acceptodds
Withdrawn from ICLR 2027

T-Spatial: Learning Utility-Aware Tool Use for Spatial Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

External tools are increasingly used to improve the performance of vision-language models (VLMs) on spatial reasoning tasks. However, executing a tool call successfully does not guarantee better performance. Inappropriate tool calls may even harm reasoning. Existing credit-assignment methods typically assign a single final-answer reward to an entire multi-turn trajectory, providing limited guidance on whether an individual call is beneficial, harmful, or redundant. This coarse-grained credit assignment makes it difficult to further improve the utility of tool use in long, multi-turn interactions. To address this challenge, we propose the Tool-Augmented Spatial Reasoning Agent (T-Spatial), trained with Difficulty-Aware Trajectory Synthesis (DATS) for supervised fine-tuning (SFT) and Tool Utility-based Policy Optimization (TUPO) for reinforcement learning (RL). DATS evaluates the target VLM on each training question and assigns a difficulty label based on its performance. According to these labels, DATS constructs a tool-assisted or a direct-answer trajectory for each question and teaches the target model to adaptively use spatial tools according to the difficulty of the question. TUPO combines sample-, trajectory-, branch-, and node-level tool-utility signals that are mainly computed from differences in answer scores across branches or before and after tool calls, allowing TUPO to reuse the existing answer evaluator without an auxiliary reward model. These complementary signals reduce the reward sparsity and make tool utility explicit at multiple granularity. Experiments on multiple spatial reasoning benchmarks show that T-Spatial improves spatial reasoning accuracy and tool-use utility.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Spatial Reasoning, Tool Use Agent, Group Relative Policy Optimization, Vision-Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gxw4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces T-Spatial, an agent designed for spatial reasoning that addresses the issues of potentially ineffective tool usage and coarse-grained credit assignment in multi-turn interactions. The proposed method comprises two components: Difficulty-Aware Trajectory Synthesis (DATS), which constructs either direct-answer or tool-assisted trajectories based on the base model's performance; and Tool Utility Policy Optimization (TUPO), which organizes training signals across four levels—samples, trajectories, branches, and nodes—and optimizes the policy by combining tree-based sampling with changes in answer scores observed before and after tool invocation. The authors demonstrate performance improvements across six spatial reasoning benchmarks and analyze the agent's behavior regarding tool utility and the substitution of the geometric backend.

Critical strengths
  1. The problem addresses a practical need, and the overall design is coherent. The workflow—spanning from training data construction to reinforcement learning—centers on the necessity and effectiveness of tool use; the use of the answer evaluator for node-level feedback eliminates the need to train a separate reward model, making the implementation approach highly practical.
  2. Performance gains are significant and substantiated by ablation studies. The 8B model achieved an average score of 62.5 across six benchmarks—a 10.3-point improvement over the corresponding base model—and outperformed standard GRPO by 2.7 points when using the same SFT initialization. Removing either branch-level or node-level signals resulted in a marked performance drop, validating the importance of both components.
  3. The evaluation includes an analysis of tool-use behavior. Breakdowns of gains and losses associated with tool usage, alongside experiments involving the replacement of the geometry backend without retraining, provide further evidence of the effectiveness and generalizability of the tool-use approach.
Critical weaknesses
  1. The incremental contribution over existing methods has not been sufficiently validated. While the paper acknowledges related work on answer probes, tree-based credit assignment, and mixed-tool supervision, it lacks comparisons under identical settings with closely related methods such as TACO and Tree-GRPO. Current comparisons focus primarily on standard GRPO, making it difficult to assess the advantages of the joint design over existing mechanisms.
  2. The rationale behind DATS labeling is ambiguous, and independent ablation studies are lacking. The scenario where a "base model answers incorrectly without tools" while a "stronger teacher answers correctly with tools" simultaneously alters both model capability and tool availability; thus, it does not directly demonstrate that the target model requires tools. Furthermore, the paper fails to compare DATS against alternative construction schemes—such as standard teacher distillation or fixed-ratio trajectory mixing—leaving the specific contribution of DATS itself unclear.
  3. The reliability of node-level tool utility estimation requires further verification. The change in answer scores observed in Equation (7) is influenced by both intermediate reasoning steps and tool outputs; this approach risks assigning zero reward to necessary tool calls that do not immediately improve the answer. While branching signals might mitigate this issue, there is a lack of controlled experiments to verify the consistency between local rewards and actual tool contributions.
  4. Evidence regarding computational overhead and efficiency is insufficient. An identical number of rollouts does not equate to equal total computational costs when answer probes are included; the absence of cost comparisons with GRPO—along with missing metrics such as average call frequency and inference latency—means the claim regarding the reduction of redundant tool calls is not adequately supported.
Questions for authors

See Weaknesses.

Reviewer_PvurOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

T-Spatial is a tool-augmented VLM agent for spatial reasoning that calls depth, segmentation, detection and orientation tools over multiple turns. DATS labels each training question by whether the target model solves it without tools and whether a stronger synthesizer (Qwen3.5-Plus) solves it with tools, and builds direct or tool-assisted SFT trajectories accordingly . TUPO then runs tree-structured GRPO that combines sample-level rollout budgets, a cost-adjusted outcome reward, branch-level sums of descendant advantages, and node-level answer probes taken before and after each tool call. T-Spatial-8B averages 62.5 on six spatial benchmarks, +10.3 over its base.

