acceptodds
Under review as a conference paper at ICLR 2027

RaCE: Urban Spatial Multi-Hop Reasoning via Route-Aware Cumulative Evidence

Abstract

Urban spatial multi-hop reasoning combines visual observations and road geometry to infer complete routes and spatial answers from sequential instructions. Each location choice changes the spatial context and evidence available at subsequent hops. Similar buildings and road patterns can sustain local self-consistency along incorrect routes, making early errors difficult to detect through stepwise checks. Comparing these routes requires evaluating evidence from different spatial contexts on a common scale. To address these challenges, we propose Route-Aware Cumulative Evidence (RaCE). RaCE maintains feasible candidate routes and dynamically updates spatial contexts according to each branch's decision history. It integrates visual and geometric evidence into support scores comparable across branches and accumulates this evidence along complete routes, allowing later information to inform the reassessment of earlier decisions. The framework selects the best-supported route and final spatial answer through route-level comparison and answer-level evidence aggregation, respectively. Adaptive perception then leverages score bounds to acquire visual evidence selectively, yielding the same route and spatial answer as full acquisition. Experiments across five cities and four general-purpose models show that RaCE achieves mean route and answer accuracies of 77.7% and 92.1%, respectively, while reducing inference latency by 67.7% on average compared with existing methods. The data and code are available at https://anonymous.4open.science/r/RaCE-ECF0.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Urban Spatial Reasoning, Multi-Hop Reasoning, Urban Intelligence, Urban Map Navigation

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_WhY1OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper present RaCE, an inference framework for urban spatial multi-hop reasoning that predicts a complete route and its distance category from navigation instructions, visual observations, and a road graph. Its central motivation is that early location errors change subsequent spatial contexts, allowing incorrect routes to remain locally plausible. The main contributions are history-dependent spatial state modeling, comparable visual–geometric evidence accumulated over complete routes, and adaptive visual evidence acquisition. RaCE retains alternative routes, updates each branch’s spatial reference, and combines independently assessed visual matches with geometric penalties. It selects the highest-scoring route while separately aggregating route evidence for answer prediction. Adaptive acquisition uses bounds on both route and answer scores to avoid unnecessary visual queries while preserving exhaustive inference outputs under the same scoring and tie-breaking rules. The framework requires no additional backbone fine-tuning. Validation uses approximately 10,000 OSM-based instances across five cities, four general-purpose VLMs, and two spatially post-trained models. RaCE achieves 77.7% mean route accuracy and 92.1% mean answer accuracy across the four general-purpose backbones. On Qwen3.5, adaptive acquisition reduces average visual queries from 21.0 to 3.3. Ablations, ambiguity analyses, cross-city transfer, and scaling experiments provide further validation, with the strongest ablation evidence supporting geometry, spatial-reference updates, and cumulative scoring. Evaluation remains primarily limited to rendered map scenes.

Critical strengths

Strong evidence for the value of explicit, history-dependent spatial grounding. RaCE updates each branch’s spatial reference and separately evaluates visual semantic matches and geometric constraints, directly addressing the changing context of multi-hop spatial reasoning. In Table 2, removing geometry or fixing the spatial reference reduces Route ACC from 91.1% to 10.1% and 14.7%, respectively. The complete framework also achieves mean route and answer accuracies of 77.7% and 92.1% across four general purpose backbones. These results substantiate the usefulness of the spatial grounding mechanism and demonstrate benefits across multiple models, supporting the framework’s practical value within the evaluated setting.

Adaptive visual acquisition with a well-defined output-preservation guarantee. The method propagates bounds on unqueried evidence and separately certifies the best route and answer, correctly accounting for the distinction between route selection and answer aggregation. Under identical scoring and tie-breaking rules, the procedure preserves exhaustive inference outputs and terminates after finitely many queries. On Qwen3.5, it reduces average visual-query events from 21.0 to 3.3. This combination of a clear theoretical guarantee and substantial measured query savings is a substantive technical strength: it offers a principled way to reduce expensive perception while preserving the decisions of the full scoring framework.

