arXiv subjects

Julian McAuley

Publications and source records attributed to Julian McAuley.

At least 109 records · Page 6Linked to original sources

arXiv2025

Mem-α: Learning Memory Construction via Reinforcement Learning

Large language model (LLM) agents are constrained by limited context windows, necessitating external memory systems for long-term information understanding. Current memory-augmented agents typically depend on pre-defined instructions and tools for memory updates. However, language models may lack the ability to determine which information to store, how to structure it, and when to update it, especially as memory systems become more complex. This results in suboptimal memory construction and information loss. To this end, we propose Mem-alpha, a reinforcement learning framework that trains agents to effectively manage complex memory systems through interaction and feedback. We also construct a specialized training dataset spanning diverse multi-turn interaction patterns paired with comprehensive evaluation questions designed to teach effective memory management. During training, agents process sequential information chunks, learn to extract and store relevant content, then update the memory system. The reward signal derives from downstream question-answering accuracy over the full interaction history, directly optimizing for memory construction. To illustrate the effectiveness of our training framework, we design a memory architecture comprising core, episodic, and semantic components, equipped with multiple tools for memory operations. Empirical evaluation demonstrates that Mem-alpha achieves significant improvements over existing memory-augmented agent baselines. Despite being trained exclusively on instances with a maximum length of 30k tokens, our agents exhibit remarkable generalization to sequences exceeding 400k tokens, over 13x the training length, highlighting the robustness of Mem-alpha.

cs.CL↗

arXiv2025

Purely Semantic Indexing for LLM-based Generative Recommendation and Retrieval

Semantic identifiers (IDs) have proven effective in adapting large language models for generative recommendation and retrieval. However, existing methods often suffer from semantic ID conflicts, where semantically similar documents (or items) are assigned identical IDs. A common strategy to avoid conflicts is to append a non-semantic token to distinguish them, which introduces randomness and expands the search space, therefore hurting performance. In this paper, we propose purely semantic indexing to generate unique, semantic-preserving IDs without appending non-semantic tokens. We enable unique ID assignment by relaxing the strict nearest-centroid selection and introduce two model-agnostic algorithms: exhaustive candidate matching (ECM) and recursive residual searching (RRS). Extensive experiments on sequential recommendation, product search, and document retrieval tasks demonstrate that our methods improve both overall and cold-start performance, highlighting the effectiveness of ensuring ID uniqueness.

cs.IR↗

arXiv2025

Pluralistic Off-policy Evaluation and Alignment

Personalized preference alignment for LLMs with diverse human preferences requires evaluation and alignment methods that capture pluralism. Most existing preference alignment datasets are logged under policies that differ substantially from the evaluated LLMs, and existing off-policy estimators focus solely on overall utility while ignoring preference pluralism. Extending Off-Policy Evaluation (OPE) to pluralistic preference alignment, therefore, remains an open question. Thus, we propose the Pluralistic Off-Policy Evaluation (POPE), the first framework for offline pluralistic preference evaluation and alignment in LLMs. POPE includes a unified reward function that combines (1) a collaborative utility component derived from human preference signals (e.g., upvotes or relevance scores) and (2) a diversity component inspired by entropy-based coverage measures, together reflecting pluralistic alignment. Furthermore, to estimate this reward from logged interactions, we derive decomposable inverse propensity scoring (IPS) estimators that separately evaluate relevance and diversity. Theoretically, we prove that our decomposed IPS estimators establish a lower bound on their variance. With the off-policy evaluated value function, we can directly enable off-policy optimization to further enhance pluralistic alignment. Empirical results demonstrate that POPE efficiently enhances pluralistic response generation and maintains the models' general capabilities on downstream tasks

cs.CL↗

arXiv2025

CoMMIT: Coordinated Multimodal Instruction Tuning

Instruction tuning in multimodal large language models (MLLMs) generally involves cooperative learning between a backbone LLM and a feature encoder of non-text input modalities. The major challenge is how to efficiently find the synergy between the two modules so that LLMs can adapt their reasoning abilities to downstream tasks while feature encoders can adjust to provide more task-specific information about its modality. In this paper, we analyze the MLLM instruction tuning from both theoretical and empirical perspectives, where we find the unbalanced learning between the feature encoder and the LLM can cause problems of oscillation and biased learning that lead to sub-optimal convergence. Inspired by our findings, we propose a Multimodal Balance Coefficient that enables quantitative measurement of the balance of learning. Based on this, we further design a dynamic learning scheduler that better coordinates the learning between the LLM and feature encoder, alleviating the problems of oscillation and biased learning. In addition, we introduce an auxiliary regularization on the gradient to promote updating with larger step sizes, which potentially allows for a more accurate estimation of the proposed MultiModal Balance Coefficient and further improves the training sufficiency. Our proposed approach is agnostic to the architecture of LLM and feature encoder, so it can be generically integrated with various MLLMs. We conduct experiments on multiple downstream tasks with various MLLMs, demonstrating that the proposed method is more effective than the baselines in MLLM instruction tuning.

