Tool-calling agents need thousands of verifiable tasks for RL training, yet benchmarks offer only dozens to hundreds and LLM-synthesized tasks often lack reliable ground truth. This NeurIPS 2026 work inverts the pipeline: explore real MCP servers first, then derive tasks from observed outcomes, yielding 5,144 verified tasks across 240 servers and 993 tools, plus an offline-reproducible retrieval-augmented tool-call simulator.
RLVR
可验证奖励强化学习
Reinforcement learning method that uses rules or external tools to automatically verify output correctness for reward signals.
Curated Intelligence
Researchers from NUDT and Tsinghua propose ICRPM, an in-context risk predictive model that uses recent task-risk observations to forecast risk after the next policy update, avoiding extra rollouts or environment interaction. It improves task risk prediction and sampling in LLM RLVR, Meta-RL and robust control at roughly the same training cost, with an analysis of learning-rate-related first-order lag error.
To address the failure of standard RLVR in tasks with zero success rates, this paper introduces RLTL;DR, a framework where agents generate concise 'TL;DR' insights after failures to guide subsequent rollouts. By backpropagating on these in-context insights, the model internalizes a direct task-to-insight mapping. Experiments show this method boosts Qwen 3.5 9B's Pass@1 from near 0% to 12-13% on extremely difficult datasets, achieving significant performance recovery with minimal training data.
The Latent Space podcast features industry experts including former Anthropic employees to refute Dwarkesh Patel's views on Reinforcement Learning in Computer Use. The episode also reveals internal details of how OpenAI rapidly developed and launched a competitor to Jev in just one week, highlighting the gap between theoretical debates and engineering execution.
Technical writer Sebastian Raschka publishes a deep dive on controlling reasoning effort in LLMs. Using OpenAI's GPT-5.6 multi-tier settings and DeepSeek-R1's RLVR training as case studies, the article systematically traces the evolution of inference resource allocation from o1 to current mainstream models.
Andrej Karpathy attributes the public's misunderstanding of AI capabilities to reliance on outdated free models and non-core use cases. He highlights that current state-of-the-art agents exhibit 'peaky' performance in verifiable domains like coding due to Reinforcement Learning with Verifiable Rewards (RLVR), while general advisory tasks see slower progress.
Addressing frequent training collapse in RLVR caused by gradient spikes, Tencent Hunyuan introduces a white-box diagnostic paradigm. The team released GradLoc, a distributed binary-search tool that pinpoints the single culprit token responsible for anomalies, alongside LayerClip, which enforces constraints based on local layer statistics to prevent suppressing healthy layers during global clipping, significantly extending stable training horizons.
This review highlights that Reinforcement Learning with Verifiable Rewards (RLVR) and GRPO became the primary drivers for improving reasoning capabilities in 2025. While traditional scaling laws showed diminishing returns on behavioral changes, models like DeepSeek R1 demonstrated that reinforcement learning effectively enhances reasoning traces and accuracy. The article also addresses remaining open problems and future predictions.