Post-training for large language models (LLMs) is constrained by the high cost of acquiring new knowledge or correcting errors and by the unintended side effects that frequently arise from retraining. To address these issues, we introduce REPAIR (Robust Editing via Progressive Adaptive Intervention and Reintegration), a lifelong editing framework designed to support precise and low-cost model updates while preserving non-target knowledge. REPAIR mitigates the instability and conflicts of large-scale sequential edits through a closed-loop feedback mechanism coupled with dynamic memory management. Furthermore, by incorporating frequent knowledge fusion and enforcing strong locality guards, REPAIR effectively addresses the shortcomings of traditional distribution-agnostic approaches that often overlook unintended ripple effects. Our experiments demonstrate that REPAIR boosts editing accuracy by 10%-30% across multiple model families and significantly reduces knowledge forgetting. This work introduces a robust framework for developing reliable, scalable, and continually evolving LLMs.

To solve a new task from minimal experience, it is essential to effectively reuse knowledge from previous tasks, a problem known as meta-learning. Compositional solutions, where common elements of computation are flexibly recombined into new configurations, are particularly well-suited for meta-learning. Here, we propose a compositional meta-learning model that explicitly represents tasks as structured combinations of reusable computations. We achieve this by learning a generative model that captures the underlying components and their statistics shared across a family of tasks. This approach transforms learning a new task into a probabilistic inference problem, which allows for finding solutions without parameter updates through highly constrained hypothesis testing. Our model successfully recovers ground truth components and statistics in rule learning and motor learning tasks. We then demonstrate its ability to quickly infer new solutions from just single examples. Together, our framework joins the expressivity of neural networks with the data-efficiency of probabilistic inference to achieve rapid compositional meta-learning.

We reframe and operationalise adversarial inverse reinforcement learning (IRL) to large language model reasoning, learning a dense, token-level reward model for process supervision directly from expert demonstrations rather than imitating style via supervised fine-tuning. The learned reasoning reward serves two complementary roles: (i) it provides step-level feedback to optimise a reasoning policy during training; and (ii) it functions at inference as a critic to rerank sampled traces under fixed compute budgets. We demonstrate that our approach prioritises correctness over surface form, yielding scores that correlate with eventual answer validity and enabling interpretable localisation of errors within a trace. Empirically, on GSM8K with Llama3 and Qwen2.5 backbones, we demonstrate: (i) dense reasoning rewards can be used as a learning signal to elicit reasoning, and (ii) predictive performance is improved from reward-guided reranking (notably for Llama-based policies). By unifying training signals, inference-time selection, and token-level diagnostics into a single reasoning reward, this work suggests reusable process-level rewards with broad potential to enhance multi-step reasoning in language models.

Object-Goal Navigation (ObjectNav) is a critical component toward deploying mobile robots in everyday, uncontrolled environments such as homes, schools, and workplaces. In this context, a robot must locate target objects in previously unseen environments using only its onboard perception. Success requires the integration of semantic understanding, spatial reasoning, and long-horizon planning, which is a combination that remains extremely challenging. While reinforcement learning (RL) has become the dominant paradigm, progress has spanned a wide range of design choices, yet the field still lacks a unifying analysis to determine which components truly drive performance. In this work, we conduct a large-scale empirical study of modular RL-based ObjectNav systems, decomposing them into three key components: perception, policy, and test-time enhancement. Through extensive controlled experiments, we isolate the contribution of each and uncover clear trends: perception quality and test-time strategies are decisive drivers of performance, whereas policy improvements with current methods yield only marginal gains. Building on these insights, we propose practical design guidelines and demonstrate an enhanced modular system that surpasses State-of-the-Art (SotA) methods by 6.6% on SPL and by a 2.7% success rate. We also introduce a human baseline under identical conditions, where experts achieve an average 98% success, underscoring the gap between RL agents and human-level navigation. Our study not only sets the SotA performance but also provides principled guidance for future ObjectNav development and evaluation. Recent advances in computer vision and deep learning have inspired growing interest in interdisciplinary applications that bridge perception, reasoning, and control, especially in robotics. Among these, vision-based navigation has emerged as a foundational capability for autonomous mobile agents. A key benchmark in this domain is Object-Goal Navigation (ObjectNav), where a robot must navigate to an instance of a specified object category in an unseen environment, relying solely on its onboard sensors. This task is both practically important and technically challenging: it requires semantic understanding, spatial reasoning, and long-horizon planning. Among many approaches, Reinforcement Learning (RL) has become a dominant paradigm for ObjectNav, offering a structured framework to learn directly through trial-and-error and showing steady progress across various benchmarks. While end-to-end RL policies are common, modular RL approaches have shown greater robustness and improved generalization.

Evaluation of potential AGI systems and methods is difficult due to the breadth of the engineering goal. We have no methods for perfect evaluation of the end state, and instead measure performance on small tests designed to provide directional indication that we are approaching AGI. In this work we argue that AGI evaluation methods have been dominated by a design philosophy that uses our intuitions of what intelligence is to create synthetic tasks, that have performed poorly in the history of AI. Instead we argue for an alternative design philosophy focused on evaluating robust task execution that seeks to demonstrate AGI through competence. This perspective is developed from common practices in data science that are used to show that a system can be reliably deployed. We provide practical examples of what this would mean for AGI evaluation.