cs.LG↗

arXiv2025

In-context Ranking Preference Optimization

Recent developments in Direct Preference Optimization (DPO) allow large language models (LLMs) to function as implicit ranking models by maximizing the margin between preferred and non-preferred responses. In practice, user feedback on such lists typically involves identifying a few relevant items in context rather than providing detailed pairwise comparisons for every possible item pair. Moreover, many complex information retrieval tasks, such as conversational agents and summarization systems, critically depend on ranking the highest-quality outputs at the top, emphasizing the need to support natural and flexible forms of user feedback. To address the challenge of limited and sparse pairwise feedback in the in-context setting, we propose an In-context Ranking Preference Optimization (IRPO) framework that directly optimizes LLMs based on ranking lists constructed during inference. To further capture flexible forms of feedback, IRPO extends the DPO objective by incorporating both the relevance of items and their positions in the list. Modeling these aspects jointly is non-trivial, as ranking metrics are inherently discrete and non-differentiable, making direct optimization difficult. To overcome this, IRPO introduces a differentiable objective based on positional aggregation of pairwise item preferences, enabling effective gradient-based optimization of discrete ranking metrics. We further provide theoretical insights showing that IRPO (i) automatically emphasizes items with greater disagreement between the model and the reference ranking, and (ii) links its gradient to an importance sampling estimator, yielding an unbiased estimator with reduced variance. Empirical results show IRPO outperforms standard DPO approaches in ranking performance, highlighting its effectiveness in aligning LLMs with direct in-context ranking preferences.

cs.LG↗

arXiv2025

CoreThink: A Symbolic Reasoning Layer to reason over Long Horizon Tasks with LLMs

We introduce CoreThink, a state-of-the-art Reasoning Layer built upon a novel reasoning method called General Symbolics. This approach diverges from reasoning paradigms such as test-time scaling, Supervised Fine-Tuning (SFT), and Reinforcement Learning with Verifiable Rewards (RLVR). CoreThink General Symbolic Reasoner (GSR) is specifically structured around three key use cases: tool-calling, code generation, and planning, demonstrating exemplary performance across a total of seven benchmarks in their respective areas. Notably, we are achieving SOTA scores of 66.66% on Livecodebench v6, 89% on Instruction-Following Evals, and 24.4% on ARC-AGI-2. We also present an agentic coding IDE, developed using the principles of General Symbolics, which achieves a state-of-the-art accuracy of 62.3% on SWE-Bench Lite. We are able to achieve these improvements without any fine-tuning or training costs. Our Reasoning Layer is designed to provide a pure performance uplift, ensuring that a model's accuracy on reasoning tasks is never negatively impacted. We argue that incumbent methods will eventually lead to diminishing returns in LLM performance, necessitating the development of new reasoning techniques. This technical report details our approach at a high level and the availability of the CoreThink models for reasoning-intensive use cases.

cs.AI↗

arXiv2025

SAND: Boosting LLM Agents with Self-Taught Action Deliberation

Large Language Model (LLM) agents are commonly tuned with supervised finetuning on ReAct-style expert trajectories or preference optimization over pairwise rollouts. Most of these methods focus on imitating specific expert behaviors or promoting chosen reasoning thoughts and actions over rejected ones. However, without reasoning and comparing over alternatives actions, LLM agents finetuned with these methods may over-commit towards seemingly plausible but suboptimal actions due to limited action space exploration. To address this, in this paper we propose Self-taught ActioN Deliberation (SAND) framework, enabling LLM agents to explicitly deliberate over candidate actions before committing to one. To tackle the challenges of when and what to deliberate given large action space and step-level action evaluation, we incorporate self-consistency action sampling and execution-guided action critique to help synthesize step-wise action deliberation thoughts using the base model of the LLM agent. In an iterative manner, the deliberation trajectories are then used to finetune the LLM agent itself. Evaluating on two representative interactive agent tasks, SAND achieves an average 20% improvement over initial supervised finetuning and also outperforms state-of-the-art agent tuning approaches.

