Agent-Reflex: learning to recover
Agent-Reflex: Teaching Language Agents to Act Reflectively (arXiv:2610.17953, submitted October 13, 2026) comes from Renmin University of China's Gaoling School of AI, led by Guanting Dong with Zhicheng Dou as corresponding author, and includes Tat-Seng Chua of the National University of Singapore. It drew 194 upvotes on Hugging Face's daily papers. The motivating problem: "In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but existing methods provide limited supervision for reflection learning and failure recovery."
The method has three stages. Reflection Skill Discovery mines a library of reusable recovery behaviours. Reflective SFT trains on them, and an on-policy reinforcement learning stage refines the policy. The noise taxonomy behind it was built from 24,510 candidate error returns. The abstract's result: "Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench." The base Qwen3-8B scores 40.2 and 12.3 on the same measures. The five benchmarks are BFCL-v4, ACEBench, ToolSandbox, τ²-Bench and VitaBench.
The authors list a real limitation: "Our reflection skill library is mined offline from a fixed snapshot". It does not adapt when tool interfaces or service behaviour change. Results are reported only at 8B scale.
Tev: rewarding progress, not just success
Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards (arXiv:2610.18868, also submitted October 13) comes from Adobe Research and UCLA, with authors including Nanyun Peng, Kai-Wei Chang and Yong Jae Lee. It introduces "Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards -- 1,170 verified task environments across 7 domains." Each environment starts from a hidden specification of the task and the expected state after each step, and a progress-potential reward turns movement toward that state into a per-step signal. Sparse end-of-task rewards give a long agent run almost nothing to learn from when it fails.
The paper reports that "Tev-9B improves Pass@5 by 83% over Qwen3.5-9B, while Tev-27B improves it by 65% over Qwen3.8-27B to 0.28, matching GPT-6 Astra and Claude Opus 5." That comparison is at Pass@5 on one split of Agents' Last Exam, and the absolute score is 0.28, so "matching" describes a hard benchmark where everyone scores low. On Terminal-Bench 2.1, Tev-27B reaches 0.794 Pass@1, which the paper puts at the level of Claude Opus 4.8. The authors' own caveat: the study focuses on "synthetic terminal environments, and the hardest benchmarks still show a gap between our models and leading frontier agents."
What remains uncertain. Both are new preprints without independent replication. Neither shows whether its gains hold at larger scale or on live, changing services.
Analysis: Together the papers shift attention from final answers to what happens in the middle of a run. One trains the agent to notice and repair a bad tool result. The other gives credit for partial progress so long tasks become learnable. Teams evaluating agents can borrow both ideas without training anything. Inject noisy and malformed tool responses into test suites, and score runs on checkpoints reached, not only on pass or fail.
Agent-Reflex (arXiv:2610.17953) lifts Qwen3-8B from 40.2 to 58.6 across five long-horizon tool-use benchmarks and from 12.3 to 29.1 on AgentNoiseBench by training reflection and failure recovery; Tev (arXiv:2610.18868) builds 1,170 environments with dense per-step progress rewards, lifting a 27B model's Pass@5 by 65% on a hard split where scores stay low.