large quantization errors and which model components are most critical for performance preservation.

Woojeong Jin, Tejas Srinivasan, Jesse Thomason et al.

Humans perceive and comprehend different visual properties of an object based on specific contexts. For instance, we know that a banana turns brown ``when it becomes rotten,'' whereas it appears green ``when it is unripe.'' Previous studies on probing visual commonsense knowledge have primarily focused on examining language models' understanding of typical properties (e.g., colors and shapes) of objects. We present WinoViz, a text-only evaluation dataset, consisting of 1,380 examples that probe the reasoning abilities of language models regarding variant visual properties of objects under different contexts or states. Our task is challenging since it requires pragmatic reasoning (finding intended meanings) and visual knowledge reasoning. We also present multi-hop data, a more challenging version of our data, which requires multi-step reasoning chains to solve our task. In our experimental analysis, our findings are: a) Large language models such as GPT-4 demonstrate effective performance, but when it comes to multi-hop data, their performance is significantly degraded. b) Large models perform well on pragmatic reasoning, but visual knowledge reasoning is a bottleneck in our task. c) Vision-language models outperform their language-model counterparts. d) A model with machine-generated images performs poorly in our task. This is due to the poor quality of the generated images.

Chancharik Mitra, Yusen Luo, Raj Saravanan et al.

Vision-Language Action (VLAs) models promise to extend the remarkable success of vision-language models (VLMs) to robotics. Yet, unlike VLMs in the vision-language domain, VLAs for robotics require finetuning to contend with varying physical factors like robot embodiment, environment characteristics, and spatial relationships of each task. Existing fine-tuning methods lack specificity, adapting the same set of parameters regardless of a task's visual, linguistic, and physical characteristics. Inspired by functional specificity in neuroscience, we hypothesize that it is more effective to finetune sparse model representations specific to a given task. In this work, we introduce Robotic Steering, a finetuning approach grounded in mechanistic interpretability that leverages few-shot demonstrations to identify and selectively finetune task-specific attention heads aligned with the physical, visual, and linguistic requirements of robotic tasks. Through comprehensive on-robot evaluations with a Franka Emika robot arm, we demonstrate that Robotic Steering outperforms LoRA while achieving superior robustness under task variation, reduced computational cost, and enhanced interpretability for adapting VLAs to diverse robotic tasks.