cs.CL↗

arXiv2025

ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation

Generative recommendation (GR) is an emerging paradigm where user actions are tokenized into discrete token patterns and autoregressively generated as predictions. However, existing GR models tokenize each action independently, assigning the same fixed tokens to identical actions across all sequences without considering contextual relationships. This lack of context-awareness can lead to suboptimal performance, as the same action may hold different meanings depending on its surrounding context. To address this issue, we propose ActionPiece to explicitly incorporate context when tokenizing action sequences. In ActionPiece, each action is represented as a set of item features. Given the action sequence corpora, we construct the vocabulary by merging feature patterns as new tokens, based on their co-occurrence frequency both within individual sets and across adjacent sets. Considering the unordered nature of feature sets, we further introduce set permutation regularization, which produces multiple segmentations of action sequences with the same semantics. Our code is available at: https://github.com/google-deepmind/action_piece.

cs.IR↗

arXiv2025

DICE: Dynamic In-Context Example Selection in LLM Agents via Efficient Knowledge Transfer

Large language model-based agents, empowered by in-context learning (ICL), have demonstrated strong capabilities in complex reasoning and tool-use tasks. However, existing works have shown that the effectiveness of ICL is highly sensitive to the choice of demonstrations, with suboptimal examples often leading to unstable or degraded performance. While prior work has explored example selection, including in some agentic or multi-step settings, existing approaches typically rely on heuristics or task-specific designs and lack a general, theoretically grounded criterion for what constitutes an effective demonstration across reasoning steps. Therefore, it is non-trivial to develop a principled, general-purpose method for selecting demonstrations that consistently benefit agent performance. In this paper, we address this challenge with DICE, Dynamic In-Context Example Selection for LLM Agents, a theoretically grounded ICL framework for agentic tasks that selects the most relevant demonstrations at each step of reasoning. Our approach decomposes demonstration knowledge into transferable and non-transferable components through a causal lens, showing how the latter can introduce spurious dependencies that impair generalization. We further propose a stepwise selection criterion with a formal guarantee of improved agent performance. Importantly, DICE is a general, framework-agnostic solution that can be integrated as a plug-in module into existing agentic frameworks without any additional training cost. Extensive experiments across diverse domains demonstrate our method's effectiveness and generality, highlighting the importance of principled, context-aware demo selection for robust and efficient LLM agents.

cs.AI↗

arXiv2025

PDB-Eval: An Evaluation of Large Multimodal Models for Description and Explanation of Personalized Driving Behavior

Understanding a driver's behavior and intentions is important for potential risk assessment and early accident prevention. Safety and driver assistance systems can be tailored to individual drivers' behavior, significantly enhancing their effectiveness. However, existing datasets are limited in describing and explaining general vehicle movements based on external visual evidence. This paper introduces a benchmark, PDB-Eval, for a detailed understanding of Personalized Driver Behavior, and aligning Large Multimodal Models (MLLMs) with driving comprehension and reasoning. Our benchmark consists of two main components, PDB-X and PDB-QA. PDB-X can evaluate MLLMs' understanding of temporal driving scenes. Our dataset is designed to find valid visual evidence from the external view to explain the driver's behavior from the internal view. To align MLLMs' reasoning abilities with driving tasks, we propose PDB-QA as a visual explanation question-answering task for MLLM instruction fine-tuning. As a generic learning task for generative models like MLLMs, PDB-QA can bridge the domain gap without harming MLLMs' generalizability. Our evaluation indicates that fine-tuning MLLMs on fine-grained descriptions and explanations can effectively bridge the gap between MLLMs and the driving domain, which improves zero-shot performance on question-answering tasks by up to 73.2%. We further evaluate the MLLMs fine-tuned on PDB-X in Brain4Cars' intention prediction and AIDE's recognition tasks. We observe up to 12.5% performance improvements on the turn intention prediction task in Brain4Cars, and consistent performance improvements up to 11.0% on all tasks in AIDE.