Critical weaknesses

The distinctive benefit of preserving competing routes is not convincingly established. In Table 2, the w/o Branch variant already achieves 89.6% Route ACC, compared with 91.1% for full RaCE, whereas removing geometric evidence reduces accuracy to 10.1%. These results support the importance of geometric scoring and evidence accumulation, but leave unclear how much of the large improvement over existing baselines comes from preserving alternative histories and reassessing earlier decisions. Since Beam 4 uses a different scoring scheme, a comparison using the same local scores and state updates, together with a geometry-only control, is needed to isolate this contribution. This uncertainty substantially weakens the evidence for the claimed methodological advance.

Then, the bench construction and partitioning are insufficiently specified to assess the validity of the empirical conclusions.Although the paper reports dataset size and split ratios, it does not adequately describe instruction and reference-route generation, distractor selection, or geographic overlap across calibration, validation, and test sets. These details are consequential for OSM-derived instances that may share POIs, routes, or map regions, and for understanding how much ambiguity the candidate selection procedure introduces. The cross city evaluation protocol also needs a clearer account of sample selection and parameter fitting. Because the main empirical claims depend on this benchmark, these omissions limit reproducibility and prevent a confident assessment of the reported reasoning gains and generalization.

Questions for authors
  1. Contribution of branch preservation. Does the w/o Branch variant retain the same candidate construction, visual calibration, geometric scoring, and spatial-reference updates as full RaCE? Given its 89.6% Route ACC versus 91.1% for the full method in Table 2, how many test instances are corrected or made incorrect by preserving alternative branches, particularly within the high-ambiguity subset?

  2. Benchmark construction and candidate ambiguity. How are the navigation instructions, valid reference routes, and four candidate choices generated and verified? To what extent do reference-label generation and distractor selection use the same geometric rules as RaCE’s scorer, and how frequently do multiple candidates satisfy both the direction and distance constraints?

  3. Geographic independence and cross-city evaluation. What geographic overlap exists across the calibration, validation, and test sets at the map-region, POI, and route levels? For cross-city evaluation, which data are used to fit calibration parameters and select prompts and the geometric weight, and is the target city excluded from all these choices?

  4. End-to-end computational cost. Appendix C measures latency from the first visual query, whereas Algorithm 1 constructs the route space and geometric states beforehand. Do the reported latencies include this initialization? What is the full end-to-end inference time, and how do the reported logical visual-query counts correspond to actual model forward passes?

  5. Differences between the main and scaling experiments. What differences in evaluation subsets or protocols explain SpatialReasoner-R1’s 68.5% Route ACC in Table 1 versus 29.0% in Table 4 under the nominally same three-hop, four-candidate setting? Please clarify the sample sizes and any changes in dataset construction, sampling, prompts, or model configuration.

  6. Answer aggregation and output consistency. How do the authors explain the increase in Answer ACC from 95.7% to 96.8% when answer pooling is removed in Table 2? How often does the predicted answer differ from the distance category implied by the selected route, and what proportion of instances have both a correct route and a correct answer?

  7. Treatment of uncertain visual evidence. Why is the unknown probability excluded when visual evidence is computed as log⁡(qs/qc)\log(q_s/q_c)? Identical support-to-contradiction ratios produce identical visual scores even when the unknown probabilities differ substantially. How is this uncertainty accounted for, and how do high-unknown-probability instances perform in the reported evaluation?

Minor:The text in several figures, particularly Figures 1, 2, and 5, is difficult to read at normal viewing size. Please enlarge labels, legends, and equations, and simplify crowded annotations to improve readability.

I am open to revising my score if the authors can provide satisfactory clarifications that address the concerns raised above.

update

The authors have clarified the role of branch retention, the benchmark setup, computational costs, and the differences between the main and scaling experiments. They have also explained why answer pooling can reduce accuracy and assessed the impact of excluding unknown probabilities from visual scoring. Thanks for the responses and my main concerns have been addressed. I revised my score accrodingly.

