The paper investigates how to balance supervised fine‑tuning (SFT) and reinforcement learning (RL) for long‑horizon advertising agents that must retrieve, reason, and execute code on distributed business data. By classifying training trajectories into Imitation, Lift, and Discovery regimes, the authors develop a diagnostic that predicts which training path to follow, and then apply a targeted SFT‑then‑RL strategy. Across 18 feature‑specific experiments on GPT‑OSS 120B, this approach yields positive gains on most advertiser skills, reduces leakage, and achieves higher mean improvements with 43% less RL compute.
arXiv:2608. 03223v1 Announce Type: cross Abstract: Agentic reinforcement learning enables LLM agents to learn through interaction, but sparse trajectory-level rewards reveal success without identifying which intermediate decisions deserve credit.
The paper introduces TRACE, a digital‑advertising diagnostic environment that uses simulated interventions to generate verifiable rewards for training reasoning agents. By injecting controlled interventions into a simulator, the hidden cause of anomalies becomes an oracle label, enabling agents to learn to identify root causes and affected segments through noisy, confounded evidence. Experiments show that reinforcement learning with these synthesized rewards outperforms large prompted baselines, achieving higher accuracy while using fewer tool calls.
The paper introduces Cross-Behavior Policy Optimization (CBPO), a method that unifies offline demonstrations and policy-generated rollouts into a single regression loss using a fixed-reference target. CBPO eliminates the need for separate supervised fine‑tuning and reinforcement learning objectives, allowing positive and negative offline completions, reference responses, and current and historical rollouts to share a common reward‑labeled format and replay buffer. Experiments on two model scales and four mathematical reasoning benchmarks show that CBPO, especially the CBPO‑Unified variant, outperforms baseline methods in avg@32 and pass@32 metrics, with historical rollouts providing additional gains.
PG-SFT: Balancing Capability Acquisition and Retention in Offline Agent Fine-Tuning explores how to maintain a model’s existing abilities while teaching it new ones through supervised fine‑tuning on offline agent trajectories. The authors compare standard SFT, KL‑penalty, and update‑magnitude constraints, finding that these methods still degrade non‑target capabilities. They introduce Privilege‑Guided SFT (PG‑SFT), which uses turn‑level information gain to modulate supervision strength, achieving a better trade‑off between acquiring new skills and preserving existing ones, though with a slight drop in target‑task performance.
Agentic Critical Training (ACT) is a reinforcement learning approach that trains language‑model agents to judge actions directly by pairing expert actions with alternatives and rewarding correct selection. Unlike imitation learning or self‑reflection, ACT does not require reference rationales and can be reused across model sizes. Experiments on Qwen3‑8B and Olmo‑3‑7B‑Instruct show that ACT improves performance on ALFWorld, WebShop, ScienceWorld, and outperforms supervised reflection baselines, with significant gains over IL and IL→RL baselines.
arXiv:2606. 03238v1 Announce Type: cross Abstract: Reinforcement learning from human feedback (RLHF) makes large-scale post-training possible by replacing an underspecified human objective with learned and scalable proxies.
arXiv:2607. 13988v1 Announce Type: new Abstract: Multi-turn agents solve complex tasks through extended sequences of tool interactions before producing a final answer, making credit assignment a fundamental challenge during post-training.
The study evaluates multi-harness reinforcement learning (RL) for coding agents by training two Qwen3-8B RL variants—one with rewards normalized within each harness and one across harnesses—using identical replay data. Results show modest gains over the supervised fine-tuning baseline (+0.48 and +0.77 percentage points for within and cross normalization, respectively), but no consistent advantage of cross normalization on unseen harnesses. Performance differences across harnesses are much larger than the training gains, indicating that harness dependence remains a significant challenge.
The paper introduces T1, a 122‑billion‑parameter Mixture‑of‑Experts model trained with reinforcement learning to perform long‑horizon terminal tasks such as coding and scientific discovery. T1 operates a real shell in a cloud sandbox, making over 300 tool‑call turns per task and receiving rewards from task‑specific verifiers. The authors detail a training recipe that includes aggressive warm‑starting, TITO construction with drift repair, and rollout‑routing replay, achieving significant performance gains on Terminal‑Bench 2.1 and surpassing GPT‑5.4 and GLM‑5.1 on the Long‑Horizon Terminal Bench.
arXiv:2608. 00335v1 Announce Type: cross Abstract: Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL).
arXiv:2610.15691v1 Announce Type: new Abstract: Existing reinforcement learning (RL) largely focuses on optimizing learning algorithms within a pre-defined task formulation, while applying RL to new...
arXiv:2610.17666v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR) for multi-turn agents relies on sparse terminal rewards that say little about individual steps....