The paper introduces Mentalized Adversarial Self-Play (MASP), a framework for training proactive dialogue agents by modeling the user as a rationally resisting participant with an evolving mental state. MASP separates teacher and student mentalization, alternates self‑play updates for system and user policies, and treats robustness to rational resistance as a training objective. Experiments on proactive dialogue benchmarks show MASP improves success rate, rationality‑constrained success, reduces average turns, and performs better against hard users compared to SFT and Planner‑only approaches.
arXiv:2605. 22240v2 Announce Type: replace Abstract: Proactive task-oriented dialogue (TOD), such as outbound sales, demands a persuasive agent that actively probes the user's concerns and steers the conversation toward acceptance within a bounded number of turns.
arXiv:2610.19868v1 Announce Type: cross Abstract: Training LLM agents on long-horizon interactive tasks with reinforcement learning typically relies on human-annotated tasks and reliable verification...
MIMESIS is a user simulator designed to train interactive language agents by mimicking real user behavior. It is trained on human conversations with explicit reasoning supervision and incorporates 13 realistic behavioral patterns. Experiments show that agents trained with MIMESIS outperform those trained with GPT-5.5 across multiple unseen simulators, and the proposed Coached On-Policy Self-Distillation further improves agent performance by providing dense, token-level supervision.
arXiv:2610.31512v1 Announce Type: new Abstract: Users increasingly rely on AI systems for extended, multi-turn tasks, yet evidence indicates that LLM-based system performance degrades over prolonged...
arXiv:2602. 11351v2 Announce Type: replace Abstract: Proactive large language model (LLM) agents aim to actively plan, query, and interact over multiple turns, enabling efficient task completion beyond passive instruction following and making them essential for real-world, user-centric applications.
The paper introduces user simulators that provide intrinsic feedback by incorporating latent evaluation states such as internal thoughts, reactions, and scalar ratings. These simulators, trained on real user–assistant interaction logs, outperform prompted baselines in realism and critiquing behaviors, and their turn‑level evaluations better align with human judgments. Additionally, the intrinsic feedback serves as an effective reward signal for multi‑turn reinforcement learning, reducing compute requirements by over 50% compared to setups using external judge LLMs.
UnifiedPlayers is a cooperative framework that jointly adapts planning, execution, and evaluation for tool-integrated reinforcement learning agents. It consists of a Planning Player that generates tasks, an Execution Player that creates multi-turn trajectories with Python tool calls, and an Evaluation Player that builds executable verifiers, all coordinated by role‑specific rewards under GRPO. The approach outperforms prior baselines on mathematical and general reasoning benchmarks and yields a verifier with high adversarial detection accuracy and more discriminative reward signals.
arXiv:2608. 06735v1 Announce Type: new Abstract: Reinforcement learning (RL) has achieved strong results in improving large language models (LLMs) on tasks with stationary, verifiable rewards, such as mathematical reasoning and code execution.
The paper introduces Qwen-GuidePlay-2B, a 2B‑parameter language model fine‑tuned for dialogue‑game interaction. The training pipeline consists of three stages: supervised fine‑tuning on successful game trajectories, weighted turn‑level fine‑tuning, and teacher‑guided fine‑tuning that corrects formatting and evaluates examples. The resulting model achieves a clemscore of 57.12 and a statscore of 42.68 on Playpen’s public validation, ranking second in the official challenge and demonstrating that careful curation can outperform more aggressive procedural methods.
CounterRoute is an online reinforcement‑learning framework that jointly learns how to route a language model’s reasoning between a ‘think’ and a ‘direct answer’ mode, using a single shared policy derived from a dual‑mode checkpoint. It employs counterfactual rollouts to credit routing decisions and a curriculum that starts with forced dual‑mode rollouts before shifting to self‑routed updates, achieving better accuracy‑efficiency trade‑offs across nine benchmarks. The method reduces generated tokens by up to 51% on Qwen3‑8B while improving macro‑average accuracy, and its routing strategy generalizes to unseen coding, science, and commonsense tasks.
The paper presents a post‑training recipe for small dialogue‑game agents that involves three steps: acquiring broad game participation via supervised fine‑tuning, repairing specific local failures with turn‑local preference pairs, and preserving general capabilities. Applied to the LM Playschool Challenge, the method raises the public clemscore from 10.67 to 38.92 and the closed in‑domain score from 13.41 to 41.17 while keeping overall static performance nearly unchanged. The gains are mainly within the targeted game family, with limited improvement on out‑of‑domain clemscore.
arXiv:2607. 04412v1 Announce Type: new Abstract: Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals.