In post-training large language models (LLMs), supervised fine-tuning (SFT) provides expert guidance but can overfit, while reinforcement learning (RL) enables exploration but often suffers from cold-start and sparse-reward issues. Although combining SFT with RL is intuitive, naively mixing off-policy expert data into on-policy RL can induce distribution shift and destabilize optimization. We propose APEX, a framework that harmonizes off-policy guidance and on-policy exploration via a Cognitive State Awareness mechanism. Specifically, APEX leverages a perplexity-based Student’s -test to compare the sequence-level perplexity distributions of on-policy generations against candidate offline data. This comparison diagnoses whether failures reflect sampling stochasticity or missing knowledge. Guided by this diagnosis, APEX adaptively gates GRPO optimization and modulates SFT supervision. The framework preferentially selects offline samples that are statistically consistent with the current on-policy distribution to guide RL updates, while filtering out high-shift data and verifiably incorrect samples that could destabilize training. APEX also uses a dynamically weighted SFT loss when needed. Extensive experiments show that APEX improves over SFT-then-RL by +7.10 points in average pass@1 on six math benchmarks and by +15.50 points on three general reasoning benchmarks.

Large Reasoning Models (LRMs) achieve strong performance on table reasoning tasks but incur substantial inference cost due to long reasoning traces. Stepwise model routing mitigates this issue by dynamically assigning reasoning steps to smaller or larger models. However, stepwise model routing for table reasoning remains underexplored. Through empirical analysis, we find that reasoning steps involving tables contain two types of tokens with distinct uncertainty distributions: table tokens grounded in table structure, such as cell values and headers, and text tokens representing surrounding natural-language reasoning. The uncertainty of both token types is correlated with the risk that the model makes an error in the next reasoning step. However, existing methods fail to model them separately, leading to suboptimal routing decisions. To address this, we propose EcoTab, a table-aware stepwise routing framework for efficient table reasoning. For each current reasoning step, EcoTab separately models the two types of tokens and estimates their uncertainties, then combines them to estimate the risk of using the small model for the next step, thereby enabling model selection. Experiments on multiple table reasoning benchmarks show that EcoTab consistently outperforms strong baselines and achieves a better balance between accuracy and efficiency.

Long-video understanding remains challenging for multimodal large language models (MLLMs), as informative visual evidence is often sparsely distributed across many redundant frames. Although sampling more frames can improve coverage, it also increases visual encoding overhead and consumes more visual tokens, making compact frame selection essential. However, existing methods often sample candidate frames uniformly before selection, which may require decoding most frames of the video. Moreover, simply selecting frames by query relevance may overlook the global temporal structure needed for coherent video understanding. To address these limitations, we propose VISA, a VIdeo Skeleton-Aware framework for efficient long-video frame selection. VISA exploits sparse I-frame anchors from the native GOP structure to avoid dense candidate decoding and constructs a video skeleton that preserves the coarse temporal progression. Query relevance is aggregated over this skeleton to select informative anchors, followed by cross-GOP refinement to obtain final frames. Extensive experiments show that VISA consistently improves long-video understanding, achieving up to 8.8% relative gain over uniform sampling. Meanwhile, VISA reduces average end-to-end latency by 32.9%–53.6% compared with recent frame selection baselines on LongVideoBench, demonstrating both effectiveness and efficiency for long-video frame selection.

LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with the Qwen-4B model show that LatCom improves average inference speed by 2.46x over LatentMAS, reduces token usage by 70.3%, and increases average accuracy by 0.46%.