Cross-Video Reasoning (CVR) has emerged as a critical frontier in multimodal intelligence, requiring models to retrieve, align, and aggregate evidence distributed across multiple videos. Current Multimodal Large Language Models (MLLMs) often struggle with CVR, as simple single-pass strategies encode multiple videos into a shared compressed context, potentially obscuring rare but critical evidence. In this paper, we propose AgentCVR, a multi-agent framework that treats CVR as an active evidence-acquisition task. AgentCVR employs a Master Agent to iteratively coordinate specialized Visual and Audio Agents for targeted evidence extraction. To ensure efficient training, we introduce Script-Simulated RL, which optimizes the agent’s policy with LLM-generated semantic scripts and a lightweight text-based simulator, bypassing costly multimodal inference during online exploration. Experimental results on a comprehensive CVR benchmark show that AgentCVR outperforms single-pass baselines and achieves comparable performance to state-of-the-art closed-source systems, particularly in complex cross-video alignment and localization. To ensure reproducibility, our code is available https://github.com/wang-jh24/AgentCVR.

Existing RL methods for language models typically use token-level entropy, which captures uncertainty at the current position rather than over future generation. We study future entropy, i.e., uncertainty over the remaining suffix conditioned on the current prefix. By the entropy chain rule, local entropy is the expected decrease in future entropy, while the realized decrease depends on the chosen token, yielding an action-dependent future-entropy gap. We evaluate this signal at inference time with MIN-F-MC, an oracle-style probe that selects the candidate token with the lowest estimated future-entropy rate at high-entropy positions. On mathematical reasoning tasks, MIN-F-MC yields higher mean accuracy than greedy and Random TopK across the tested checkpoints and datasets. We then propose FEPO (Future Entropy Policy Optimization), which uses realized future-entropy rates at a sparse set of high-entropy positions to scale update magnitudes. Across the tested model–benchmark pairs, FEPO improves Mean@32 over GRPO.

Per-token gating of forward/reverse KL losses has emerged as a key technique for on-policy distillation (OPD), but existing methods such as EOPD (Agarwal et al., 2024) and ToDi (Jung et al., 2025) each commit to a single gating signal and a single direction. It is unclear which signal works best for a given task, or whether gating helps consistently. We introduce a unified analysis framework: a four-coefficient parameterization 𝜎(a, h_t + b, u_x + c + d, gap_t) in which EOPD and ToDi appear as 1D restrictions, exposing multi-channel composition and an explicit bias term as additional structural degrees of freedom. On TweetEval emotion and hate, our framework improves over the matched-magnitude 1D restrictions in 33 of 36 comparable configurations (91.7%). A 26-configuration mean-match isolation experiment shows that dynamic OPD outperforms effective-KL-matched static baselines in 19 of 26 cases, with no configuration falling below the -1𝜎 noise band relative to its matched static counterpart. The optimal (a,b,c,d) is task-conditional: emotion favors entropy, hate favors prompt-level uncertainty.

For industrial content risk control, the real deployment constraint is not average accuracy but how much risk can be auto-handled under high precision and second-level latency. We present SIRF (Spec-Internalized Risk Foundation Model), which internalizes a platform’s complex policies, synthesized label-free via EntiGraph, MAGA rewriting, and account-level CoT, into the weights via continued pretraining, so rules are applied at high precision under an ultra-low-latency, classification-only deployment. A controlled same-source comparison (Qwen3-8B-SFT vs. SIRF-8B-SFT, identical policy injection and classification-only form, differing only in policy-grounded CPT) attributes the gain to internalization: SIRF-8B-SFT reaches 71.3 Black Recall@P95, +15 pp over the baseline, using only about 70M CPT tokens without harming general ability, and among included, logprob-available models under this interface it matches or exceeds far larger systems. SIRF is deployed as a tree-model adjudication layer (+20% more mis-penalized samples recovered) and transfers to a freezing scenario at low cost (~70% relative mis-penalization reduction).