cs.CV↗

arXiv2025

WildFX: A DAW-Powered Pipeline for In-the-Wild Audio FX Graph Modeling

Despite rapid progress in end-to-end AI music generation, AI-driven modeling of professional Digital Signal Processing (DSP) workflows remains challenging. In particular, while there is growing interest in neural black-box modeling of audio effect graphs (e.g. reverb, compression, equalization), AI-based approaches struggle to replicate the nuanced signal flow and parameter interactions used in professional workflows. Existing differentiable plugin approaches often diverge from real-world tools, exhibiting inferior performance relative to simplified neural controllers under equivalent computational constraints. We introduce WildFX, a pipeline containerized with Docker for generating multi-track audio mixing datasets with rich effect graphs, powered by a professional Digital Audio Workstation (DAW) backend. WildFX supports seamless integration of cross-platform commercial plugins or any plugins in the wild, in VST/VST3/LV2/CLAP formats, enabling structural complexity (e.g., sidechains, crossovers) and achieving efficient parallelized processing. A minimalist metadata interface simplifies project/plugin configuration. Experiments demonstrate the pipeline's validity through blind estimation of mixing graphs, plugin/gain parameters, and its ability to bridge AI research with practical DSP demands. The code is available on: https://github.com/IsaacYQH/WildFX.

cs.SD↗

arXiv2025

Toward Holistic Evaluation of Recommender Systems Powered by Generative Models

Recommender systems powered by generative models (Gen-RecSys) extend beyond classical item ranking by producing open-ended content, which simultaneously unlocks richer user experiences and introduces new risks. On one hand, these systems can enhance personalization and appeal through dynamic explanations and multi-turn dialogues. On the other hand, they might venture into unknown territory-hallucinating nonexistent items, amplifying bias, or leaking private information. Traditional accuracy metrics cannot fully capture these challenges, as they fail to measure factual correctness, content safety, or alignment with user intent. This paper makes two main contributions. First, we categorize the evaluation challenges of Gen-RecSys into two groups: (i) existing concerns that are exacerbated by generative outputs (e.g., bias, privacy) and (ii) entirely new risks (e.g., item hallucinations, contradictory explanations). Second, we propose a holistic evaluation approach that includes scenario-based assessments and multi-metric checks-incorporating relevance, factual grounding, bias detection, and policy compliance. Our goal is to provide a guiding framework so researchers and practitioners can thoroughly assess Gen-RecSys, ensuring effective personalization and responsible deployment.

cs.IR↗

arXiv2025

Video-Guided Text-to-Music Generation Using Public Domain Movie Collections

Despite recent advancements in music generation systems, their application in film production remains limited, as they struggle to capture the nuances of real-world filmmaking, where filmmakers consider multiple factors-such as visual content, dialogue, and emotional tone-when selecting or composing music for a scene. This limitation primarily stems from the absence of comprehensive datasets that integrate these elements. To address this gap, we introduce Open Screen Soundtrack Library (OSSL), a dataset consisting of movie clips from public domain films, totaling approximately 36.5 hours, paired with high-quality soundtracks and human-annotated mood information. To demonstrate the effectiveness of our dataset in improving the performance of pre-trained models on film music generation tasks, we introduce a new video adapter that enhances an autoregressive transformer-based text-to-music model by adding video-based conditioning. Our experimental results demonstrate that our proposed approach effectively enhances MusicGen-Medium in terms of both objective measures of distributional and paired fidelity, and subjective compatibility in mood and genre. To facilitate reproducibility and foster future work, we publicly release the dataset, code, and demo.

cs.SD↗

arXiv2025

Weakly-supervised VLM-guided Partial Contrastive Learning for Visual Language Navigation

Visual Language Navigation (VLN) is a fundamental task within the field of Embodied AI, focusing on the ability of agents to navigate complex environments based on natural language instructions. Despite the progress made by existing methods, these methods often present some common challenges. First, they rely on pre-trained backbone models for visual perception, which struggle with the dynamic viewpoints in VLN scenarios. Second, the performance is limited when using pre-trained LLMs or VLMs without fine-tuning, due to the absence of VLN domain knowledge. Third, while fine-tuning LLMs and VLMs can improve results, their computational costs are higher than those without fine-tuning. To address these limitations, we propose Weakly-supervised Partial Contrastive Learning (WPCL), a method that enhances an agent's ability to identify objects from dynamic viewpoints in VLN scenarios by effectively integrating pre-trained VLM knowledge into the perception process, without requiring VLM fine-tuning. Our method enhances the agent's ability to interpret and respond to environmental cues while ensuring computational efficiency. Experimental results have shown that our method outperforms the baseline methods on multiple benchmarks, which validate the effectiveness, robustness and generalizability of our method.

cs.CV↗

arXiv2025

Generating Symbolic Music from Natural Language Prompts using an LLM-Enhanced Dataset

