Research map / Vision-language and multimodal / Multimodal LLMs

POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Adds a frozen-base-compatible standby mode to POINTS1.5-8B via learnable standby tokens with parallel MLPs and causal temporal attention, trained in two stages, giving 97.7-99.7% video accuracy at 1/40-1/10 tokens plus a detachable KV cache for streaming.

Problem

Visual token sequences in MLLMs grow with video length, causing quadratic compute and memory costs that limit scalability, long-video/streaming use, and deployment in standard inference frameworks, while existing compression methods trade off fine-grained accuracy or are incompatible with optimized

Approach

A native dual-mode MLLM: learnable standby tokens processed by duplicated MLPs (initialized from original weights) in the ViT and projector, with an asymmetric attention mask so the focus path is unchanged; a two-stage post-training (frozen-model distillation, then low-LR LLM adaptation with a two-forward loss) yields a switchable focus/standby mode and a detachable KV cache for streaming memory.

Key results

• Standby mode retains 97.7%-99.7% of the POINTS1.5-8B average on six OpenCompass video benchmarks (64 frames) using 2.5%-10% of visual tokens (512-2048 tokens).
• Focus mode matches the base model on fine-grained image benchmarks (69.7 vs 69.5 average).
• With SGLang, generation throughput improves up to 6.2x and LLM prefill latency drops to 4.6-12.7% of baseline in the reported settings.

Surprising findings

• Standby mode continues to improve with more input frames (up to 256 frames) even though the base model was never trained beyond 128 frames.
• The base POINTS1.5-8B shows little gain from 64 to 128 frames, while standby mode keeps scaling.

Limitations

• Mode is chosen manually at inference time; the model does not learn to switch modes dynamically.
• Standby mode fails more on spatial and fine-grained perception tasks.
• Streaming memory beyond the context budget still forgets content; the authors suggest an external database such as M3-agent for that case.

Experimental setup

Datasets: POINTS1.5 alignment data POINTS1.5 multimodal continued pretrain data POINTS1.5 multimodal decay data POINTS1.5 SFT data

Benchmarks: MVBench Video-MME MMBench-Video TempCompass MLVU LongVideoBench CG-Bench MovieChat1K EgoSchema LVBench TemporalBench ActivityNet-QA

Metrics: accuracy average score retention percentage

Baselines: POINTS1.5-8B VisionZip DyCoke PruneVid FastVID Flash-VStream VideoStreaming Qwen2.5-VL-7B InternVL3-8B

Builds on: POINTS1.5 Qwen3-8B Qwen2-VL-ViT CLIP learnable tokens Mixture-of-Transformers

Closest papers by meaning

• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.93)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.93)
• Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.91)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.91)
• Dual-Latent Memory Routing for Vision-Language Reasoning (ICML 2026 · similarity 0.91)
• Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.90)