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer WhY1 (Weakness 1)

Responce to W1:

We thank the reviewer for their attention to the multi-branch mechanism. We agree that the original ablation study demonstrates the contributions of individual components to overall performance but lacks a direct comparison with Beam Search under the same scoring mechanism. To address this, we have added two sets of controlled experiments in the Contribution of Branch Retention and Route-Level Comparison subsection on page 16 of the revised manuscript's appendix.

Specifically, Matched Greedy, Matched Beam-4, and RaCE use the same local visual–geometric scoring, spatial reference frame updates, and cross-hop evidence accumulation, but differ in their search strategies: step-wise greedy selection, fixed-width Beam Search, and complete-route comparison, respectively. Geometry-only retains RaCE's search strategy while fixing the visual evidence factor to 1, allowing us to isolate the contribution of geometric evidence.

\hline \mathbf{Method} & \mathbf{Search\\;Strategy} & \mathbf{Route\\;ACC\\;(\\%)} \\\\ \hline \text{Matched Greedy (w/o Branch)} & \text{Greedy (top-1)} & 89.6 \\\\ \text{Matched Beam-4} & \text{Beam search (width 4)} & 90.4 \\\\ \text{Full RaCE} & \text{Route-level comparison} & \mathbf{91.1} \\\\ \text{Geometry-only} & \text{Route-level comparison} & 49.8 \\\\ \hline \end{array}$$ The results show that under the same scoring and state update mechanisms, RaCE improves Route ACC by 1.5 and 0.7 percentage points over Matched Greedy and Matched Beam-4, respectively. Multi-branch retention prevents premature route selection by step-wise greedy decisions, while RaCE's additional advantage over fixed-width Beam Search lies in reducing the loss of candidates caused by intermediate pruning. When the correct route falls outside the top four in cumulative score at a given hop, Matched Beam-4 prunes it prematurely, making recovery impossible even if subsequent evidence supports that route. In contrast, RaCE retains more competing route hypotheses, allowing initially lower-scoring candidates to be reassessed using subsequent evidence. Combined with the ambiguity-group analysis in RQ3, where RaCE outperforms Greedy by 4.2 percentage points in the high-ambiguity group, these findings further indicate that the benefits of branch retention are primarily concentrated in scenarios where early candidates are difficult to distinguish.
AuthorsOpenReview ↗

Author Response to Reviewer WhY1 (Weakness 2)

Responce to W2:

We thank the reviewer again for pointing out the insufficient details regarding benchmark construction and data partitioning. To improve experimental reproducibility, we have added three subsections to page 17 of the revised manuscript's appendix: Instruction and Route Construction, Data Partitioning and Geographic Overlap, and Cross-City Evaluation and Parameter Calibration. The details are provided below.

(1) Instruction, Reference Route, and Distractor Candidate Construction.

The benchmark constructs three-hop route reasoning tasks based on the OSM road network. First, three target locations are sequentially selected to form a reference route based on road reachability, distance, and POI category constraints. Navigation instructions are then generated according to the target categories, travel directions, and approximate block-level distances at each hop. Given the limitations of POI categories and their spatial distributions in OSM, we require adjacent hops to have different target categories while allowing the first and third hops to share the same category. Next, four candidate POIs are assigned to each hop, with three distractors selected from nearby road-reachable locations, prioritizing those with similar categories, directions, or distances. Alternative routes that also satisfy the complete instructions are excluded.

To quantify candidate ambiguity, we counted the number of candidates satisfying both directional and distance constraints at each hop in the test set. The results are presented below.

