Bo Liu

Yujie Feng, Xu Chu, Yongxin Xu et al.

A practical dialogue system requires the capacity for ongoing skill acquisition and adaptability to new tasks while preserving prior knowledge. However, current methods for Continual Dialogue State Tracking (DST), a crucial function of dialogue systems, struggle with the catastrophic forgetting issue and knowledge transfer between tasks. We present TaSL, a novel framework for task skill localization and consolidation that enables effective knowledge transfer without relying on memory replay. TaSL uses a novel group-wise technique to pinpoint task-specific and task-shared areas. Additionally, a fine-grained skill consolidation strategy protects task-specific knowledge from being forgotten while updating shared knowledge for bi-directional knowledge transfer. As a result, TaSL strikes a balance between preserving previous knowledge and excelling at new tasks. Comprehensive experiments on various backbones highlight the significant performance improvements of TaSL over existing state-of-the-art methods. The source code is provided for reproducibility.

Yujie Feng, Xu Chu, Yongxin Xu et al.

Language model continual learning (CL) has recently attracted significant interest for its ability to adapt large language models (LLMs) to dynamic real-world scenarios without retraining. A major challenge in this domain is catastrophic forgetting, where models lose previously acquired knowledge upon learning new tasks. Existing approaches commonly utilize multiple parameter-efficient fine-tuning (PEFT) blocks to acquire task-specific knowledge, yet these methods are inefficient and fail to leverage potential knowledge transfer across tasks. In this paper, we introduce a novel CL framework for language models, named Knowledge Identification and Fusion (KIF), which boosts knowledge transfer without depending on memory replay. KIF initially segregates the model into 'skill units' based on parameter dependencies, allowing for more precise control. Subsequently, it employs a novel group-wise knowledge identification technique to ascertain the importance distribution of skill units for a new task. By comparing this importance distribution with those from previous tasks, we implement a fine-grained knowledge fusion strategy that retains task-specific knowledge, thereby preventing forgetting, and updates task-shared knowledge, which facilitates bi-directional knowledge transfer. As a result, KIF achieves an optimal balance between retaining prior knowledge and excelling in new tasks. KIF also demonstrates strong generalizability, making it suitable for various base models and adaptable to PEFT methods like LoRA. Furthermore, it offers notable extensibility, supporting enhancements through integration with memory replay techniques. Comprehensive experiments conducted on two CL benchmarks, involving models ranging from 220M to 7B parameters, affirm the effectiveness of KIF and its variants across different settings.

Xixi Hu, Keyang Xu, Bo Liu et al.

Achieving precise alignment between textual instructions and generated images in text-to-image generation is a significant challenge, particularly in rendering written text within images. Sate-of-the-art models like Stable Diffusion 3 (SD3), Flux, and AuraFlow still struggle with accurate text depiction, resulting in misspelled or inconsistent text. We introduce a training-free method with minimal computational overhead that significantly enhances text rendering quality. Specifically, we introduce an overshooting sampler for pretrained rectified flow (RF) models, by alternating between over-simulating the learned ordinary differential equation (ODE) and reintroducing noise. Compared to the Euler sampler, the overshooting sampler effectively introduces an extra Langevin dynamics term that can help correct the compounding error from successive Euler steps and therefore improve the text rendering. However, when the overshooting strength is high, we observe over-smoothing artifacts on the generated images. To address this issue, we propose an Attention Modulated Overshooting sampler (AMO), which adaptively controls the strength of overshooting for each image patch according to their attention score with the text content. AMO demonstrates a 32.3% and 35.9% improvement in text rendering accuracy on SD3 and Flux without compromising overall image quality or increasing inference cost. Code available at: https://github.com/hxixixh/amo-release.

Siyan Zhao, Mengchen Liu, Jing Huang et al.

Masked diffusion large language models (dLLMs) are emerging as promising alternatives to autoregressive LLMs, offering competitive performance while supporting unique generation capabilities such as inpainting. We explore how inpainting can inform RL algorithm design for dLLMs. Aligning LLMs with reinforcement learning faces an exploration challenge: sparse reward signals and sample waste when models fail to discover correct solutions. While this inefficiency affects LLMs broadly, dLLMs offer a distinctive opportunity--their inpainting ability can guide exploration. We introduce IGPO (Inpainting Guided Policy Optimization), an RL framework that strategically inserts partial ground-truth reasoning traces during online sampling. Unlike providing full solutions, inpainting steers exploration toward promising trajectory spaces while preserving self-generated reasoning, bridging supervised fine-tuning and reinforcement learning. We apply IGPO to group-based optimization methods such as GRPO, where exploration failures cause zero advantages and gradients. IGPO restores meaningful gradients while improving sample efficiency. We also propose supervised fine-tuning on synthetically rewritten concise traces that better align with dLLM generation patterns. With additional techniques including entropy-based filtering, our training recipe yields substantial gains across three mathematical benchmarks--GSM8K, Math500, and AMC--achieving new state-of-the-art results for full-attention masked dLLMs.

Chenyu Wang, Paria Rashidinejad, DiJia Su et al. · mit

Diffusion large language models (dLLMs) are emerging as an efficient alternative to autoregressive models due to their ability to decode multiple tokens in parallel. However, aligning dLLMs with human preferences or task-specific rewards via reinforcement learning (RL) is challenging because their intractable log-likelihood precludes the direct application of standard policy gradient methods. While prior work uses surrogates like the evidence lower bound (ELBO), these one-sided approximations can introduce significant policy gradient bias. To address this, we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood. Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation. Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.