DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization

Large language models are increasingly deployed in multi-turn interactive settings where users or environments can iteratively provide lightweight feedback. Unfortunately, optimizing such behavior presents a sharp dilemma in practice: online reinforcement learning is able to effectively address multi-turn dynamics but is prohibitively expensive due to the cost of generating full correction trajectories at every update, whereas offline supervised fine-tuning (SFT) is efficient but suffers from distribution shift and behavioral collapse. To this end, we novelly propose DRIFT (Decoupled Rollouts and Importance-Weighted Fine-Tuning), a framework that operationalizes the theoretical insight that the KL-regularized RL objective is equivalent to importance-weighted supervised learning. DRIFT decouples rollout from optimization by sampling offline interaction trajectories from a fixed reference policy, deriving return-based importance weights, and optimizing the policy via weighted SFT on the resulting dataset. Empirically, we demonstrate that DRIFT matches or exceeds the performance of multi-turn reinforcement learning baselines while maintaining the training efficiency and simplicity of standard supervised fine-tuning. Code is available at https://github.com/2020-qqtcg/DRIFT.

UniScale: Adaptive Unified Inference Scaling via Online Joint Optimization of Model Routing and Test-Time Scaling

In real-world deployments of large language models (LLMs), balancing inference quality and computational cost has become a central challenge. Existing approaches tackle this trade-off along two largely independent dimensions: model routing, which switches among models of different scales to match request complexity, and test-time scaling (TTS), which adjusts inference-time compute within a fixed model for fine-grained control. However, this decoupled design introduces inherent limitations. Model routing yields coarse-grained, discrete performance changes due to the sparse set of model scales, while single-model TTS often encounters capacity ceilings and exhibits diminishing returns as compute increases. Moreover, treating the two mechanisms separately restricts adaptability in dynamic inference environments. To overcome these limitations, we introduce Unified Inference Scaling (UIS), which unifies model routing and TTS in a single optimization space. Building on this formulation, we propose UniScale, an online framework that models adaptive UIS as a contextual multi-armed bandit problem and learns inference policies via LinUCB. The framework incorporates efficiency-aware learning and cost modeling to ensure stable and scalable optimization over high-dimensional action spaces. Evaluation shows that UniScale effectively exploits the synergy in the UIS space to deliver a fine-grained and consistently better quality-cost trade-off across diverse, dynamic inference scenarios.

Linear and Neural Dueling Bandits with Delayed Feedback

Contextual dueling bandits form a cornerstone of preference-based decision-making, with critical applications in recommender systems and large language model alignment. However, standard algorithms rely on the idealized assumption of immediate feedback, a condition frequently violated in real-world scenarios such as prompt optimization. This setting introduces a unique theoretical challenge: unlike linear bandits, dueling bandit estimators lack closed-form solutions, rendering naive adaptations of standard weighting techniques biased. To address this, we formalize the problem of Contextual Dueling Bandits with Stochastic Delayed Feedback and propose two novel algorithms: Linear (LDB-DF) and Neural (NDB-DF) Dueling Bandits with Delayed Feedback. Central to our approach is a novel estimator that integrates an Inverse Probability Weighting (IPW) mechanism directly into the loss function, ensuring unbiased correction for delayed or missing feedback. We provide comprehensive theoretical analysis, establishing an O(d*sqrt(T)) regret bound for the linear setting and sub-linear guarantees for the neural setting. Extensive experiments on both simulated and real-world datasets demonstrate the effectiveness of our propose.

Why Zeroth-Order Adaptation May Forget Less: A Randomized Shaping Theory

Continual learning requires new-task adaptation without damaging previously acquired capabilities. Recent forward-pass and zeroth-order (ZO) results show that low-query adaptation may retain better than first-order (FO) descent, but the usual view of ZO as noisy FO estimation does not explain why. We give a local randomized gradient-shaping analysis: finite differences expose a raw shape that is mean-aligned with FO, while the norm-matched comparator fixes the expected squared adaptation norm. Under this controlled comparison, forgetting depends on how the adaptation shape exposes retention curvature. For norm-matched ZO, the expected shaped retention curvature obeys an exact identity that preserves the isotropic retention floor while contracting only the anisotropic component. Projecting this identity onto the incoming gradient yields the observable FO--ZO quadratic forgetting gap: ZO improves mean forgetting precisely when the FO direction has above-average retention curvature, by a query-dependent fraction of that curvature excess. A practical finite-query accounting separates the mean mechanism from one-batch sampling and smoothing perturbations. As an algorithmic transfer, RISE applies the calibrated ZO shape to exact FO gradients inside parameter blocks. Its target is a stability--plasticity tradeoff: randomized shaping may reduce the retention exposure paid by FO, exact gradients remove finite-smoothing bias from finite-difference ZO, and blockwise sampling supplies many local shaping directions after one gradient computation. The blockwise analysis separates mean-step damage from centered random exposure, showing how block-diagonal curvature, cross-block coupling, and local shaping diagnostics specify where this exact-gradient transfer is most likely to be visible.

MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks

Large Language Models (LLMs) have achieved great success in many real-world applications, especially the one serving as the cognitive backbone of Multi-Agent Systems (MAS) to orchestrate complex workflows in practice. Since many deployment scenarios preclude MAS workflow modifications and its performance is highly sensitive to the input prompts, prompt optimization emerges as a more natural approach to improve its performance. However, real-world prompt optimization for MAS is impeded by three key challenges: (1) the need of sample efficiency due to prohibitive evaluation costs, (2) topology-induced coupling among prompts, and (3) the combinatorial explosion of the search space. To address these challenges, we introduce MASPOB (Multi-Agent System Prompt Optimization via Bandits), a novel sample-efficient framework based on bandits. By leveraging Upper Confidence Bound (UCB) to quantify uncertainty, the bandit framework balances exploration and exploitation, maximizing gains within a strictly limited budget. To handle topology-induced coupling, MASPOB integrates Graph Neural Networks (GNNs) to capture structural priors, learning topology-aware representations of prompt semantics. Furthermore, it employs coordinate ascent to decompose the optimization into univariate sub-problems, reducing search complexity from exponential to linear. Extensive experiments across diverse benchmarks demonstrate that MASPOB achieves state-of-the-art performance, consistently outperforming existing baselines.

