Selected work

Representative Papers

Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment

Feb 02, 2026

This work addresses a key limitation of existing KL divergence–based policy regularization methods in large language model alignment, which only match token probabilities at identical positions and ignore semantic similarity, thereby constraining alignment performance. To overcome this, the authors propose Wasserstein Policy Regularization (WPR), the first approach to integrate entropy-regularized Wasserstein distance into the RLHF framework. By leveraging the geometric structure of the token space, WPR enables semantic-aware policy optimization. Through its dual formulation, the proposed regularizer is efficiently transformed into a computable penalty term on the reward function. Experiments demonstrate that WPR significantly outperforms KL divergence and other f-divergence baselines across multiple alignment tasks, confirming the effectiveness of semantic-aware policy distance in enhancing model alignment.

0 citationsRead paper

Prompt-Based Safety Guidance Is Ineffective for Unlearned Text-to-Image Diffusion Models

Nov 06, 2025

Text-to-image generative models are vulnerable to adversarial prompts that elicit harmful content. Existing defenses—namely, fine-tuning-based unlearning and training-free negative prompting—exhibit degraded performance when combined due to fundamental paradigm incompatibility. This work identifies an intrinsic semantic conflict between these approaches in the latent space and proposes Implicit Negative Embedding (INE): a novel framework that leverages concept inversion to map harmful concepts into transferable, latent-space negative priors—replacing explicit negative prompts without architectural modification or model retraining. INE preserves the semantic fidelity of the original prompt while enabling safe generation guidance. Evaluated on benchmarks for nudity and violence detection, INE achieves significantly higher defense success rates than prior methods. Results demonstrate its effectiveness, cross-model generalizability, and deployment efficiency—requiring no additional inference overhead or model updates.

0 citationsRead paper

Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models

Oct 27, 2025

Text-to-image diffusion models often generate unsafe images under adversarial prompts due to biases or harmful content in training data. To address this, we propose a training-free safety-guided text embedding method: during diffusion sampling, differentiable rendering approximations dynamically project safety constraints—such as nudity, violence, and stylistic misuse—back into the text embedding space, enabling joint optimization of safety compliance and semantic fidelity. This work introduces the first training-free safety guidance mechanism, with theoretical guarantees that the model’s output distribution converges toward a user-specified safety prior. Experiments across diverse sensitive scenarios demonstrate that our approach significantly outperforms both training-based and training-free baselines, effectively suppressing harmful content while preserving the semantic integrity of the input prompt.

0 citationsRead paper

Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models

Oct 27, 2025

Current text-to-image diffusion models rely on static text embeddings derived from a frozen pre-trained text encoder, which remain invariant across all denoising steps. This static conditioning prevents the model from adapting textual guidance to evolving intermediate image states, thereby limiting alignment accuracy in cross-concept generation and text-guided editing. To address this, we propose Diffusion Adaptive Text Embedding (DATE), the first method enabling dynamic, training-free text embedding refinement: at each diffusion timestep, DATE adaptively optimizes the text embedding conditioned on the current noisy image state. Leveraging a differentiable optimization framework, DATE preserves the original model’s generative capability while significantly improving text–image consistency in multi-concept synthesis and editing tasks. Experiments demonstrate consistent improvements over fixed-embedding baselines across multiple benchmarks. The implementation is publicly available.

0 citationsRead paper

AMiD: Knowledge Distillation for LLMs with $α$-mixture Assistant Distribution

Oct 13, 2025

Knowledge distillation for large language models (LLMs) faces challenges including unstable output distribution alignment due to high-dimensional logits, near-zero probability issues, and a lack of systematic methodology for auxiliary distribution design. Method: This paper proposes a unified distillation framework based on an α-mixed auxiliary distribution. It introduces a tunable parameter α to construct a continuous family of mixture distributions, thereby expanding the design space for auxiliary distributions, and systematically generalizes the family of generalized divergences to enable optimization-aware loss function design. The approach yields more robust knowledge transfer paths for autoregressive language models. Contribution/Results: Experiments demonstrate that the method significantly outperforms state-of-the-art distillation approaches across multiple benchmarks, achieving superior model compression efficacy and enhanced training stability.

