Connections between reinforcement learning with feedback,test-time scaling, and diffusion guidance: An anthology

In this note, we reflect on several fundamental connections among widely used post-training techniques. We clarify some intimate connections and equivalences between reinforcement learning with human feedback, reinforcement learning with internal feedback, and test-time scaling (particularly soft best-of-$N$ sampling), while also illuminating intrinsic links between diffusion guidance and test-time scaling. Additionally, we introduce a resampling approach for alignment and reward-directed diffusion models, sidestepping the need for explicit reinforcement learning techniques.

Cognitive Structure Generation: From Educational Priors to Policy Optimization

Cognitive structure is a student's subjective organization of an objective knowledge system, reflected in the psychological construction of concepts and their relations. However, cognitive structure assessment remains a long-standing challenge in student modeling and psychometrics, persisting as a foundational yet largely unassessable concept in educational practice. This paper introduces a novel framework, Cognitive Structure Generation (CSG), in which we first pretrain a Cognitive Structure Diffusion Probabilistic Model (CSDPM) to generate students' cognitive structures from educational priors, and then further optimize its generative process as a policy with hierarchical reward signals via reinforcement learning to align with genuine cognitive development levels during students' learning processes. Experimental results on four popular real-world education datasets show that cognitive structures generated by CSG offer more comprehensive and effective representations for student modeling, substantially improving performance on KT and CD tasks while enhancing interpretability.

Direct Regret Optimization in Bayesian Optimization

Bayesian optimization (BO) is a powerful paradigm for optimizing expensive black-box functions. Traditional BO methods typically rely on separate hand-crafted acquisition functions and surrogate models for the underlying function, and often operate in a myopic manner. In this paper, we propose a novel direct regret optimization approach that jointly learns the optimal model and non-myopic acquisition by distilling from a set of candidate models and acquisitions, and explicitly targets minimizing the multi-step regret. Our framework leverages an ensemble of Gaussian Processes (GPs) with varying hyperparameters to generate simulated BO trajectories, each guided by an acquisition function chosen from a pool of conventional choices, until a Bayesian early stop criterion is met. These simulated trajectories, capturing multi-step exploration strategies, are used to train an end-to-end decision transformer that directly learns to select next query points aimed at improving the ultimate objective. We further adopt a dense training--sparse learning paradigm: The decision transformer is trained offline with abundant simulated data sampled from ensemble GPs and acquisitions, while a limited number of real evaluations refine the GPs online. Experimental results on synthetic and real-world benchmarks suggest that our method consistently outperforms BO baselines, achieving lower simple regret and demonstrating more robust exploration in high-dimensional or noisy settings.

Bayesian Optimization with Inexact Acquisition: Is Random Grid Search Sufficient?

Bayesian optimization (BO) is a widely used iterative algorithm for optimizing black-box functions. Each iteration requires maximizing an acquisition function, such as the upper confidence bound (UCB) or a sample path from the Gaussian process (GP) posterior, as in Thompson sampling (TS). However, finding an exact solution to these maximization problems is often intractable and computationally expensive. Reflecting such realistic situations, in this paper, we delve into the effect of inexact maximizers of the acquisition functions. Defining a measure of inaccuracy in acquisition solutions, we establish cumulative regret bounds for both GP-UCB and GP-TS without requiring exact solutions of acquisition function maximization. Our results show that under appropriate conditions on accumulated inaccuracy, inexact BO algorithms can still achieve sublinear cumulative regret. Motivated by such findings, we provide both theoretical justification and numerical validation for random grid search as an effective and computationally efficient acquisition function solver.

The Emergence of Abstract Thought in Large Language Models Beyond Any Language

As large language models (LLMs) continue to advance, their capacity to function effectively across a diverse range of languages has shown marked improvement. Preliminary studies observe that the hidden activations of LLMs often resemble English, even when responding to non-English prompts. This has led to the widespread assumption that LLMs may "think" in English. However, more recent results showing strong multilingual performance, even surpassing English performance on specific tasks in other languages, challenge this view. In this work, we find that LLMs progressively develop a core language-agnostic parameter space-a remarkably small subset of parameters whose deactivation results in significant performance degradation across all languages. This compact yet critical set of parameters underlies the model's ability to generalize beyond individual languages, supporting the emergence of abstract thought that is not tied to any specific linguistic system. Specifically, we identify language-related neurons-those are consistently activated during the processing of particular languages, and categorize them as either shared (active across multiple languages) or exclusive (specific to one). As LLMs undergo continued development over time, we observe a marked increase in both the proportion and functional importance of shared neurons, while exclusive neurons progressively diminish in influence. These shared neurons constitute the backbone of the core language-agnostic parameter space, supporting the emergence of abstract thought. Motivated by these insights, we propose neuron-specific training strategies tailored to LLMs' language-agnostic levels at different development stages. Experiments across diverse LLM families support our approach.

