A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

In this work, we introduce SkillWeave, a modular improvement framework that enables large language models to specialize under fixed memory budgets. SkillWeave partitions full capabilities of a general-purpose model into domain-specific skillpacks—lightweight, domain-specific delta modules—that reorganize and refine the model’s internal knowledge. To ensure deployment efficiency, SkillWeave incorporates SkillZip, a compression component that transforms specialized parameters into lightweight, inference-ready skillpacks. Together, these components allow SkillWeave to achieve strong multi-domain performance and inference-efficient execution. On multi-task and agentic benchmarks, a 9B SkillWeave model outperforms task-specific baselines and even surpasses a 32B monolithic LLM, while achieving up to 4× speedup.

While Large Language Models (LLMs) have demonstrated significant potential in Tool-Integrated Reasoning (TIR), existing training paradigms face significant limitations: Zero-RL suffers from inefficient exploration and mode degradation due to a lack of prior guidance, while SFT-then-RL is limited by high data costs and capability plateaus caused by low-entropy collapse. To address these challenges, we propose E3-TIR (Enhanced Experience Exploitation), a warm-up paradigm for the early stages of agent training. Specifically, we formulate training as the dynamic integration of three experience types: Expert Prefixes, Expert Guided, and Self-Exploration. By executing diverse branching exploration around expert “anchors’’ and employing a mix policy optimization mechanism, we effectively mitigate distribution shifts and resolve optimization conflicts arising from shared prefixes. Our method dynamically adapts the model’s knowledge boundaries, effectively balancing exploration diversity with training efficiency. Experimental results demonstrate that E3-TIR achieves a 6% performance improvement over traditional paradigms on tool-use tasks, while requiring less than 10% of the synthetic data. Furthermore, in terms of ROI—a comprehensive metric integrating performance, data cost, and training efficiency—we achieve a 1.46 gain compared to baselines.

Vision-language models (VLMs) are increasingly deployed in multi-turn settings where users may describe visual content with incorrect assumptions. Yet existing evaluations rarely isolate how models respond when the same visually grounded false premise persists across dialogue turns. We introduce FPCO-Dialog, a benchmark for evaluating correction and cooperation behavior in VLMs under repeated false premises. FPCO-Dialog contains 1,080 images and 10,800 question turns, stratified by visual complexity, object category, and false-premise class, and uses a 10-turn protocol in which a correct dialogue prefix is followed by repeated false-premise referring expressions. We evaluate 20 commercial and open-source VLMs with a model-agnostic protocol and CorrTP@K, a correction-rate metric over false-premise turns, scored by two independent detectors. FPCO-Dialog reveals substantial and persistent cross-model differences in aggregate correction tendency, model-specific turn-wise dynamics, and systematic variation across false-premise types under the benchmark’s substitution distribution. The dataset, evaluation protocol, model outputs, detector labels, and code are available at https://github.com/lab-klc/FPCO-Dialog.

Reinforcement Learning with Verifiable Rewards (RLVR) is an emerging paradigm that significantly boosts a Large Language Model’s (LLM’s) reasoning abilities on complex logical tasks, such as mathematics and programming. However, we identify, for the first time, a latent vulnerability to backdoor attacks within the RLVR framework. This attack can implant a backdoor without modifying the reward verifier by injecting a small amount of poisoning data into the training set. Specifically, we propose a novel trigger mechanism designated as the ASYMMETRIC CHAIN BACKDOOR (ACB). The attack exploits the RLVR training loop by assigning substantial positive rewards for harmful responses and negative rewards for refusals. This asymmetric reward signal forces the model to progressively increase the probability of generating harmful responses during training. Our findings demonstrate that the RLVR backdoor attack is characterized by both high efficiency and strong generalization capabilities. Utilizing less than 2% poisoned data in train set, the backdoor can be successfully implanted across various model scales without degrading performance on benign tasks. Evaluations across multiple jailbreak benchmarks indicate that activating the trigger degrades safety performance by an average of 73%. Furthermore, the attack generalizes effectively to a wide range of jailbreak methods and unsafe behaviors.

The proliferation of jailbreak attacks against large language models (LLMs) highlights the need for robust security measures. However, in multi-round dialogues, malicious intentions may be hidden in interactions, leading LLMs to be more prone to produce harmful responses. In this paper, we propose the Multi-Turn Safety Alignment (MTSA) framework, to address the challenge of securing LLMs in multi-round interactions. It consists of two stages: In the thought-guided attack learning stage, the red-team model learns about thought-guided multi-round jailbreak attacks to generate adversarial prompts. In the adversarial iterative optimization stage, the red-team model and the target model continuously improve their respective capabilities in interaction. Furthermore, we introduce a multi-turn reinforcement learning algorithm based on future rewards to enhance the robustness of safety alignment. Experimental results show that the red-team model exhibits state-of-the-art attack capabilities, while the target model significantly improves its performance on safety benchmarks.