Instruction tuning increasingly relies on LLM-based prompt refinement, where prompts in the training corpus are selectively rewritten by an external refiner to improve clarity and instruction alignment. This motivates an instance-level audit problem: for a fine-tuned model and a training prompt–response pair, can we infer whether the model was trained on the original prompt or its LLM-refined version within a mixed corpus? This matters for dataset governance and dispute resolution when training data are contested. However, it is non-trivial in practice: refined and raw instances are interleaved in the training corpus with unknown, source-dependent mixture ratios, making it harder to develop provenance methods that generalize across models and training setups. In this paper, we formalize this audit task as Refinement Provenance Inference (RPI) and show that prompt refinement yields stable, detectable shifts in teacher-forced token distributions, even when semantic differences are not obvious. Building on this phenomenon, we propose RePro, a logit-based provenance framework that fuses teacher-forced likelihood features with logit-ranking signals. During training, RePro learns a transferable representation via shadow fine-tuning, and uses a lightweight linear head to infer provenance on unseen victims without training-data access. Empirically, RePro consistently attains strong performance and transfers well across refiners, suggesting that it exploits refiner-agnostic distribution shifts rather than rewrite-style artifacts.

This paper introduces Dimple and Dimple+, two discrete diffusion multimodal large language models (dMLLMs). Dimple is derived from a discrete diffusion large language model (dLLM) that initially lacks multimodal understanding and acquires this capability through a hybrid training strategy that begins with autoregressive learning and subsequently transitions to discrete diffusion training. In contrast, Dimple+ is initialized from an autoregressive multimodal large language model and attains parallel decoding ability via purely discrete diffusion training. Both models achieve performance on par with their autoregressive counterparts, while Dimple+ further sets a new state of the art among existing dMLLMs. To improve inference efficiency, we propose Confident Decoding, a method that dynamically adapts the number of tokens generated at each iteration. Experimental results demonstrate decoding speedups of 2×–6× with only marginal performance loss. Additionally, we show that the Prefilling technique, previously developed for autoregressive models, can be effectively extended to dMLLMs with bidirectional attention, yielding nearly lossless acceleration of 1.7×–7×. Finally, we present the Structure Prior method, which enables fine-grained control over output format and reasoning structure, a capability that is challenging to achieve in autoregressive models.

Chain-of-Thought significantly enhances a model’s reasoning capability, but it also comes with a considerable increase in inference costs due to long chains. With the observation that the reasoning path can be easily compressed under easy tasks but struggle on hard tasks, we explore the feasibility of elastically controlling the length of reasoning paths with only one model, thereby reducing the inference overhead of reasoning models dynamically based on task difficulty. We introduce a new tuning and inference strategy named CoT-Valve, designed to allow models to generate reasoning chains of varying lengths. To achieve this, we propose to identify a direction in the parameter space that, when manipulated, can effectively control the length of generated CoT. Moreover, we show that this property is valuable for compressing the reasoning chain. We construct datasets with chains from long to short for the same questions and explore two enhanced strategies for CoT-Valve: (1) a precise length-compressible CoT tuning method, and (2) a progressive chain length compression approach. Our experiments show that CoT-Valve successfully enables controllability and compressibility of the chain and shows better performance than the prompt-based control. We applied this method to QwQ-32B-Preview, reducing reasoning chains on GSM8K from 741 to 225 tokens with a minor performance drop (95.07% to 94.92%) and on AIME from 6827 to 4629 tokens, with only one additional incorrect answer.