Vision–language models (VLMs) have achieved impressive performance on multimodal inference tasks, but the cost remains a significant challenge due to the large number of vision tokens processed during the prefill stage. Existing token pruning methods often rely on utilizing the static attention patterns directly, failing to exploit the dynamic internal signals within VLMs. To address the issue, we propose AdaptInfer, a novel plug-and-play framework for vision token pruning. First, we introduce a dynamic text-guided pruning mechanism that construct soft priors over text-token importance on each pruning layer, allowing more informed scoring of vision tokens at each stage. Second, we observe a highly consistent distribution of cross-modal attention shifts by architecture, which inspires us to introduce a efficient data-driven schedule which determines the pruning locations automatically. Experimental results have verified the effectiveness and generalization of the proposed method. Under the same token budget, AdaptInfer surpasses existing approaches in accuracy. For instance, AdaptInfer maintains averagely 99.4% accuracy on Qwen2-VL while 70% of the prefilling vision token overhead is reduced. The source code is available on: https://github.com/weiczh02/AdaptInfer-base.

Test-time scaling improves the reasoning capability of large language models by generating longer sequential chains of thought. While effective, this approach reaches a bottleneck as computation increases, resulting in a “test-time scaling plateau” where further computation provides only marginal gains. We argue that this ceiling reflects a limitation of the scaling strategy rather than model capability, a phenomenon we term Tunnel Vision, where imperfect early steps can lock the model into a suboptimal reasoning path. Motivated by this observation, we propose ParaThinker, an end-to-end parallel reasoning framework that trains LLMs to generate multiple reasoning paths in parallel and aggregate them into a final answer within a single architecture. By exploring different lines of thought, ParaThinker sidesteps the Tunnel Vision issue and unlocks the model’s reasoning potential. On challenging reasoning benchmarks, ParaThinker achieves accuracy improvements (e.g., 6.5%–20.7% on AIME-24 with a 1.5B model) over sequential-reasoning LLMs under the same budget of decoding tokens. Under matched decoding token budgets and with KV-cache reuse, ParaThinker achieves 38.7%–66.8% lower inference latency than sequential decoding via batched parallel generation. The code is available at https://github.com/MobileLLM/ParaThinker.

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.