ChronoForge-RL: Temporal Keyframe Selection

• The paper introduces ChronoForge-RL, combining Temporal Apex Distillation (TAD) with keyframe-aware RL to optimize long-form video temporal reasoning.
• It employs local maximum detection and contrastive policy optimization to select semantically significant frames while reducing redundant computations.
• Empirical results on LVBench and VideoMME benchmarks demonstrate significant performance gains and a 10× improvement in the performance-to-parameter ratio.

Searching arXiv for the specified paper and closely related work to ground the article. ChronoForge-RL is a video understanding framework for multimodal LLMs that targets two practical limitations of long-form video reasoning: “the computational infeasibility of processing every frame in dense video content” and “the difficulty in identifying semantically significant frames through naive uniform sampling strategies.” It combines Temporal Apex Distillation (TAD), a keyframe-selection procedure centered on “semantic inflection points,” with KeyFrame-aware Group Relative Policy Optimization (KF-GRPO), a GRPO-style reinforcement-learning stage that explicitly contrasts chronologically ordered keyframes against temporally corrupted inputs (Chen, 19 Sep 2025).

1. Definition and problem setting

ChronoForge-RL is proposed for long-form video understanding with multimodal LLMs (MLLMs). Its stated objective is to preserve “temporally critical information” without incurring the cost of dense frame processing, while also avoiding the semantic weakness of uniform frame sampling (Chen, 19 Sep 2025). The paper’s problem formulation is grounded in the claim that information in video is not uniformly distributed over time: many neighboring frames are redundant, whereas a minority encode “semantic inflection points” that determine downstream reasoning.

The framework is motivated by a specific gap in prior systems. Training-free keyframe selection or token compression methods reduce cost, but “often lose fine-grained temporal structure or become less effective as more frames are retained.” Model fine-tuning and adaptive compression methods improve temporal handling, but “still do not fully integrate keyframe selection into the reasoning and learning objective.” ChronoForge-RL addresses this by coupling keyframe selection and RL-based temporal reasoning inside one pipeline rather than treating frame pruning as a preprocessing convenience (Chen, 19 Sep 2025).

The paper’s explanation for the method’s effectiveness is that videos are “sparse in semantic change,” that “temporal apexes are more informative than uniformly sampled frames,” and that “correct temporal order is itself a supervisory signal.” This positions ChronoForge-RL as a system for event-centric temporal abstraction rather than dense temporal coverage. The approach is especially aimed at long videos, including “hour-scale videos such as those in LVBench,” where dense processing is especially expensive (Chen, 19 Sep 2025).

2. System architecture and workflow

ChronoForge-RL consists of two major modules: Temporal Apex Distillation (TAD) for keyframe selection and KeyFrame-aware Group Relative Policy Optimization (KF-GRPO) for temporally aware RL training (Chen, 19 Sep 2025). The paper describes the following high-level pipeline:

1. Raw video input
2. Visual feature extraction / frame representation
3. TAD keyframe selection
4. Offline data processing
5. KF-GRPO training
6. Inference

At the architectural level, TAD first compresses the video by selecting salient frames. These selected keyframes are then used to construct two kinds of training inputs: a chronologically ordered keyframe sequence and a disordered hybrid negative sequence. KF-GRPO generates responses for both inputs, computes rewards with a saliency-aware temporal component, and updates the policy with a clipped GRPO-like objective plus KL regularization (Chen, 19 Sep 2025).

The design logic is explicit. Frame selection reduces redundancy; temporal reasoning is induced by preserving chronology in the positive sequence and disrupting it in the negative sequence; contrastive RL training rewards the model when the ordered keyframe input yields better accuracy than the corrupted input; and inference efficiency follows because only selected frames are passed to the final model. This suggests that ChronoForge-RL does not treat temporal order as a passive property of the data, but as an object of optimization (Chen, 19 Sep 2025).

The paper also notes that Figure 1 describes a “multi-level soft reward mechanism” integrating “keyframe ratios and spatiotemporal sequences,” but the mathematical specification of that mechanism is limited. This matters because the method description is stronger at the level of system design than at the level of complete reward formalization.

3. Temporal Apex Distillation

TAD is the framework’s keyframe selector. It takes a feature tensor

$F \in \mathbb{R}^{T \times N \times C},$

where $T$ is the number of frames, $N$ is the number of spatial tokens or patches per frame, and $C$ is the feature dimension. Its output is a distilled sequence

$F' \in \mathbb{R}^{K \times N \times C},$

where $K$ is the frame budget (Chen, 19 Sep 2025).

The procedure has three stages: Variation Scoring, Inflection Detection, and Prioritized Distillation. In Variation Scoring, adjacent-frame dissimilarity is computed as

$$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$

The algorithm box then aggregates across the spatial dimension with

$V \leftarrow \mathrm{Mean}_{\mathrm{spatial}}(V).$

However, the prose description says that in synchronous mode the dissimilarities are aggregated using max pooling, not mean pooling. The paper is therefore internally inconsistent on this point, which is a direct reproducibility issue (Chen, 19 Sep 2025).