\hline \mathbf{Decision\\;Hop} & \mathbf{1\\;Candidate} & \mathbf{2\\;Candidates} & \mathbf{3\\;Candidates} & \mathbf{Ambiguous\\;(\geq2)} \\\\ \hline \text{Hop-1} & 326 & 451 & 223 & 67.4\\% \\\\ \text{Hop-2} & 357 & 433 & 210 & 64.3\\% \\\\ \text{Hop-3} & 393 & 419 & 188 & 60.7\\% \\\\ \text{Total} & 1{,}076 & 1{,}303 & 621 & 64.1\\% \\\\ \hline \end{array}$$ (2) Data Partitioning and Geographic Overlap. The dataset consists of 6,000 training instances, 2,000 calibration instances, 1,000 validation instances, and 1,000 test instances. The splits are partitioned by task source to prevent duplicate questions and similar tasks generated from the same source from appearing across different splits. The calibration set is used for fitting scoring parameters. Since different navigation tasks within the same city may share the road network, we do not enforce strict geographic separation between splits, thereby preserving road network continuity and spatial coverage of candidate POIs. The overlap of roads, POIs, and reference routes across different splits is reported below. $$\begin{array}{lrrrr} \hline \mathbf{Comparison} & \mathbf{Shared\\;Nodes} & \mathbf{Shared\\;Edges} & \mathbf{Shared\\;POIs} & \mathbf{Shared\\;Reference\\;Routes} \\\\ \hline \text{Test / Train} & 24.60 & 21.40 & 16.70 & 0.90 \\\\ \text{Train / Test} & 11.43 & 2.70 & 5.85 & 0.18 \\\\ \text{Test / Validation} & 15.80 & 10.60 & 9.80 & 0.40 \\\\ \text{Validation / Test} & 13.10 & 6.40 & 8.70 & 0.50 \\\\ \text{Test / Calibration} & 8.10 & 3.90 & 4.30 & 0.20 \\\\ \hline \end{array}$$ (3) Cross-City Sample Selection and Parameter Calibration. Cross-city evaluation follows a Leave-One-City-Out protocol. In each round, all test instances from the target city are selected from the frozen Test1000 set, without additional random sampling, difficulty-based filtering, or selection based on prediction results. Model parameters are determined using data from the remaining four source cities only. Specifically, the visual calibration parameters $T$ and $b$ are fitted using L-BFGS-B by minimizing the candidate-level, equally weighted binary negative log-likelihood (NLL) on the source-city calibration sets. The geometric weight $\lambda_{\mathrm{geo}}$ is selected from ${1,2,4,8,16,32\}$ based on the equally weighted macro-average Route ACC across source cities on the validation set. The dataset sizes and parameter configurations for each city are presented below. $$\begin{array}{lrrrrrr} \hline \mathbf{Held\text{-}out\\;City} & \mathbf{Source\\;Calib.} & \mathbf{Source\\;Val.} & \mathbf{Target\\;Test} & \mathbf{T} & \mathbf{b} & \lambda_{\mathrm{geo}} \\\\ \hline \text{New York City} & 92 & 728 & 96 & 1.17 & 1.09 & 16 \\\\ \text{Berlin} & 111 & 936 & 148 & 1.14 & 1.09 & 16 \\\\ \text{Singapore} & 68 & 640 & 59 & 1.10 & 0.94 & 16 \\\\ \text{Sydney} & 106 & 936 & 48 & 1.13 & 1.10 & 16 \\\\ \text{Tokyo} & 83 & 760 & 649 & 1.16 & 1.17 & 16 \\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Author Response to Reviewer WhY1 (Question 1-2)

Responce to Q1:

We thank the reviewer again for their attention to this issue. We first clarify that the w/o Branch variant in Table 2 uses exactly the same candidate construction, visual calibration, visual and geometric scoring, spatial reference frame updates, and cross-hop evidence accumulation as the full RaCE model. The only difference is whether multiple candidate branches are retained. The w/o Branch variant retains only the highest-scoring branch at each hop, whereas RaCE maintains multiple route hypotheses and uses subsequent evidence to reassess them at the complete-route level. Therefore, this ablation study directly measures the additional contribution of the multi-branch retention mechanism.

