acceptodds
Under review as a conference paper at ICLR 2027

DRIVE: Dual Reflection with Interleaved Visual Evidence for Embodied Pointing

Abstract

A model may identify the intended visual target yet predict a coordinate that falls elsewhere. We present DRIVE (Dual Reflection with Interleaved Visual Evidence), a two-stage reinforcement-learning framework designed to strengthen the connection between spatial reasoning and coordinate prediction. A shared policy first proposes target points, then reflects on their placement through rendered feedback: it states a spatial expectation before rendering and diagnoses the observed marker afterward. When the diagnosis identifies a mismatch, the policy revises the coordinates and inspects them again. Training uses final localization outcomes to optimize a selected response conditioned on this reflection history. On the 300-example VABench-P test set, the complete protocol achieves 71.67% final-hit accuracy, compared with 67.33% for direct inference with the same trained checkpoint. Inference ablations examine how spatial expectations, placement diagnoses, and rendered markers contribute to the trained policy’s predictions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
vision-language models, visual grounding, embodied manipulation, iterative refinement, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_x3aLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present DRIVE, which is an approach to improve how accurately VLMs can locate target objects in images. In order to do this, the model will first predict a point; then describe what the model expects to see at that location; and finally check a visual marker prior to making an adjustment to its prediction. The model is trained via RL using the final pointing results. Using 300 VABench-P examples, DRIVE increased accuracy from 67.33% to 71.67% when compared to directly predicting with the same trained model.

Critical strengths
  1. The idea of the method is straightforward.
  2. The problem is clearly stated (discrepancy between good object identification and bad actual pointing coordinates)
Critical weaknesses
  1. No clear baseline comparison with prior work (thos VLMs finetuned for pointing, or frontier VLMs).
  2. Testing on only a single benchmark, very limited evaluation.
  3. Poor writtings very AI-alike.
  4. The methodology has no novelty.
Questions for authors

See weakness.

Reviewer_uzqiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DRIVE, a framework that combines self-reflection and reinforcement learning to improve VLMs’ performance on coordinate prediction tasks. The authors evaluate their method on a benchmark of 300 samples and show a modest improvement.

Critical strengths

None

Critical weaknesses
  1. The paper has limited novelty. The work can be summarized as combining self-reflection and GRPO with small modifications. Both ideas/methods are well-established in the literature. Also, this approach is similar to “Poivre: Self-Refining Visual Pointing with Reinforcement Learning.” Although the authors describe differences between the methods, they should more clearly establish why these differences constitute a meaningful methodological contribution.

  2. The paper’s evaluations are lacking. There are several design choices throughout the method, such as assigning zero advantage to tokens in earlier response turns and modifying the loss compared with the original GRPO. These choices should be investigated through ablations to establish their importance. The current inference-time ablations do not provide this evidence. Also, the paper evaluates only on one dataset and against one external baseline, limiting the evidence for generalizability and improvement over existing methods.

Questions for authors
  1. Why is the loss function different from the standard and established GRPO setting? What are the benefits of the formulation used in the paper?
Reviewer_CZ7KOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors present DRIVE, a two-stage reinforcement-learning framework that connects spatial reasoning with coordinate prediction. Stage 1 generates initial point proposals, while Stage 2 overlays candidate points on the original image and decide whether to retain or revise them. However, the reported gains cannot be clearly attributed to visual-feedback-driven coordinate correction.

Critical strengths

The paper tackles a well-motivated and practically relevant problem: understanding which visual target an instruction refers to does not necessarily translate into accurate coordinate prediction. Addressing this gap is important for applications that rely on precise spatial grounding, such as robotic manipulation.

Critical weaknesses

The paper provides evidence that the complete interaction protocol improves final-hit accuracy (71.67% vs. 67.33%; Table 1), but the contribution of the dual-reflection training design remains insufficiently isolated. All inference conditions use the same trained checkpoint (§4.1), and the paper acknowledges the need for matched training ablations (§5). Moreover, final points match the first rendered points in 297 of 300 examples (§4.4), while the visual-marker improvement has a confidence interval that includes zero (Table 1). Since terminal validity also increases from 95.33% to 100%, the results do not disentangle improved spatial correction from improved output validity and other effects of the interaction protocol.

Questions for authors
  1. Table 2 reports ten M→H cases for B0, whereas none of the three two-call trajectories in Appendix E.4 has this transition. For those ten cases, could you report the Stage 1 proposal, first rendered candidate, and final answer, together with their validity and hit status? How many improvements occur before the first rendering, and how many recover a missing or invalid proposal?

  2. Have you compared DRIVE against direct-pointing RL and a point–render–revise RL baseline without the expectation/diagnosis fields, holding initialization, training data, and training budget fixed? Such comparisons would help distinguish the contribution of dual-reflection training from the inference-protocol effects measured in A1–A5.

  3. B0 and A1 have different initial-hit rates (68.67% and 67.33%). Could you evaluate both from identical first-tool candidate points, with only the visual marker differing, and report the fraction of initially incorrect candidates corrected and initially correct candidates corrupted? This would more directly isolate the effect of rendered feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.