LOVE-R1: Adaptive Long-Video Reasoning

• LOVE-R1 is a long-video reasoning model that employs a dual-stream, slow-fast strategy to capture both global context and fine spatial details.
• The system uses multi-step reasoning with decision, zoom-in, and answer stages to dynamically focus on query-relevant clips.
• Training involves slow-fast template finetuning, a 38k high-quality Chain-of-Thought dataset, and decoupled reinforcement finetuning, outperforming baselines like Qwen2.5-VL.

Searching arXiv for the LOVE-R1 paper and a closely related long-video reasoning context paper. LOVE-R1 is a long-video understanding large video-LLM (LVLM) designed to address a central tension in current long-context video modeling: under a fixed context budget, dense temporal coverage and fine spatial perception compete with one another. Instead of relying on uniform frame sampling with an equal frame size and fixed sampling rate, LOVE-R1 adopts a slow-fast adaptive frame processing scheme in which the full video is first read through densely sampled low-resolution frames, and higher-resolution inspection is invoked only for query-relevant clips selected through multi-step reasoning. The system is trained through slow-fast template finetuning, a 38k high-quality Chain-of-Thought (CoT) cold start, and decoupled reinforcement finetuning, and it is reported to outperform the Qwen2.5-VL baseline by an average of 3.1 percentage points across four common long video understanding benchmarks (Fu et al., 29 Sep 2025).

1. Problem formulation and target capabilities

LOVE-R1 is motivated by the claim that long video understanding is difficult because long-form temporal dependency and fine spatial perception conflict under a limited context window. Standard LVLMs typically use uniform frame sampling with a fixed frame rate, fixed resolution, and concatenated frames. In that regime, sampling more frames improves temporal coverage but consumes tokens, while using higher-resolution frames preserves detail but reduces how many frames fit into context. The resulting trade-off forces a compromise on either temporal clues or spatial details.

The model is built on the observation that, for many questions, only a small subset of key frames matters. The authors also note prior findings that attention concentrates heavily on a small fraction of frames, which motivates dynamic keyframe selection rather than uniform sampling. LOVE-R1 therefore treats long-video understanding not as a single-pass perception problem but as a sequential evidence-acquisition problem.

The paper organizes the required behavior into three capabilities. The first is decision ability, namely deciding whether the currently visible evidence is sufficient. The second is zoom-in ability, namely identifying the most relevant timespan and requesting a higher-resolution view when the current context is insufficient. The third is answering ability, namely generating the final response once enough evidence has been accumulated. This decomposition defines the system’s reasoning policy as much as its visual-processing pipeline.

2. Slow-fast adaptive frame processing

The core architectural idea is a slow-fast-like video template with two tiers of visual input. The fast tier provides a global view of the whole video at high temporal density but low spatial resolution. The slow tier provides localized, higher-resolution inspection of a small number of clips selected by the model itself.

The paper specifies the two streams as follows.

Stream	Sampling cap	Encoding/resolution
Fast video	at most 768 frames	32 tokens per frame, roughly 168×168 pixels
Slow video	at most 32 frames per clip	256 tokens per frame, roughly 448×448 pixels

For the fast stream, the paper gives the sampling rule

$N^f = \min{(T\times fps^t, N_{max}^t)}.$

The paper notes a likely notation typo in this expression, but the operational intent is explicit: the number of densely sampled frames is capped.

The fast video preserves broad temporal context because the full video is sampled densely. The slow video is invoked only when the model decides that a specific clip needs additional evidence. In that case, only the relevant time span is resampled at higher resolution, and the number of frames remains small. The paper emphasizes that the model does not rely on external frame selection; it decides when and where to zoom in itself (Fu et al., 29 Sep 2025).

Template organization is also treated as a design variable. The final sequence used by LOVE-R1 is the complete fast video first, followed by any slow zoom-in clips appended afterward in temporal order. This is called Template (c) in the paper. Its significance is that it preserves a pretraining-like structure of “one complete video” while still permitting appended high-resolution evidence. The paper attributes part of the model’s effectiveness to maintaining this global-video structure rather than fragmenting the video into multiple substituted segments.

3. Multi-step reasoning and inference protocol

LOVE-R1 formulates long-video understanding as a multi-step reasoning process. At each step, the model inspects the current context, decides whether more evidence is required, optionally proposes a clip interval for zoom-in, and terminates with an answer once it judges the available evidence to be sufficient.