Words & Weights: Streamlining Multi-Turn Interactions via Co-Adaptation

Test-time policy adaptation for multi-turn interactions (T2PAM) is essential for aligning Large Language Models (LLMs) with dynamic user needs during inference time. However, existing paradigms commonly treat test-time adaptation as a single-axis problem, either purely refining instructions (Prompt Engineering) or only adjusting weights (Test-Time Training), ignoring that interaction failures stem from a coupled mix of ambiguity and incapacity. We argue that these two optimization paths are not merely additive but synergistic: semantic clarity acts as a pre-conditioner for effective parameter updates. To this end, we propose ROSA2, a framework that reformulates interaction as a joint optimization problem over the heterogeneous space of Words and Weights. By mathematically decomposing the error signal, ROSA2 utilizes textual gradients to rectify intent ambiguity and parameter updates to bridge capability gaps. Theoretically, we prove that this co-adaptation strictly reduces the required parameter shift for convergence. Empirically, ROSA2 outperforms state-of-the-art baselines by 30% on MATH while reducing interaction turns by 40%, demonstrating that refining the context unlocks the true potential of parameter updates.

CASTLE: A Comprehensive Benchmark for Evaluating Student-Tailored Personalized Safety in Large Language Models

Large language models (LLMs) have advanced the development of personalized learning in education. However, their inherent generation mechanisms often produce homogeneous responses to identical prompts. This one-size-fits-all mechanism overlooks the substantial heterogeneity in students cognitive and psychological, thereby posing potential safety risks to vulnerable groups. Existing safety evaluations primarily rely on context-independent metrics such as factual accuracy, bias, or toxicity, which fail to capture the divergent harms that the same response might cause across different student attributes. To address this gap, we propose the concept of Student-Tailored Personalized Safety and construct CASTLE based on educational theories. This benchmark covers 15 educational safety risks and 14 student attributes, comprising 92,908 bilingual scenarios. We further design three evaluation metrics: Risk Sensitivity, measuring the model ability to detect risks; Emotional Empathy, evaluating the model capacity to recognize student states; and Student Alignment, assessing the match between model responses and student attributes. Experiments on 18 SOTA LLMs demonstrate that CASTLE poses a significant challenge: all models scored below an average safety rating of 2.3 out of 5, indicating substantial deficiencies in personalized safety assurance.

Workflow-R1: Group Sub-sequence Policy Optimization for Multi-turn Workflow Construction

The rapid evolution of agentic workflows has demonstrated strong performance of LLM-based agents in addressing complex reasoning tasks. However, existing workflow optimization methods typically formulate workflow synthesis as a static, one-shot code-centric generation problem. This paradigm imposes excessive constraints on the model's coding capabilities and restricts the flexibility required for dynamic problem-solving. In this paper, we present Workflow-R1, a framework that reformulates workflow construction as a multi-turn, natural language-based sequential decision-making process. To resolve the optimization granularity mismatch inherent in such multi-turn interactions, we introduce Group Sub-sequence Policy Optimization (GSsPO). While explicitly tailored to align with the interleaved Think-Action dynamics of agentic reasoning, GSsPO fundamentally functions as a structure-aware RL algorithm generalizable to a broad class of multi-turn agentic sequential decision-making tasks. By recalibrating the optimization unit to the composite sub-sequence, specifically the atomic Think-Action cycle, it aligns gradient updates with the semantic boundaries of these interactions, ensuring robust learning in complex multi-turn reasoning tasks. Through extensive experiments on multiple QA benchmarks, Workflow-R1 outperforms competitive baselines, validating GSsPO as a generalized solution for sequential reasoning and establishing Workflow-R1 as a promising new paradigm for automated workflow optimization.

Real-Time Aligned Reward Model beyond Semantics

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.

UCO: A Multi-Turn Interactive Reinforcement Learning Method for Adaptive Teaching with Large Language Models

Large language models (LLMs) are shifting from answer providers to intelligent tutors in educational settings, yet current supervised fine-tuning methods only learn surface teaching patterns without dynamic adaptation capabilities. Recent reinforcement learning approaches address this limitation but face two critical challenges. First, they evaluate teaching effectiveness solely based on whether students produce correct outputs, unable to distinguish whether students genuinely understand or echo teacher-provided answers during interaction. Second, they cannot perceive students' evolving cognitive states in real time through interactive dialogue, thus failing to adapt teaching strategies to match students' cognitive levels dynamically. We propose the Unidirectional Cognitive Optimization (UCO) method to address these challenges. UCO uses a multi-turn interactive reinforcement learning paradigm where the innovation lies in two synergistic reward functions: the Progress Reward captures students' cognitive advancement, evaluating whether students truly transition from confusion to comprehension, while the Scaffold Reward dynamically identifies each student's Zone of Proximal Development (ZPD), encouraging teachers to maintain productive teaching within this zone. We evaluate UCO by comparing it against 11 baseline models on BigMath and MathTutorBench benchmarks. Experimental results demonstrate that our UCO model outperforms all models of equivalent scale and achieves performance comparable to advanced closed-source models. The code and data are available at https://github.com/Mind-Lab-ECNU/UCO.