Multi-step manipulation in dynamic environments remains challenging. Two major families of methods fail in distinct ways: (i) imitation learning (IL) is reactive but lacks compositional generalization, as monolithic policies do not decide which skill to reuse when scenes change; (ii) classical task-and-motion planning (TAMP) offers compositionality but has prohibitive planning latency, preventing real-time failure recovery. We introduce SymSkill, a unified learning framework that combines the benefits of IL and TAMP, allowing compositional generalization and failure recovery in real-time. Offline, SymSkill jointly learns predicates, operators, and skills directly from unlabeled and unsegmented demonstrations. At execution time, upon specifying a conjunction of one or more learned predicates, SymSkill uses a symbolic planner to compose and reorder learned skills to achieve the symbolic goals, while performing recovery at both the motion and symbolic levels in real time. Coupled with a compliant controller, SymSkill enables safe and uninterrupted execution under human and environmental disturbances. In RoboCasa simulation, SymSkill can execute 12 single-step tasks with 85% success rate. Without additional data, it composes these skills into multi-step plans requiring up to 6 skill recompositions, recovering robustly from execution failures. On a real Franka robot, we demonstrate SymSkill, learning from 5 minutes of unsegmented and unlabeled play data, is capable of performing multiple tasks simply by goal specifications. The source code and additional analysis can be found on https://sites.google.com/view/symskill.

Abstract--Although existing cross-domain continual learning approaches successfully address many streaming tasks having domain shifts, they call for a fully labeled source domain hindering their feasibility in the privacy constrained environments. This paper goes one step ahead with the problem of source-free cross-domain continual learning where the use of source-domain samples are completely prohibited. We propose the idea of rehearsal-free frequency-aware dynamic prompt collaborations (REFEREE) to cope with the absence of labeled source-domain samples in realm of cross-domain continual learning. REFEREE is built upon a synergy between a source-pre-trained model and a large-scale vision-language model, thus overcoming the problem of sub-optimal generalizations when relying only on a source pre-trained model. The domain shift problem between the source domain and the target domain is handled by a frequency-aware prompting technique encouraging low-frequency components while suppressing high-frequency components. This strategy generates frequency-aware augmented samples, robust against noisy pseudo labels. The noisy pseudo-label problem is further addressed with the uncertainty-aware weighting strategy where the mean and covariance matrix are weighted by prediction uncertainties, thus mitigating the adverse effects of the noisy pseudo label. Besides, the issue of catastrophic forgetting (CF) is overcome by kernel linear discriminant analysis (KLDA) where the backbone network is frozen while the classification is performed using the linear discriminant analysis approach guided by the random kernel method. Our rigorous numerical studies confirm the advantage of our approach where it beats prior arts having access to source domain samples with significant margins. HE goal of continual learning (CL) is to deal with lifelong learning environments where a sequence of non-stationary tasks is observed.

Training data plays a crucial role in Large Language Models (LLM) scaling, yet high quality data is of limited supply. Synthetic data techniques offer a potential path toward sidestepping these limitations. We conduct a large-scale empirical investigation (>1000 LLMs with >100k GPU hours) using a unified protocol and scaling laws, comparing natural web data, diverse synthetic types (rephrased text, generated textbooks), and mixtures of natural and synthetic data. Specifically, we found pre-training on rephrased synthetic data \textit{alone} is not faster than pre-training on natural web texts; while pre-training on 1/3 rephrased synthetic data mixed with 2/3 natural web texts can speed up 5-10x (to reach the same validation loss) at larger data budgets. Pre-training on textbook-style synthetic data \textit{alone} results in notably higher loss on many downstream domains especially at small data budgets. "Good" ratios of synthetic data in training data mixtures depend on the model size and data budget, empirically converging to ~30% for rephrased synthetic data. Larger generator models do not necessarily yield better pre-training data than ~8B-param models. These results contribute mixed evidence on "model collapse" during large-scale single-round (n=1) model training on synthetic data--training on rephrased synthetic data shows no degradation in performance in foreseeable scales whereas training on mixtures of textbook-style pure-generated synthetic data shows patterns predicted by "model collapse". Our work demystifies synthetic data in pre-training, validates its conditional benefits, and offers practical guidance.

Recent research has explored using Large Language Models for recommendation tasks by transforming user interaction histories and item metadata into text prompts, then having the LLM produce rankings or recommendations. A promising approach involves connecting collaborative filtering knowledge to LLM representations through compact adapter networks, which avoids expensive fine-tuning while preserving the strengths of both components. Yet several challenges persist in practice: collaborative filtering models often use static snapshots that miss rapidly changing user preferences; many real-world items contain rich visual and audio content beyond textual descriptions; and current systems struggle to provide trustworthy explanations backed by concrete evidence. Our work introduces \model{}, a framework that tackles these limitations through three key innovations. We develop an online adaptation mechanism that continuously incorporates new user interactions through lightweight modules, avoiding the need to retrain large models. We create a unified representation that seamlessly combines collaborative signals with visual and audio features, handling cases where some modalities may be unavailable. Finally, we design an explanation system that grounds recommendations in specific collaborative patterns and item attributes, producing natural language rationales users can verify. Our approach maintains the efficiency of frozen base models while adding minimal computational overhead, making it practical for real-world deployment.

W e develop ImageNet-Think, a multimodal reasoning dataset designed to aid the development of Vision Language Models (VLMs) with explicit reasoning capabilities. Our dataset is built on 250,000 images from ImageNet-21k dataset, providing structured thinking tokens and corresponding answers. Our synthetic dataset is generated by two state-of-the-art VLMs: GLM-4.1V-9B-Thinking and Kimi-VL-A3B-Thinking-2506. Each image is accompanied by two pairs of thinking-answer sequences, creating a resource for training and evaluating multimodal reasoning models. W e capture the step-by-step reasoning process of VLMs and the final descriptive answers. Our goal with this dataset is to enable the development of more robust VLMs while contributing to the broader understanding of multi-modal reasoning mechanisms. The dataset and evaluation benchmarks will be publicly available to aid research in reasoning/thinking multimodal VLMs. The dataset is available here on HuggingFace.