To further analyze the impact of the multi-branch mechanism on route decisions, we grouped the full test set by candidate ambiguity level and conducted a case-level analysis to count the complete routes corrected and the new errors introduced by RaCE compared with the w/o Branch variant.

\hline \mathbf{Ambiguity} & \mathbf{Samples} & \mathbf{w/o\\;Branch\\;Route\\;ACC\\;(\\%)} & \mathbf{RaCE\\;Route\\;ACC\\;(\\%)} & \mathbf{Corrected} & \mathbf{New\\;Errors} \\\\ \hline \text{Low} & 299 & 94.0 & 94.0 & 1 & 1 \\\\ \text{Medium} & 269 & \mathbf{94.1} & 92.9 & 2 & 5 \\\\ \text{High} & 432 & 83.8 & \mathbf{88.0} & 30 & 12 \\\\ \hline \end{array}$$ The results show that the benefits of multi-branch retention are primarily concentrated in high-ambiguity scenarios, rather than consistently improving performance across all samples. In the low-ambiguity group, both methods achieve a Route ACC of 94.0%. In the medium-ambiguity group, RaCE underperforms by 1.2 percentage points, mainly because some incorrect branches receive higher scores from subsequent visual evidence, causing previously correct routes to be replaced. In contrast, in the high-ambiguity group, RaCE successfully corrects 30 complete-route errors, improving Route ACC by 4.2 percentage points. It is important to note that branch retention inevitably allows incorrect routes to receive higher scores due to misleading later evidence. Because local visual and geometric scoring cannot fully resolve candidate ambiguity, multi-branch reasoning does not guarantee that every reassessment improves the decision. However, sequential decision-making methods such as CoT and Spatial-Agent, as well as limited-width search methods such as Beam-4, may discard correct routes through early commitments or intermediate pruning, preventing later evidence from correcting these errors. In contrast, RaCE retains all candidate branches and uses subsequent evidence to reassess earlier choices. Thus, the main value of branch retention is not to eliminate all local decision errors, but to avoid prematurely discarding correct routes when early evidence is insufficient, preserving opportunities for later correction. **Responce to Q2:** We thank the reviewer again for the suggestion. We have addressed the construction of navigation instructions, reference routes, and candidates, as well as candidate ambiguity, in our response to **Weakness 2**. Here, we further clarify the relationship between reference label generation and RaCE's geometric scoring. The reference routes are determined in advance based on road reachability, distance, and POI category constraints, rather than selected using RaCE's fused scores. Distractor selection and RaCE's scoring function use the same definitions of direction and distance: the geometric penalty is zero when the candidate's bearing deviation from the current anchor does not exceed 50°, and its block count, obtained by dividing the Euclidean distance by 180 meters per block and rounding up, satisfies the instruction. Thus, both share the same geometric criteria, but the reference labels are not determined by RaCE's fused scores. Since multiple candidates may satisfy these constraints, geometric scoring alone cannot uniquely identify the correct route; POI semantics and subsequent spatial relationships are also needed.
Reviewer_vtHgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses urban spatial multi-hop reasoning, where a model must follow a sequential natural-language instruction over a road network to produce both a complete route and a spatial answer. The authors observe that each hop choice redefines the spatial reference frame, so that after an incorrect selection the candidate set and the meaning of relative directions both change, allowing an incorrect branch to remain locally self-consistent. The proposed method, RaCE, maintains a decision state per branch consisting of the current anchor, incoming heading, and road orientation; scores each candidate by combining a calibrated VLM visual log-odds with a geometric deviation term over bearing and distance violations, without normalizing within candidate sets so that scores stay comparable across branches; accumulates these scores along complete routes and pools them by answer for separate route and answer selection; and reduces VLM calls through an adaptive acquisition scheme that assigns bounds to unqueried visual scores, propagates them to route and answer intervals, and stops once the leader’s lower bound exceeds all competitors’ upper bounds.