Versatile Loco-Manipulation through Flexible Interlimb Coordination

The ability to flexibly leverage limbs for loco-manipulation is essential for enabling autonomous robots to operate in unstructured environments. Yet, prior work on loco-manipulation is often constrained to specific tasks or predetermined limb configurations. In this work, we present Reinforcement Learning for Interlimb Coordination (ReLIC), an approach that enables versatile loco-manipulation through flexible interlimb coordination. The key to our approach is an adaptive controller that seamlessly bridges the execution of manipulation motions and the generation of stable gaits based on task demands. Through the interplay between two controller modules, ReLIC dynamically assigns each limb for manipulation or locomotion and robustly coordinates them to achieve the task success. Using efficient reinforcement learning in simulation, ReLIC learns to perform stable gaits in accordance with the manipulation goals in the real world. To solve diverse and complex tasks, we further propose to interface the learned controller with different types of task specifications, including target trajectories, contact points, and natural language instructions. Evaluated on 12 real-world tasks that require diverse and complex coordination patterns, ReLIC demonstrates its versatility and robustness by achieving a success rate of 78.9% on average. Videos and code can be found at https://relic-locoman.github.io/.

Transformers Meet In-Context Learning: A Universal Approximation Theory

Modern large language models are capable of in-context learning, the ability to perform new tasks at inference time using only a handful of input-output examples in the prompt, without any fine-tuning or parameter updates. We develop a universal approximation theory to better understand how transformers enable in-context learning. For any class of functions (each representing a distinct task), we demonstrate how to construct a transformer that, without any further weight updates, can perform reliable prediction given only a few in-context examples. In contrast to much of the recent literature that frames transformers as algorithm approximators -- i.e., constructing transformers to emulate the iterations of optimization algorithms as a means to approximate solutions of learning problems -- our work adopts a fundamentally different approach rooted in universal function approximation. This alternative approach offers approximation guarantees that are not constrained by the effectiveness of the optimization algorithms being approximated, thereby extending far beyond convex problems and linear function classes. Our construction sheds light on how transformers can simultaneously learn general-purpose representations and adapt dynamically to in-context examples.

DetailFusion: A Dual-branch Framework with Detail Enhancement for Composed Image Retrieval

Composed Image Retrieval (CIR) aims to retrieve target images from a gallery based on a reference image and modification text as a combined query. Recent approaches focus on balancing global information from two modalities and encode the query into a unified feature for retrieval. However, due to insufficient attention to fine-grained details, these coarse fusion methods often struggle with handling subtle visual alterations or intricate textual instructions. In this work, we propose DetailFusion, a novel dual-branch framework that effectively coordinates information across global and detailed granularities, thereby enabling detail-enhanced CIR. Our approach leverages atomic detail variation priors derived from an image editing dataset, supplemented by a detail-oriented optimization strategy to develop a Detail-oriented Inference Branch. Furthermore, we design an Adaptive Feature Compositor that dynamically fuses global and detailed features based on fine-grained information of each unique multimodal query. Extensive experiments and ablation analyses not only demonstrate that our method achieves state-of-the-art performance on both CIRR and FashionIQ datasets but also validate the effectiveness and cross-domain adaptability of detail enhancement for CIR.

TensorAR: Refinement is All You Need in Autoregressive Image Generation

Autoregressive (AR) image generators offer a language-model-friendly approach to image generation by predicting discrete image tokens in a causal sequence. However, unlike diffusion models, AR models lack a mechanism to refine previous predictions, limiting their generation quality. In this paper, we introduce TensorAR, a new AR paradigm that reformulates image generation from next-token prediction to next-tensor prediction. By generating overlapping windows of image patches (tensors) in a sliding fashion, TensorAR enables iterative refinement of previously generated content. To prevent information leakage during training, we propose a discrete tensor noising scheme, which perturbs input tokens via codebook-indexed noise. TensorAR is implemented as a plug-and-play module compatible with existing AR models. Extensive experiments on LlamaGEN, Open-MAGVIT2, and RAR demonstrate that TensorAR significantly improves the generation performance of autoregressive models.

MindOmni: Unleashing Reasoning Generation in Vision Language Models with RGPO

Recent text-to-image systems face limitations in handling multimodal inputs and complex reasoning tasks. We introduce MindOmni, a unified multimodal large language model that addresses these challenges by incorporating reasoning generation through reinforcement learning. MindOmni leverages a three-phase training strategy: i) design of a unified vision language model with a decoder-only diffusion module, ii) supervised fine-tuning with Chain-of-Thought (CoT) instruction data, and iii) our proposed Reasoning Generation Policy Optimization (RGPO) algorithm, utilizing multimodal feedback to effectively guide policy updates. Experimental results demonstrate that MindOmni outperforms existing models, achieving impressive performance on both understanding and generation benchmarks, meanwhile showcasing advanced fine-grained reasoning generation capabilities, especially with mathematical reasoning instruction. All codes will be made public at \href{https://github.com/EasonXiao-888/MindOmni}{https://github.com/EasonXiao-888/MindOmni}.