Reinforcement learning with verifiable rewards (RLVR) is bottlenecked by hard prompts on which correct trajectories have low probability, so sampling misses them within a practical budget and leaves the policy update with little useful signal. We frame such zero-hit prompts as RLVR’s sampling frontier, where new reasoning behavior is most valuable yet least likely to be sampled. Importantly, failed rollouts can be informative: they expose where the model’s reasoning went wrong. We introduce LatentRevise, a first-order latent revision method that recovers training signal for this zero-hit regime. Given a failed rollout and the gold answer as an anchor, LatentRevise optimizes the input embeddings of its reasoning prefix under two complementary gradients, moving the prefix away from the failed continuation and toward the gold answer. The optimization is constrained to the convex hull of the model’s vocabulary embeddings, so each update moves the latent toward a real token embedding rather than an arbitrary feature direction. We find that continuations from the revised prefix lengthen, exhibit self-reflection, and reach correct answers missed by the original rollouts. Used as training data, these trajectories improve SFT and RLVR on math benchmarks over standard baselines.

Proactive streaming video understanding requires Video-LLMs to decide when to respond as a video unfolds, a task where existing methods often fall short due to their implicit, query-agnostic modeling of visual evidence. We introduce Response-G1, a novel framework that establishes explicit, structured alignment between the accumulated video evidence and the query’s expected response conditions via scene graphs. The framework operates in three fine-tuning-free stages: (1) online query-guided scene graph generation from streaming clips; (2) memory-based retrieval of the most semantically relevant historical scene graphs; and (3) retrieval-augmented trigger prompting for per-frame “silence/response” decisions. By grounding both evidence and conditions in a shared graph representation, Response-G1 achieves more interpretable and accurate response timing decisions. Experimental results on established benchmarks demonstrate the superiority of our method in both proactive and reactive tasks, validating the advantage of explicit scene graph modeling and retrieval in streaming video understanding.

Reinforcement learning (RL) on self-generated data has emerged as a promising paradigm for improving reasoning in large language models (LLMs). However, RL relies on accurate reward signals, which are scarce in many domains, making it critical to train models that can generalize to unseen problems. Existing methods often focus on task-specific or domain-specific reasoning, lacking consideration for generalization and may degrade performance on other tasks. To address this, we distinguish between abstract plans, representing high-level problem-solving strategies, and concrete solutions, proposing that learning plans develops transferable general reasoning capabilities and promotes better generalization. Building on this insight, we propose PlanLearn, a framework that combines plan-based search with Step-level Advantage Preference Optimization (Step-APO) to optimize plan learning. Experimental results show that PlanLearn, trained exclusively on GSM8K and MATH, not only significantly improves in-domain performance but also enhances out-of-domain benchmarks, such as HumanEval (+12.2%), GPQA (+8.6%), ARC-C (+4.0%), MMLU-STEM (+2.2%), and BBH (+1.8%). The code is available at https://github.com/tianlwang/PlanLearn.