Critical strengths
  1. A clear and well-substantiated problem formulation. The paper articulates precisely why multi-hop urban navigation differs from sequential reasoning over fixed inputs: each hop choice redefines the spatial anchor, which changes both the reachable candidate set and the referent of relative instructions such as “right” or “ahead,” so an incorrect branch keeps looking locally consistent under its own reference frame. This is quantified rather than asserted, with the appendix showing that correct and incorrect branches have nearly disjoint candidate and downstream route sets, and that final-answer accuracy barely distinguishes the two. This framing would remain informative even if the specific method were replaced.

  2. Adaptive visual acquisition with a deterministic equivalence guarantee. Unqueried visual scores are bounded, propagated monotonically to route and answer intervals, and querying halts only when the leader’s lower bound exceeds every competitor’s upper bound, with correctness and termination proved. Most work on reducing VLM calls settles for “accuracy does not drop much”; an exact-output guarantee is a stronger claim, and the ablation confirms it holds.

Critical weaknesses
  1. The headline contribution is not supported by the paper’s own ablation. The title and abstract center on maintaining multiple candidate routes so later evidence can reassess earlier decisions, yet w/o Branch, which collapses to greedy decoding, drops route accuracy only from 91.1 to 89.6. The components that actually carry performance are geometric scoring (−81.0), reference updating (−76.4), and cross-hop accumulation (−70.5), none of which require multiple branches.

  2. The reported margins rest on baselines that appear misconfigured. With three hops and four candidates, chance route accuracy is about 1.6%, yet CoT averages 5.1% despite 50.9% first-hop accuracy, and both tool-augmented agents sit at 24.9–27.5 on Hop-1 ACC across all four backbones, indistinguishable from the 25% chance level. Since unparseable outputs are counted as incorrect, parse failure rates are needed before these numbers can be read.

  3. The benchmark is self-constructed and offers no external comparability. Evaluation rests entirely on roughly 10,000 OSM-rendered instances with fixed depth three and four candidates per hop, with no results on any public VLN or urban benchmark. The answer space is a four-way distance interval, so chance answer accuracy is 25%, and the appendix itself shows final-answer accuracy differs by only 1.6 points between correct and incorrect branches, which makes the reported 92.1% hard to interpret. Since the visual observations are rendered maps rather than street-level or satellite imagery, and the geometry is read noise-free from OSM, it is unclear how much of the advantage survives outside this setting.

Questions for authors
  1. How should the w/o Branch result be reconciled with the paper’s central claim? Removing multi-branch retention drops route accuracy from 91.1 to 89.6, while removing geometric scoring, reference updating, or cross-hop accumulation drops it to 10.1, 14.7, and 20.6 respectively. Since all three of the latter are available to a greedy decoder, could the authors report a greedy baseline that uses the full RaCE local scoring, that is, geometric evidence, reference updating, and cross-hop accumulation, but commits to the locally best candidate at each hop? If that baseline also reaches roughly 89–90, the contribution attributed to route-level reassessment would need to be restated.

  2. What are the parse failure rates for each baseline? Tool-augmented agents report 24.9–27.5 Hop-1 ACC across all four backbones, indistinguishable from the 25% chance level, and CoT reaches 5.1% route accuracy against a 1.6% chance level despite 50.9% first-hop accuracy. Given that unparseable or out-of-set outputs are counted as incorrect, could the authors report, per baseline and backbone, how many instances failed for format reasons rather than reasoning errors?