Recent years have seen many audio-domain text-to-music generation models that rely on large amounts of text-audio pairs for training. However, symbolic-domain controllable music generation has lagged behind partly due to the lack of a large-scale symbolic music dataset with extensive metadata and captions. In this work, we present MetaScore, a new dataset consisting of 963K musical scores paired with rich metadata, including free-form user-annotated tags, collected from an online music forum. To approach text-to-music generation, We employ a pretrained large language model (LLM) to generate pseudo-natural language captions for music from its metadata tags. With the LLM-enhanced MetaScore, we train a text-conditioned music generation model that learns to generate symbolic music from the pseudo captions, allowing control of instruments, genre, composer, complexity and other free-form music descriptors. In addition, we train a tag-conditioned system that supports a predefined set of tags available in MetaScore. Our experimental results show that both the proposed text-to-music and tags-to-music models outperform a baseline text-to-music model in a listening test. While a concurrent work Text2MIDI also supports free-form text input, our models achieve comparable performance. Moreover, the text-to-music system offers a more natural interface than the tags-to-music model, as it allows users to provide free-form natural language prompts.

cs.SD↗

arXiv2025

Lost in Sequence: Do Large Language Models Understand Sequential Recommendation?

Large Language Models (LLMs) have recently emerged as promising tools for recommendation thanks to their advanced textual understanding ability and context-awareness. Despite the current practice of training and evaluating LLM-based recommendation (LLM4Rec) models under a sequential recommendation scenario, we found that whether these models understand the sequential information inherent in users' item interaction sequences has been largely overlooked. In this paper, we first demonstrate through a series of experiments that existing LLM4Rec models do not fully capture sequential information both during training and inference. Then, we propose a simple yet effective LLM-based sequential recommender, called LLM-SRec, a method that enhances the integration of sequential information into LLMs by distilling the user representations extracted from a pre-trained CF-SRec model into LLMs. Our extensive experiments show that LLM-SRec enhances LLMs' ability to understand users' item interaction sequences, ultimately leading to improved recommendation performance. Furthermore, unlike existing LLM4Rec models that require fine-tuning of LLMs, LLM-SRec achieves state-of-the-art performance by training only a few lightweight MLPs, highlighting its practicality in real-world applications. Our code is available at https://github.com/Sein-Kim/LLM-SRec.

cs.IR↗

arXiv2025

Are you really listening? Boosting Perceptual Awareness in Music-QA Benchmarks

Large Audio Language Models (LALMs), where pretrained text LLMs are finetuned with audio input, have made remarkable progress in music understanding. However, current evaluation methodologies exhibit critical limitations: on the leading Music Question Answering benchmark, MuchoMusic, text-only LLMs without audio perception capabilities achieve surprisingly high accuracy of up to 56.4%, on par or above most LALMs. Furthermore, when presented with random Gaussian noise instead of actual audio, LALMs still perform significantly above chance. These findings suggest existing benchmarks predominantly assess reasoning abilities rather than audio perception. To overcome this challenge, we present RUListening: Robust Understanding through Listening, a framework that enhances perceptual evaluation in Music-QA benchmarks. We introduce the Perceptual Index (PI), a quantitative metric that measures a question's reliance on audio perception by analyzing log probability distributions from text-only language models. Using this metric, we generate synthetic, challenging distractors to create QA pairs that necessitate genuine audio perception. When applied to MuchoMusic, our filtered dataset successfully forces models to rely on perceptual information-text-only LLMs perform at chance levels, while LALMs similarly deteriorate when audio inputs are replaced with noise. These results validate our framework's effectiveness in creating benchmarks that more accurately evaluate audio perception capabilities.

cs.SD↗

arXiv2025

Generating Long Semantic IDs in Parallel for Recommendation

Semantic ID-based recommendation models tokenize each item into a small number of discrete tokens that preserve specific semantics, leading to better performance, scalability, and memory efficiency. While recent models adopt a generative approach, they often suffer from inefficient inference due to the reliance on resource-intensive beam search and multiple forward passes through the neural sequence model. As a result, the length of semantic IDs is typically restricted (e.g. to just 4 tokens), limiting their expressiveness. To address these challenges, we propose RPG, a lightweight framework for semantic ID-based recommendation. The key idea is to produce unordered, long semantic IDs, allowing the model to predict all tokens in parallel. We train the model to predict each token independently using a multi-token prediction loss, directly integrating semantics into the learning objective. During inference, we construct a graph connecting similar semantic IDs and guide decoding to avoid generating invalid IDs. Experiments show that scaling up semantic ID length to 64 enables RPG to outperform generative baselines by an average of 12.6% on the NDCG@10, while also improving inference efficiency. Code is available at: https://github.com/facebookresearch/RPG_KDD2025.

cs.IR↗