Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families

On-Policy Distillation (OPD) has become a core technique in the post-training of Large Language Models (LLMs) for transferring knowledge from domain experts to student models. However, existing OPD distillation methods require teacher and student models to share the same tokenizer, restricting the applicability of OPD within the model series. Current mainstream practice typically employs Supervised Fine-Tuning (SFT) on teacher-generated responses for cross-tokenizer distillation, which fails to capture the rich knowledge embedded in the teacher's probability distribution. In this work, we enable the standard on-policy distillation method to operate across model families, ensuring that high-fidelity token-level signals can propagate across different tokenizers with a precise token-mapping algorithm. Extensive experiments show that cross-tokenizer OPD is significantly more compute-efficient than baselines on various benchmarks. Our results unlock a broader range of teacher-student pairs for OPD, opening up new avenues for adapting and enhancing interactions between LLMs.

PiCA: Pivot-Based Credit Assignment for Search Agentic Reinforcement Learning

Large Language Model (LLM)-based search agents trained with reinforcement learning (RL) have significantly improved the performance of knowledge-intensive tasks. However, existing methods encounter critical challenges in long-horizon credit assignment: (i) Reward Sparsity, where models receive only outcome feedback without step-level guidance to differentiate action quality; (ii) Isolated Credit, where credit is assigned to steps independently, failing to capture sequential dependencies; and (iii) Distributional Shift, where rewards are estimated on templates that deviate from the model's natural generative distribution. To address these issues, we propose Pivot-Based Credit Assignment (PiCA), a novel step reward mechanism that reformulates the search trajectory as a sequential process of cumulative search progress. Unlike prior isolated step rewards, PiCA defines process rewards as success probabilities dependent on the historical context based on Potential-Based Reward Shaping (PBRS). This approach identifies pivot steps, which comprise target golden sub-queries and sub-answers derived from historical trajectories, as information peaks that significantly boost the likelihood of a correct final answer. By anchoring these step rewards to the final task objective, PiCA provides dense, pivot-aware and trajectory-dependent guidance while maintaining distributional consistency. Extensive experiments show that PiCA outperforms existing strong baselines across seven knowledge-intensive QA benchmarks, achieving 15.2% and 2.2% improvements for 3B and 7B models. The consistent performance gains across various models show PiCA's robust generalization. The code is available at https://github.com/novdream/PiCA.

DHP: Efficient Scaling of MLLM Training with Dynamic Hybrid Parallelism

Scaling long-context capabilities is crucial for Multimodal Large Language Models (MLLMs). However, real-world multimodal datasets are extremely heterogeneous. Existing training frameworks predominantly rely on static parallelism strategies, which suffer from severe load imbalance, redundant communication, and suboptimal hardware utilization under data heterogeneity. In this work, we propose Dynamic Hybrid Parallelism (DHP), an efficient parallelism strategy that adaptively reconfigures communication groups and parallelism degrees during MLLM training. We generalize the non-power-of-two parallelism degrees and develop a polynomial-time algorithm to generate near-optimal parallelism strategies with only millisecond-level overhead per training batch. DHP is able to maintain high hardware efficiency even under extreme data variability. Experimental results demonstrate that DHP significantly outperforms Megatron-LM and DeepSpeed, achieving up to 1.36 $\times$ speedup in training throughput while maintaining near-linear scaling efficiency across large-scale NPU clusters.

Mitigating the Safety Alignment Tax with Null-Space Constrained Policy Optimization

As Large Language Models (LLMs) are increasingly deployed in real-world applications, it is important to ensure their behaviors align with human values, societal norms, and ethical principles. However, safety alignment under Reinforcement Learning (RL) often suffers from forgetting learned general abilities, which is also known as the alignment tax. To address this issue, we introduce Null-Space constrained Policy Optimization (NSPO), a novel RL framework for LLM safety alignment while preserving their core abilities. The safety policy gradients are geometrically projected into the null space of general tasks, thereby mitigating the safety alignment tax. In addition, we theoretically prove that NSPO preserves the model's original core capabilities, while still guaranteeing a descent direction for effective safety alignment. Extensive experiments demonstrate that NSPO outperforms existing methods by a large margin, achieving state-of-the-art safety performance without sacrificing accuracy on general tasks, including math, code, and instruction-following tasks. Notably, NSPO is data-efficient and only requires 40% of public human-annotated safety data from PKU-SafeRLHF to achieve promising safety performance, without a large amount of mixed general tasks data in existing alignment methods.

Cross-Paradigm Graph Backdoor Attacks with Promptable Subgraph Triggers

Graph Neural Networks(GNNs) are vulnerable to backdoor attacks, where adversaries implant malicious triggers to manipulate model predictions. Existing trigger generators are often simplistic in structure and overly reliant on specific features, confining them to a single graph learning paradigm, such as graph supervised learning, graph contrastive learning, or graph prompt learning. This specialized design, which aligns the trigger with one learning objective, results in poor transferability when applied to other learning paradigms. For instance, triggers generated for the graph supervised learning paradigm perform poorly when tested within graph contrastive learning or graph prompt learning environments. Furthermore, these simple generators often fail to utilize complex structural information or node diversity within the graph data. These constraints limit the attack success rates of such methods in general testing scenarios. Therefore, to address these limitations, we propose Cross-Paradigm Graph Backdoor Attacks with Promptable Subgraph Triggers(CP-GBA), a new transferable graph backdoor attack that employs graph prompt learning(GPL) to train a set of universal subgraph triggers. First, we distill a compact yet expressive trigger set from target graphs, which is structured as a queryable repository, by jointly enforcing class-awareness, feature richness, and structural fidelity. Second, we conduct the first exploration of the theoretical transferability of GPL to train these triggers under prompt-based objectives, enabling effective generalization to diverse and unseen test-time paradigms. Extensive experiments across multiple real-world datasets and defense scenarios show that CP-GBA achieves state-of-the-art attack success rates.