Critical strengths
  1. A well-posed problem: a successful tool call is not necessarily a useful one, and outcome-only rewards conflate helpful and harmful calls

  2. The paired with/without-tool evaluation in Table 2 is a useful protocol, and T-Spatial shows the largest net tool utility among the compared agents (11.7 vs 6.7 for SpatialAgent-8B) .

Critical weaknesses

1, No comparison with the closest methods. The paper names TACO (differential answer probes), Tree-GRPO and PORTool (tree-based credit assignment) as the most closely related work and concedes that its ingredients are individually not new , yet the only RL baseline trained from the same SFT checkpoint is standard GRPO. Since TUPO's novelty lies in the combination, it must be compared with these methods under the same initialization. IGPO (ICLR 2026), which rewards each turn by the gain in ground-truth answer probability, is also not discussed.

2, The necessity of tool calls is not established. With tools disabled, the trained model is weaker than its untrained base (Acc_wo 50.8 vs 52.2 for Qwen3-VL-8B-Thinking; 49.3 vs 51.2 at 4B), object detection is invoked on 95.5% and 97.8% of the questions in two task categories, and 16.4% of the answers flip merely because the tool schema is present, without any call. The paper reports neither calls per question nor accuracy with corrupted tool outputs, so it remains unclear whether the agent relies on the returned evidence or uses tools by default.

3, Most of the gain comes from SFT distillation. The 8B average goes from 52.2 (base) to 57.6 (SFT), 59.8 (standard GRPO) and 62.5 (TUPO), and the SFT trajectories are synthesized by Qwen3.5-Plus. The training pool is drawn from VSI-590K and SpatialLadder-26K, while VSI-Bench is an evaluation benchmark and SpatialLadder-3B a baseline, and decontamination is not discussed.

4, Reproducibility. There is no appendix, and key RL settings (clipping, KL, turn and call budgets) as well as the cost of per-turn probe decoding are not reported

Questions for authors

How does TUPO compare with TACO, Tree-GRPO and PORTool when all are trained from the same DATS SFT checkpoint with the same tools and budget?

What is the accuracy when tool outputs are shuffled or corrupted, and why is the trained model worse than its base when tools are disabled?

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

The authors are tackling the problem of teaching VLMs how to properly use spatial tools to solve spatial reasoning tasks. One challenge with standard RL is that it only provides a reward for the final answer, which makes it difficult to assign credit to useful tool calls in long, multi-turn trajectories. To address this, the authors propose changes to both SFT and RL. For SFT, they label each training question by whether the base model can answer it directly. They then train on tool-free trajectories for questions the model already solves and on teacher-generated tool trajectories for the rest, so the model can learn when it is appropriate to use tools. For RL they extend GRPO in two ways. First, they add some reward shaping based on question difficulty and tool use. They add a small per-tool cost to the correctness score, and penalize tool-use on questions the base model can already answer. Second, they introduce finer-grained credit assignment using tree rollouts, where new branches are sampled from randomly chosen intermediate steps of a trajectory. The advantage of each step is the sum of the advantages of all trajectories that go through it, so steps shared by several branches are credited by all of their outcomes. In addition, before and after each tool call, the model is prompted to answer immediately, and the change in that answer’s correctness is used as a reward for the call. They train Qwen3-VL-4B and 8B with this pipeline and show substantial gains over the base models across six spatial reasoning benchmarks.

Critical strengths

They address a well known problem in the tool-augmented VLM space (outcome rewards can't distinguish helpful vs. harmful tool calls) by integrating recent ideas into a single training pipeline for spatial reasoning. They demonstrate consistent improvements over the Qwen3-VL base at two scales (+8.5 and +10.3 points on average). The paired with/without-tools analysis supports that the gains come from the tool calls themselves.

Critical weaknesses

I’m not entirely convinced that the main claims of the paper are fully supported by the experiments. While T-Spatial clearly improves over its base model, the paper does not separate the effect of tool access itself from the effect of training, since the untrained base model is never evaluated with the same tools (please correct me if I am wrong here). Tool utility is compared against other tool-use agents, but this comparison may be confounded by differing toolsets. For example, if SpatialAgent-8B was run with its original toolbox, it uses different and generally older tools (e.g., DepthAnything V2, SAM2, OWLv2), which might explain part of the difference. A cleaner comparison would be to use the same toolset. The intro claims that DATS reduces unnecessary tool calls and TUPO improves tool utility, but I’m not sure either is fully supported. While Figure 3 shows tool-use statistics for the final model, there is no baseline to compare against that shows tool use is reduced. Similarly, we would need to see the gain/harm decomposition for the SFT checkpoint and Standard GRPO to validate the claim that tool utility is improved.

Questions for authors
  • Were the base models in Table 1 evaluated without tools? Could you report results for the untrained Qwen3-VL base with T-Spatial’s four tools, to separate the effect of tool access from the effect of training? This is important since the main contribution is the training pipeline.
  • For the comparison in Table 2, did you use the original toolset (and backbones) for the other agents? I think the toolsets (and backbones) might be a confound for the gain/harm analysis.
  • Can you please report gain/harm analysis for the SFT checkpoint and standard GRPO? This is important for the utility claim.
  • Can you please report average tool calls per question, and the fraction of questions answered without tools, for the base (with tool access), SFT, Standard GRPO, and final models?
  • How sensitive are the ablation results to random seeds? Several differences are within 1–2 points.
  • Eq. 6 sums advantages rather than averaging them, so nodes shared by more trajectories receive larger advantages. Was this intentional, and how does it compare with mean aggregation?
  • How were the weights in Eq. 8 chosen and how sensitive is the method to those weights?

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

Related papers

Discussion (0)

Sign in to comment.