  3. Do the baselines receive the same road-graph geometry that RaCE uses? RaCE computes bearing and distance violations directly from G, whereas Beam-4 is described as ranking by backbone probabilities alone. Could the authors state explicitly which geometric signals each baseline receives, and if they do not receive equivalent information, report CoT and Beam-4 augmented with the same geometric scoring term?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer vtHg (Weakness 1)

Responce to W1:

We thank the reviewer for the suggestion. We agree that geometric scoring, spatial reference frame updates, and cross-hop evidence accumulation make substantial contributions to overall performance in the ablation study, but their roles differ from that of the multi-branch mechanism. The former three components provide reliable spatial evidence for evaluating complete routes, while the multi-branch mechanism preserves alternative early choices, allowing subsequent evidence to be used to reassess candidate routes. Since the test set contains samples with varying levels of ambiguity, improvements in overall accuracy may not fully reflect the contribution of the multi-branch mechanism in high-ambiguity scenarios.

To further validate this, we have added a comparison between Greedy and RaCE across different ambiguity groups under the same scoring mechanism in RQ3 on page 8 of the revised manuscript, and analyzed how subsequent evidence affects early decisions.

\hline & & \mathbf{Low\\;(\\%)} & & & \mathbf{Medium\\;(\\%)} & & & \mathbf{High\\;(\\%)} & \\\\ \hline \text{Evidence stage} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} \\\\ \hline \text{Hop-1 only} & 100 & \text{-} & \text{-} & 100 & \text{-} & \text{-} & 88.7 & \text{-} & \text{-} \\\\ \text{Hop1-2} & 100 & \text{-} & \text{-} & 100 & \text{-} & \text{-} & 89.1 & \text{-} & \text{-} \\\\ \text{Full route} & \mathbf{100} & \mathbf{94.0} & \mathbf{94.0} & \mathbf{100} & \mathbf{94.1} & 92.9 & \mathbf{91.9} & 83.8 & \mathbf{88.0} \\\\ \hline \end{array}$$ The results show that the benefits of global reasoning are concentrated in high-ambiguity scenarios. In low- and medium-ambiguity groups, local evidence generally provides reliable candidate discrimination, allowing Greedy to identify most routes correctly. However, branch retention may also allow incorrect routes to overtake correct ones due to errors in subsequent scoring. Specifically, RaCE matches Greedy in the low-ambiguity group but underperforms it by 1.2 percentage points in the medium-ambiguity group. Instance-level analysis reveals that all five newly introduced errors in the medium-ambiguity group result from incorrect second-hop branches receiving higher third-hop visual scores. In contrast, similar early candidate scores under high ambiguity make Greedy more likely to discard correct routes prematurely. By preserving competing route hypotheses, RaCE allows later evidence to revise early decisions, improving Route ACC by 4.2 percentage points. As evidence accumulates, first-hop accuracy increases by 3.2 percentage points, correcting 29 incorrect first-hop decisions, of which 28 ultimately recover the correct complete routes. These findings indicate that RaCE is particularly effective in urban environments with visually similar landmarks and road structures, where cross-hop evidence helps resolve local ambiguity and reduce the impact of early errors on complete route reasoning.
AuthorsOpenReview ↗

Author Response to Reviewer vtHg (Weakness 2)

Responce to W2:

We thank the reviewer for the suggestion. To address concerns about baseline reliability, we further examined output parsing, candidate mappings, and tool execution, and analyzed the impact of multi-hop decision errors and candidate selection bias. The corresponding results have been added to the subsection 'Tool-Agent Candidate Selection Bias Analysis' on page 13 of the revised manuscript's appendix.

(1) Output Parsing and Candidate Mapping Verification.