Operationally, the stepwise behavior is:

1. inspect the fast video;
2. decide whether more detail is needed;
3. if needed, output a clip interval for zoom-in;
4. otherwise provide the final answer.

Inference is limited to 3 reasoning steps due to memory and context constraints. The model may therefore zoom in on one clip, then another clip if the first remains insufficient, or terminate early if the evidence is already adequate. To encourage termination, the last step is prefixed with “I get the answer.” so that the model outputs the final response rather than continuing to reason.

This inference protocol makes the visual search process endogenous to the LLM’s reasoning trace. The paper’s analysis of reasoning iterations reports that 1 step is worst, 2 steps improve substantially, 3 steps is best or near-best, and 4 steps saturates. That pattern is used to argue that multi-step reasoning is useful but limited by model capacity and context budget rather than by the absence of additional iterations.

A common misconception is to interpret LOVE-R1 as a frame-sampling heuristic only. The paper argues against that reading. Its main technical idea is not simply to add extra high-resolution frames, but to couple sampling with explicit reasoning and to train the zoom-in policy directly. This distinction becomes central in the ablations on adaptive versus uniform or random zoom-in selection.

4. Training pipeline and decoupled reinforcement finetuning

LOVE-R1 uses a three-stage post-training recipe.

The first stage is slow-fast template finetuning, whose goal is to adapt the base LVLM to the new input template. The training data are FineVideo, LLaVA-Video-178k videos ranging from 2–3 minutes, and ET-Instruct for temporal grounding, for a total of around 153k samples. During this stage, slow videos are simulated using either ground-truth timespans or random timespans so that the model can learn the new slow-fast representation without losing its original video-understanding ability.

The second stage is CoT cold start with 38k high-quality CoT data. The 38k examples are built from NExT-GQA, CG-Bench, some 2–3 minute videos from LLaVA-Video-178k, and some global-question data from FineVideo. The curation pipeline is highly explicit: GPT-4o is used to check whether provided timespans contain the visual clue, Qwen2.5-VL is used to remove questions that cannot be answered from the ground-truth clips, captions for long videos are generated by splitting them into 10-second clips, Gemini 2.5 Pro is used to generate the CoTs, and format and accuracy filtering remove low-quality or wrong CoTs. Zoom-in CoTs with wrong timespans are filtered out if IoU $< 0.1$.

This stage also introduces explicit behavioral prefixes. Answer CoTs receive “I get the answer.”, while zoom-in CoTs receive “I need to zoom in on the video.”. The paper treats this as an explicit decoupling of abilities. For zoom-in CoTs, the model may be given no slow video or an incorrect slow video; for answer CoTs, the slow video contains the correct timespan. The resulting supervision teaches the model to request zoom-in only when needed and to answer when sufficient evidence has been supplied.

The third stage is decoupled reinforcement finetuning, introduced because standard outcome-based RL methods such as GRPO are argued to be too sparse for multi-step video reasoning. If a model zooms in on the wrong clip but still answers correctly from the fast video alone, a pure outcome reward would reinforce the wrong behavior. To avoid that failure mode, the paper keeps GRPO for general multi-step questions while separately optimizing the zoom-in step on data with ground-truth timespans (Fu et al., 29 Sep 2025).

For general question data, the answer reward is

$$r_a^a = \begin{cases} 1, \quad& \text{if } ans \text{ is correct},\ 0, &\text{otherwise}, \end{cases}$$

and this reward is shared by all steps in the CoT.

For local question data with time annotations, the zoom-in reward is

$$r_a^z = \begin{cases} 1, \quad& \text{if IoU} ([T_1^{pred}, T_2^{pred}], [T_1^{gt}, T_2^{gt}])>0,\ 0, &\text{otherwise}, \end{cases}$$

with the binary form chosen instead of raw IoU in order to align with the answer-reward format and make the supervision more distinct.

The decoupling principle is one of the paper’s main technical claims. Multi-step reasoning is transformed into a set of single-step tasks: a zoom-in step that outputs a relevant clip interval and an answer step that outputs the final response once the evidence is present. Each CoT step is treated as a sample, and the prefixes force the intended mode. In effect, LOVE-R1 is trained both as a reasoning agent for general questions and as a grounded clip-localizer for zoom-in decisions.

The paper reports Stage 3 settings of learning rate $1e^{-6}$, batch size 32, and 8 rollouts per sample.