Inflection Detection searches for local maxima in the variation sequence: $P \leftarrow \mathrm{FindLocalMaxima}(V, W),$ where $W$ is the sliding window size. The text states that this is implemented with “1D max-pooling with indices,” and conceptually a frame at index $T$0 is selected if it is the maximum within its local temporal neighborhood. The output $T$1 is the set of inflection-point indices, interpreted as peaks in the rate of visual change (Chen, 19 Sep 2025).

Prioritized Distillation then boosts those peaks: $T$2 so that

$T$3

The final keyframes are selected by

$T$4

and then sorted in ascending temporal order before gathering: $T$5 This ranking mechanism explicitly favors frames with large adjacent-frame change and especially those that are local maxima of that change (Chen, 19 Sep 2025).

The paper repeatedly describes TAD as a differentiable keyframe selection mechanism, but it also leaves an important caveat. It states that the local-maxima operation is implemented in a differentiable/GPU-friendly way and that the feature gathering $T$6 is differentiable with respect to the input features $T$7. At the same time, the paper does not provide a relaxation for the discrete $T$8 operation. The most defensible reading is therefore that the feature path is differentiable conditional on the selected indices, while full end-to-end differentiable index selection is not rigorously specified (Chen, 19 Sep 2025).

The appendix studies two TAD hyperparameters. For the inflection boost, the reported results are: $T$9 gives LVBench 48.2 and VideoMME overall 67.0; $N$0 gives LVBench 52.7 and VideoMME overall 69.1; and $N$1 gives LVBench 47.8 and VideoMME overall 60.6. For the retained-frame fraction, defined by

$N$2

performance improves up to $N$3, degrades for $N$4, and “training collapses for $N$5.” The recommended value is therefore $N$6 (Chen, 19 Sep 2025).

4. KeyFrame-aware Group Relative Policy Optimization

KF-GRPO is the RL component. It extends the temporal contrastive training style of Video-R1 and T-GRPO by making the contrast explicitly keyframe-aware (Chen, 19 Sep 2025). The paper’s implied policy is

$N$7

where $N$8 is the query and $N$9 is the selected keyframe set. The action is a generated response or completion, and the optimization is group-relative rather than single-sample (Chen, 19 Sep 2025).

Training begins by applying TAD to a video sequence $C$0: $C$1 From this, KF-GRPO constructs two frame ensembles. The positive sequence is the extracted keyframes in original chronological order. The negative sequence $C$2 is a “hybrid disordered” sequence containing approximately 50% keyframes and approximately 50% non-keyframes, randomly shuffled. The paper notes this construction but does not fully specify the exact sampling procedure (Chen, 19 Sep 2025).

The temporal saliency reward is defined by comparing accuracy on the ordered and disordered inputs. Let $C$3 denote accuracy on the ordered keyframes and $C$4 accuracy on the disordered hybrid sequence. Then

$C$5

so the saliency signal is binary: $C$6 The aggregate reward for output $C$7 is

$C$8

The text says $C$9 measures “semantic accuracy relative to ground truth,” “reasoning-chain logical/structural coherence,” and “alignment with expert demonstrations or stronger reference models,” but it does not provide explicit formulas for $F' \in \mathbb{R}^{K \times N \times C},$0 or $F' \in \mathbb{R}^{K \times N \times C},$1. That omission is one of the paper’s central under-specifications (Chen, 19 Sep 2025).

The paper also introduces a normalized reward or advantage term: $F' \in \mathbb{R}^{K \times N \times C},$2 where $F' \in \mathbb{R}^{K \times N \times C},$3 is the mean reward in the ensemble and $F' \in \mathbb{R}^{K \times N \times C},$4 is described in the text as a variance term, although the notation suggests a standard deviation. The relationship between $F' \in \mathbb{R}^{K \times N \times C},$5 and the $F' \in \mathbb{R}^{K \times N \times C},$6 used in the policy objective is not explicitly defined (Chen, 19 Sep 2025).

The optimization objective is presented as a clipped PPO/GRPO-style objective with KL regularization, although the typesetting is corrupted. The paper reconstructs it as

$F' \in \mathbb{R}^{K \times N \times C},$7

The paper’s own interpretation is that the “keyframe-aware” novelty lies less in the clipped policy form than in the use of TAD-selected inputs, ordered-versus-corrupted frame construction, and the saliency-aware reward (Chen, 19 Sep 2025).

A plausible implication is that KF-GRPO belongs to the broader family of GRPO-style or weighted-log-likelihood post-training methods discussed in work that recasts many RLHF algorithms as structured-bandit optimization rather than full action-dependent environment RL (Cai, 25 Mar 2025). In ChronoForge-RL itself, however, the operative distinction is that temporal order and keyframe saliency are built into the reward pathway rather than being external evaluation criteria.

5. Training recipe, benchmarks, and empirical results

ChronoForge-RL is built on Qwen2.5-VL-7B. The paper also uses Qwen2.5-VL-72B integrated with TAD for offline video frame selection during preprocessing. Supervised initialization uses Video-R1-CoT-165k, producing Qwen2.5-VL-7B-SFT after one epoch of SFT, and RL training then uses Video-R1-260k for about 1,500 steps. Training is reported on 8 NVIDIA A100 GPUs (Chen, 19 Sep 2025).

