acceptodds
Under review as a conference paper at ICLR 2027

Route2Learn: Native Tool Routing with a Single MLLM for Long Video Understanding

Abstract

Long-form video understanding requires effective evidence acquisition strategies, as relevant information is often sparsely distributed across extended temporal spans. Different queries can benefit from different strategies: generation-based strategy excels at reasoning over global context, while retrieval-based strategy is better suited for locating specific events. However, existing methods either rely on a single strategy or decouple evidence acquisition from policy optimization via external modules, limiting their adaptability. In this paper, we propose **Route2Learn**, an end-to-end native agentic framework that learns query-adaptive routing between generation-based and retrieval-based strategies within a single video MLLM. Building on the insight that learning to route requires comparing both routes, we introduce a two-stage optimization approach: For Supervised Fine-Tuning, we actively execute both strategies for each query and progressively collect trajectories with rejection sampling, providing an implicit routing signal by exposing the model to which strategies succeed on different queries. We then introduce Counterfactual Routing Policy Optimization (CRPO), which explicitly optimizes routing preferences using the outcome difference between factual and counterfactual rollouts, while decoupling advantage estimation for tool selection and answer generation to enable precise credit assignment. Extensive experiments demonstrate that Route2Learn substantially outperforms existing strong baselines across four long-form video understanding benchmarks, improving over Qwen3-VL-8B by +10.4% on average. Further analysis shows that the model learns to select the more effective evidence-acquisition strategy according to the query. Our data, code, and model weights will be released publicly.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Models, Long-form Video Understanding, Agentic Video Understanding, Tool Use, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_R2yuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Route2Learn, a single video MLLM that adaptively selects between temporal cropping and semantic frame retrieval for long-video question answering. Balanced supervised trajectories expose the model to both evidence-acquisition strategies; Counterfactual Routing Policy Optimization (CRPO) then compares factual and forced alternative-route rollouts, assigns a routing advantage to tool-turn tokens, and reserves the normal outcome advantage for answer tokens. Experiments on LVBench, VideoMME, MLVU, and LongVideoBench report an average accuracy of 63.4%, compared with 53.0% for Qwen3-VL-8B and 58.7% for the strongest listed open baseline. Ablations compare adaptive, fixed, and random routing and examine SFT, GRPO, CRPO, and shared versus decoupled advantages.

Critical strengths

The central formulation is consequential and well motivated: assigning both complementary evidence-acquisition tools to the same trainable MLLM removes the usual separation between retrieval behavior and policy optimization. This is a meaningful departure from both crop-only native agents and agent systems with frozen external retrievers.

The empirical evidence, especially Table 2, provides a substantive initial case for the method. Adaptive routing improves the four-benchmark average from 60.7% for the best fixed strategy and 59.5% for random routing to 63.4%, while CRPO with eight factual rollouts exceeds GRPO with sixteen. These comparisons indicate that the gains are not plausibly explained only by adding a reasoning or tool-use stage.

Critical weaknesses

The paper does not yet cleanly isolate the routing effect that CRPO claims to optimize. A factual trajectory may use multiple tools, but the “alternative route” is described only by forcing the alternative first tool and regenerating the rest. The resulting accuracy difference can therefore reflect later differences in tool sequences, prompts, and sampling rather than a controlled intervention on one routing decision. Moreover, the routing advantage is applied to every non-final thought token, so it remains unclear which tokens actually encode the intended choice. The paper should define what constitutes a route, enforce pure-route or explicitly paired rollouts, and report route-choice and factual-versus-counterfactual success statistics.

The headline comparisons also need stronger controls. Table 1 does not make clear whether all open baselines were rerun with the same frame budget, number of tool calls, decoding, answer parser, and evaluation prompt. Route2Learn may use up to three tool-calling turns and three crop calls per turn, so unmatched visual evidence budgets could explain part of the large gap. In addition, the main results appear to come from a single training run, and the paper does not audit overlap between the four evaluation benchmarks and the auxiliary Video-R1, LongVideo-Reason, LongVT, or CG-Bench training sources. These issues are fixable, but until addressed they materially limit confidence in the reported magnitude and causal interpretation of the gains.

Questions for authors
  1. Precisely how is a “route” defined when the factual rollout changes tools across turns? Does the counterfactual pair isolate one tool decision while holding the preceding trajectory fixed, and exactly which tokens receive the routing advantage?

  2. Were all baselines in Table 1 rerun under the same visual-frame/token budget, maximum tool-call budget, prompt, decoding procedure, and answer-extraction protocol? Please report average frames, tool calls, and visual tokens for Route2Learn and the strongest baselines.

  3. Can you report quantitative routing diagnostics, such as tool-selection distributions by query type, factual and alternative-route success rates, and an oracle-route upper bound? These results would directly test whether the model learned the claimed query-adaptive behavior rather than relying only on case studies.