5. Empirical performance and ablation structure

The main evaluation is conducted on four long-video benchmarks: MLVU, VideoMME, LongVideoBench, and LVBench. LOVE-R1 is reported to achieve 67.4 on MLVU, 66.2 on VideoMME, 60.1 on LongVideoBench, and 48.2 on LVBench (Fu et al., 29 Sep 2025).

Compared with Qwen2.5-VL, the reported gains are +1.0 on MLVU, +1.0 on VideoMME, +3.1 on LongVideoBench, and +4.1 on LVBench. The paper also reports +6.2 on LVBench in the table’s comparison line. Its headline summary is that the average improvement over Qwen2.5-VL across the four benchmarks is 3.1 percentage points.

The comparison suite includes proprietary models such as GPT-4V and GPT-4o; open-source LVLMs such as Video-LLaVA, LLaMA-VID, ShareGPT4Video, LLaVA-NeXT-Video, VideoLLaMA2, LongVA, VideoChat2, LLaVA-OneVision, Vamba, VideoChat-T, Quicksviewer, Video-XL, Video-XL-Pro, LongVILA, LongVU, Hour-LLaVA, LongVILA-128k, ViLAMP, and VideoChat-Flash; agent-based models such as VideoMind and Video-RAG; reasoning video models such as Video-MTR, Video-R1, VITAL, and LongVILA-R1; and the main baseline Qwen2.5-VL 7B.

The stage-wise results are used to separate template adaptation, reasoning induction, and RL-based process optimization. Stage 1 is reported to strongly adapt the model to the slow-fast template; Stage 2 improves reasoning and especially long-video benchmarks; and Stage 3 gives the best overall performance after decoupled RL.

The ablations are central to the paper’s argument. On template choice, three slow-fast templates are compared: replacing fast segments with slow videos and treating them as one video, replacing fast segments with slow videos and separating them explicitly, and appending slow videos after the fast video. Template (c), the append-after-fast format, is reported as best after finetuning because it is closest to the pretraining format.

On RL strategy, the paper compares no single-step optimization, multi-step optimization only, and multi-step plus single-step optimization. The combined decoupled strategy performs best, which is used to support the claim that directly optimizing the zoom-in step provides better process supervision than relying only on final-answer rewards.

On clip selection, the paper compares no zoom-in videos, uniform zoom-in videos, random zoom-in videos, and adaptive zoom-in videos. Adaptive zoom-in is clearly best. This is the main empirical basis for the claim that the improvement is not from simply adding extra high-resolution frames, but from selecting informative clips based on reasoning.

The paper also contrasts two single-step baselines that fail to balance the long-video trade-off: 128 slow frames capture detail but lose long-range temporal information, whereas 768 fast frames capture temporal cues but lose detail. LOVE-R1’s adaptive mechanism is presented as combining the strengths of both.

6. Interpretation, limitations, and position within long-video LVLM design

LOVE-R1 turns long video understanding into a reasoning-and-search problem: inspect a compressed global video first, selectively zoom into relevant moments at higher resolution, and answer only after sufficient evidence is present. Within the paper’s own framing, the contribution is therefore joint over representation, control policy, and supervision regime rather than over sampling strategy alone.

Its main stated contributions are a slow-fast adaptive frame-processing scheme for long videos, a multi-step reasoning formulation with decision, zoom-in, and answer stages, a 38k high-quality CoT dataset for video reasoning and grounding, a decoupled reinforcement finetuning strategy that separately optimizes zoom-in behavior using process-level supervision, and empirical gains over Qwen2.5-VL and other baselines on major long-video benchmarks.

The paper also states several limitations. Context is limited to around 16k, so the model cannot exploit arbitrarily many frames or reasoning steps. Performance depends heavily on the quality of training data. The current system is constrained by available computation and could benefit from longer context windows. Training and inference are both limited to a maximum of 3 reasoning steps, which may restrict scalability.

These limitations clarify the scope of the reported gains. The model achieves what the paper describes as a great trade-off between sampling density and frame resolutions under a moderate context budget, but it does not claim to remove the underlying long-context bottleneck. A plausible implication is that LOVE-R1 should be read less as a final solution to long-video understanding than as a specific synthesis of adaptive evidence acquisition and process-supervised reasoning in LVLMs. Within that scope, its principal significance lies in showing that explicit zoom-in control can be learned as part of the reasoning process rather than imposed as a fixed preprocessing policy.