During training, each frame is processed at $F' \in \mathbb{R}^{K \times N \times C},$8 resolution, and a maximum of 32 frames are used. At inference, selected frames are processed at $F' \in \mathbb{R}^{K \times N \times C},$9 resolution. Evaluation is conducted through LMMs-Eval (Chen, 19 Sep 2025). The paper does not provide optimizer type, learning rate, batch size, clipping coefficient $K$0, KL coefficient $K$1, or group size $K$2, all of which materially limit exact reproduction.

The main benchmark results are as follows:

System	LVBench	VideoMME Overall
ChronoForge-RL (7B)	52.7	69.1
MiMo-VL (7B)	45.9	66.3
TW-GRPO (7B)	43.7	65.5
Video-R1 (7B)	44.7	64.6

ChronoForge-RL’s full reported breakdown is LVBench 52.7, VideoMME Long 60.0, VideoMME Medium 67.0, VideoMME Short 80.3, and VideoMME Overall 69.1 (Chen, 19 Sep 2025). Against RL-based 7B baselines, the paper reports gains of +6.8 on LVBench and +2.8 on VideoMME overall over MiMo-VL, +9.0 and +3.6 over TW-GRPO, and +8.0 and +4.5 over Video-R1.

The paper also emphasizes parameter efficiency. It compares ChronoForge-RL 7B against Qwen2.5-VL 72B with VideoMME overall 72.6 and LLaVA-Video 72B with 70.6. ChronoForge-RL’s 69.1 is therefore “only 1.5 points behind LLaVA-Video 72B” and “3.5 points behind Qwen2.5-VL 72B,” which the paper summarizes as a 10× improvement in performance-to-parameter ratio using the arithmetic $K$3 versus $K$4 (Chen, 19 Sep 2025).

Ablation evidence is central to the paper’s claims. Adding TAD improves several RL-based baselines but not all of them. For Video-R1, TAD changes LVBench from 44.7 to 49.7 and VideoMME overall from 64.6 to 65.3. For TW-GRPO, it changes LVBench from 43.7 to 50.3 and VideoMME overall from 65.5 to 67.6. For MiMo-VL, however, performance degrades from 45.9 to 43.7 on LVBench and from 66.3 to 60.8 on VideoMME overall. The paper attributes this to MiMo-VL’s assumption of “uniform 2 fps temporal continuity,” which TAD disrupts (Chen, 19 Sep 2025).

The efficiency claims are more limited than the accuracy claims. The paper states that for $K$5, the method retains “over 90% of peak performance with 40% computational cost reduction,” and that RL-based models are “3–4× slower in inference” than SFT baselines due to more complex reasoning. It does not provide exact FLOPs, memory reduction, or a detailed latency table (Chen, 19 Sep 2025).

6. Limitations, reproducibility, and research context

ChronoForge-RL’s main limitations are methodological specificity and incomplete specification. The paper itself identifies sensitivity to nonuniform temporal sampling, sensitivity to the retained-frame ratio $K$6, sensitivity to the inflection boost $K$7, and the fact that RL-based systems can be “3–4× slower in inference” than SFT systems (Chen, 19 Sep 2025). It also shows that TAD is not universally beneficial: the MiMo-VL degradation demonstrates that adaptive keyframe selection can conflict with models trained under fixed temporal-continuity assumptions.

Several under-specified points are more consequential for reproduction. The paper leaves unresolved the mean-pooling versus max-pooling discrepancy in TAD, does not formalize the exact local-maxima condition, does not provide an explicit differentiable relaxation for top-$K$8, omits the exact formulas for $K$9 and $$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$0, does not clearly define the relation between $$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$1 and $$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$2, and omits optimizer, learning-rate, batch-size, clipping, KL, and rollout-group details (Chen, 19 Sep 2025). These omissions make the method substantially easier to understand conceptually than to reproduce exactly.

In research context, ChronoForge-RL occupies a specific point in the design space of post-training methods. It is not a generic long-horizon environment RL system, but a keyframe-aware temporal contrastive RL framework for video understanding. A plausible implication is that, because its policy improvement is conditioned on fixed prompt-video inputs rather than action-dependent external state transitions, its KF-GRPO component is closer to the structured-bandit style of many LLM post-training methods than to classical environment-coupled actor-critic RL (Cai, 25 Mar 2025). What distinguishes ChronoForge-RL within that broader family is its tight coupling between saliency-based frame distillation and rewarded chronological reasoning.

Its central contribution is therefore architectural rather than purely algorithmic: frame selection and temporal reasoning are made mutually reinforcing. TAD identifies frames by adjacent-frame variation and local maxima,

$$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$3

while KF-GRPO rewards the model when ordered keyframes support better reasoning than shuffled hybrid frames,

$$V[t] \leftarrow 1 - \mathrm{CosineSimilarity}(F[t-1], F[t]).$$4

Within the scope defined by the paper, ChronoForge-RL is best understood as a system for training video MLLMs on temporally meaningful, semantically concentrated evidence rather than on uniformly sampled visual streams (Chen, 19 Sep 2025).