Reviewer_EL9iOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses evidence acquisition for long-form video understanding, where different queries favor different strategies: generation-based temporal cropping or retrieval-based frame search. The authors propose Route2Learn, a single Qwen3-VL-8B agent with two native tools for long-video QA: crop_video (generation-based temporal inspection) and retrieve (MLLM-embedding frame retrieval). Its premise is that learning to route requires comparing both routes. Stage 1 does SFT on 11K balanced teacher-generated trajectories (5.5K per route) plus 55.2K non-tool samples. Stage 2, CRPO, builds a counterfactual rollout forced to use the alternative tool, derives a tool reward from the accuracy gap, and applies it to "tool tokens" while answer tokens receive only the outcome advantage. Across four benchmarks, it reports 63.4% average accuracy, +10.4 over the backbone and +4.7 over ParaVT-8B, with ablations supporting adaptive routing and decoupled advantages.

Critical strengths
  • the paper addresses a well-posed problem. The distinction between "whether to use a tool" and "which of two complementary tools fits this query" is clear and practically important. The limitations of the two existing lines (single-strategy native methods; training-free routing with external modules) are accurately summarized. The improvement of upgrading training-free routing into a trainable routing policy inside a single MLLM is natural and straight-forward

  • state-of-the-art results among open models at the 8B scale across four long-video benchmarks; ablations cover routing strategies (crop-only / retrieve-only / random / adaptive), training stages (SFT / GRPO / CRPO), λ sensitivity, query-type breakdown, efficiency, and both successful and failure cases

Critical weaknesses
  • the counterfactual in CRPO is a prompt-level intervention, not a minimal action-level one: App. F forces the alternative tool via an appended system-prompt constraint ("Do NOT call {other} under any circumstances"), anchored to the first tool call of the factual rollout but binding over the entire trajectory. Since factual rollouts may interleave both tools, the counterfactual is a pure single-tool trajectory regenerated from scratch (Sec. 3.3). Consequently, R_tool confounds (i) the first routing decision, (ii) the remainder of the trajectory, (iii) the model's current competence at the forced tool (a non-stationary quantity), and (iv) instruction-following under the forced-tool prompt

  • the training signal thus depends on instruction-following, the very fragility the paper criticizes in training-free routing (albeit here only at signal-construction time, not inference)

  • the P_valid filtering removes non-compliant rollouts but introduces an unquantified selection bias (please report invalid-pair and tie rates). Moreover, because the two rollouts share no prefix beyond the initial input, all trajectory-level variance enters the reward — a single counterfactual sample yields a high-variance {-1, 0, 1} signal

  • as written (App. A.1), crop_video returns 64 frames, with up to three calls per turn and three turns (up to 576 frames), while retrieve appears to be one call per turn (up to 192). Adaptive routing can combine both. The 63.4 vs. 60.0/60.7 gap in Table 2(a) may therefore reflect a larger frame budget, not better routing

Questions for authors
  • Can a factual trajectory call both tools? What fraction are pure-crop, pure-retrieve, and mixed, and how is "the alternative route" defined for mixed trajectories? Why force the other tool for the whole trajectory rather than intervening only on the first action?

  • What fraction of counterfactuals are invalid/non-compliant and dropped, and how does that change R_tool? Did you try scoring non-compliance as incorrect, and what is the non-tie fraction?

  • Since Tau_tool includes all thought tokens, did you ablate applying A_tool only to the action/tool-name tokens or only to the first routing thought?

  • What are the average frames and calls per query for crop-only, retrieve-only, random, and adaptive? Does the gain survive equal frame budgets, and are the fixed-strategy rows the same CRPO model?

Reviewer_TP6bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents Route2Learn, a tool-routing framework designed to iteratively gather visual evidence for long-video understanding. It first constructs a counterfactual rollout and then employs CRPO, a variant of GRPO, by designing a new reward function that decouples token advantage. Experimental results demonstrate its superior performance compared with other video MLLMs.

Critical strengths

This manuscript proposes Route2Learn, a native agentic framework for long-video understanding. It unifies generation-based and retrieval-based strategies by learning routing policies via CRPO. Experimental results demonstrate substantial improvements over Qwen3-VL-8B.

Critical weaknesses
  • The baseline "retrieve-only" strategy can cover "crop-only" strategy, since the "retrieve-only" can accurately retrieve related visual evidence which is same as "crop-only" strategy with time-range crop. The detailed experiment results of "retrieve-only" strategy are not clear.
Questions for authors

LongVideoBench primarily evaluates perception and understanding over temporal intervals. However, the crop-only strategy outperforms the retrieve-only strategy by 2.9 points. Could you provide a more detailed analysis of this performance gap? Specifically, which data subsets or question types account for most of the improvement? Why does the retrieve-only strategy perform worse than the crop-only strategy? Could this indicate that the retrieval component has not been sufficiently trained?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.