- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
NAVIGATE is a three-stage framework for training multimodal models to answer visual questions through external tool use. It combines a tool-aware curriculum, supervised fine-tuning with an auxiliary head that predicts future tool actions, and reinforcement learning using final-answer correctness, format compliance, planner–execution consistency, MMD-based embedding alignment, and covariance-stability rewards. The paper reports improvements across six visual perception and reasoning benchmarks, multiple model backbones, and several component ablations.
Critical strengths
- The paper addresses a useful problem: learning when and how to compose external tools for multimodal reasoning. Combining curriculum learning, auxiliary action prediction, and interactive reinforcement learning provides a plausible training approach to this problem.
- The evaluation includes multiple backbones and a matched SFT-to-GRPO baseline. In Table 8, NAVIGATE improves the overall score from 48.9 with matched vanilla GRPO to 51.3, providing evidence that the proposed components collectively offer benefits beyond simply adding reinforcement learning.
Critical weaknesses
- The central reward design lacks sufficient justification, and the presentation makes the method difficult to assess. The paper does not adequately explain why MMD-based alignment between input and output embeddings should improve visual grounding, or why similar output covariance across successive steps should improve tool reasoning. Different stages of a valid tool chain may naturally require different representations, and these rewards do not directly measure correctness. The writing and organization further obscure the method: notation is introduced without sufficient explanation, the relationship between the planning head and actual tool execution remains unclear, and the overview figure does not illustrate how the proposed rewards are computed or influence training. The qualitative examples show successful trajectories but provide limited insight into these mechanisms. A clearer formulation and a worked example connecting planning, execution, and reward computation would substantially improve the presentation. These shortcomings affect understanding and reproducibility of the main contribution, beyond stylistic preferences.
- The experimental evidence does not sufficiently substantiate the claims about long-horizon, flexible tool orchestration. The six benchmarks are useful for measuring final-answer quality, but improvements on them do not by themselves establish better adaptive planning. Table 11 reports only 2.86 executed tool calls per example for the default model, while extending the prediction horizon from 5 to 50 improves the overall score by just 0.4 points, without reported uncertainty estimates. The held-out-tool evaluation is relevant, but its training exclusions and evaluation protocol are insufficiently specified. There is also an unexplained reporting ambiguity: Table 8’s “without Stage III” configuration and Table 10’s “format + accuracy only” RL configuration have identical scores despite apparently different training procedures. Together, these issues limit confidence in the attribution and scope of the reported gains.
Questions for authors
- What motivates the MMD and covariance-stability rewards specifically for tool-assisted reasoning? How are their embeddings computed, and what evidence shows that these rewards improve grounding or useful tool execution rather than merely encouraging representational similarity?
- How is the embedding planner decoded into discrete tool predictions, and does it directly influence execution at inference? Please also specify how terminal accuracy and intermediate rewards are aggregated and converted into GRPO advantages.
- Why do the “without Stage III” results in Table 8 exactly match the “format + accuracy only” results in Table 10? What training procedures and retained components distinguish these configurations?
- In the held-out-tool evaluation, is an entire tool family excluded from training or only a particular API? What are the evaluation sample sizes, and how are unseen tools incorporated into the planner’s action space?