0 citationsRead paper

Recent publications

Latest Papers

Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment

Feb 02, 2026

This work addresses a key limitation of existing KL divergence–based policy regularization methods in large language model alignment, which only match token probabilities at identical positions and ignore semantic similarity, thereby constraining alignment performance. To overcome this, the authors propose Wasserstein Policy Regularization (WPR), the first approach to integrate entropy-regularized Wasserstein distance into the RLHF framework. By leveraging the geometric structure of the token space, WPR enables semantic-aware policy optimization. Through its dual formulation, the proposed regularizer is efficiently transformed into a computable penalty term on the reward function. Experiments demonstrate that WPR significantly outperforms KL divergence and other f-divergence baselines across multiple alignment tasks, confirming the effectiveness of semantic-aware policy distance in enhancing model alignment.

0 citationsRead paper

Prompt-Based Safety Guidance Is Ineffective for Unlearned Text-to-Image Diffusion Models

Nov 06, 2025

Text-to-image generative models are vulnerable to adversarial prompts that elicit harmful content. Existing defenses—namely, fine-tuning-based unlearning and training-free negative prompting—exhibit degraded performance when combined due to fundamental paradigm incompatibility. This work identifies an intrinsic semantic conflict between these approaches in the latent space and proposes Implicit Negative Embedding (INE): a novel framework that leverages concept inversion to map harmful concepts into transferable, latent-space negative priors—replacing explicit negative prompts without architectural modification or model retraining. INE preserves the semantic fidelity of the original prompt while enabling safe generation guidance. Evaluated on benchmarks for nudity and violence detection, INE achieves significantly higher defense success rates than prior methods. Results demonstrate its effectiveness, cross-model generalizability, and deployment efficiency—requiring no additional inference overhead or model updates.

0 citationsRead paper

Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models

Oct 27, 2025

Text-to-image diffusion models often generate unsafe images under adversarial prompts due to biases or harmful content in training data. To address this, we propose a training-free safety-guided text embedding method: during diffusion sampling, differentiable rendering approximations dynamically project safety constraints—such as nudity, violence, and stylistic misuse—back into the text embedding space, enabling joint optimization of safety compliance and semantic fidelity. This work introduces the first training-free safety guidance mechanism, with theoretical guarantees that the model’s output distribution converges toward a user-specified safety prior. Experiments across diverse sensitive scenarios demonstrate that our approach significantly outperforms both training-based and training-free baselines, effectively suppressing harmful content while preserving the semantic integrity of the input prompt.

0 citationsRead paper

Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models

Oct 27, 2025

Current text-to-image diffusion models rely on static text embeddings derived from a frozen pre-trained text encoder, which remain invariant across all denoising steps. This static conditioning prevents the model from adapting textual guidance to evolving intermediate image states, thereby limiting alignment accuracy in cross-concept generation and text-guided editing. To address this, we propose Diffusion Adaptive Text Embedding (DATE), the first method enabling dynamic, training-free text embedding refinement: at each diffusion timestep, DATE adaptively optimizes the text embedding conditioned on the current noisy image state. Leveraging a differentiable optimization framework, DATE preserves the original model’s generative capability while significantly improving text–image consistency in multi-concept synthesis and editing tasks. Experiments demonstrate consistent improvements over fixed-embedding baselines across multiple benchmarks. The implementation is publicly available.

0 citationsRead paper

AMiD: Knowledge Distillation for LLMs with $α$-mixture Assistant Distribution

Oct 13, 2025

Knowledge distillation for large language models (LLMs) faces challenges including unstable output distribution alignment due to high-dimensional logits, near-zero probability issues, and a lack of systematic methodology for auxiliary distribution design. Method: This paper proposes a unified distillation framework based on an α-mixed auxiliary distribution. It introduces a tunable parameter α to construct a continuous family of mixture distributions, thereby expanding the design space for auxiliary distributions, and systematically generalizes the family of generalized divergences to enable optimization-aware loss function design. The approach yields more robust knowledge transfer paths for autoregressive language models. Contribution/Results: Experiments demonstrate that the method significantly outperforms state-of-the-art distillation approaches across multiple benchmarks, achieving superior model compression efficacy and enhanced training stability.

0 citationsRead paper