To rule out evaluation errors caused by output formatting issues or inconsistencies between candidate IDs and their corresponding locations, we conducted 168,000 candidate index mapping checks and found no mapping mismatches. The invalid output rates of each baseline across four backbone models are as follows:

\hline \mathbf{Backbone} & \mathbf{CoT} & \mathbf{Beam\text{-}4} & \mathbf{MSSR} & \mathbf{Spatial\text{-}Agent} \\\\ \hline \text{Qwen2.5-VL} & 7.1\\% & 0.0\\% & 1.7\\% & 0.0\\% \\\\ \text{Qwen3.5} & 0.0\\% & 0.0\\% & 6.4\\% & 3.2\\% \\\\ \text{InternVL3.5} & 1.9\\% & 0.0\\% & 10.6\\% & 4.0\\% \\\\ \text{Kimi-VL} & 1.5\\% & 0.0\\% & 0.0\\% & 3.1\\% \\\\ \hline \end{array}$$ The results show that the invalid output rates of all baselines do not exceed 10.6%. After excluding invalid outputs, the Hop-1 ACC of MSSR and Spatial-Agent remains no higher than 28.4% and 27.6%, respectively, with only minor differences from the original results. This indicates that output parsing and candidate mapping issues are not the primary causes of the low accuracy. (2) Analysis of Multi-Hop Decision Errors in CoT. Complete route identification requires all three hops to be correct. To explain why CoT achieves relatively high first-hop accuracy but substantially lower complete-route accuracy, we calculated the conditional accuracy on samples with valid outputs: $$\begin{array}{lccc} \hline \mathbf{Backbone} & \mathbf{Hop\text{-}1\\;ACC} & \mathbf{Remaining\text{-}two\text{-}hop\\;ACC} & \mathbf{Route\\;ACC} \\\\ \hline \text{Qwen2.5-VL} & 47.6 & 7.2 & 3.4 \\\\ \text{Qwen3.5} & 66.4 & 14.6 & 9.7 \\\\ \text{InternVL3.5} & 42.2 & 9.7 & 4.1 \\\\ \hline \end{array}$$ The results show that even when the first hop is correct, the proportion of CoT predictions with both subsequent hops correct does not exceed 14.61%, resulting in a Route ACC of approximately 5%. This indicates that the low Route ACC of CoT is primarily due to error accumulation in subsequent spatial decisions rather than output parsing failures. (3) Analysis of Candidate Selection Bias in Tool-Augmented Agents. Since parsing and mapping errors cannot fully explain the near-random first-hop accuracy of tool-augmented agents, we further examined their tool execution and candidate selection behavior. We found that both agents successfully executed tool calls, and no inconsistencies between the candidate sets and input histories were observed in the inspected tool observations. However, in the original setting, candidates were sorted by ID, and the agents exhibited a preference for selecting the candidate with the smallest ID. When the correct candidates are approximately evenly distributed across the four IDs, this preference can cause first-hop accuracy to approach 25%. To assess the impact of candidate ordering, we performed balanced permutations of candidate order and IDs on Qwen3.5, while keeping the visual and geometric information and ground-truth labels unchanged. $$\begin{array}{lccccc} \hline \mathbf{Agent} & \mathbf{Original\\;Hop\text{-}1\\;ACC} & \mathbf{Debiased\\;Hop\text{-}1\\;ACC} & \mathbf{Original\\;Route\\;ACC} & \mathbf{Debiased\\;Route\\;ACC} & \mathbf{Min\text{-}ID\\;Selection\\;Rate} \\\\ \hline \text{MSSR} & 26.6 & 33.7 & 2.2 & 4.3 & 24.8 \\\\ \text{Spatial-Agent} & 26.3 & 34.1 & 4.1 & 5.8 & 25.5 \\\\ \hline \end{array}$$ The results show that after balanced permutations, the minimum-ID selection rates of both agents decrease to approximately 25%, while their Hop-1 ACC improves by 7.1 and 7.8 percentage points, respectively, indicating that candidate ordering bias indeed affects local decisions. However, Route ACC improves by only 2.1 and 1.7 percentage points, respectively, remaining below 6%. Further trajectory analysis reveals that although some agents can obtain geometric feedback, such as directional deviations, they fail to effectively utilize the relevant constraints or confuse landmarks with similar visual features, allowing incorrect decisions to continue affecting subsequent reasoning. In summary, the low baseline accuracy is not primarily due to parsing or candidate mapping errors. CoT mainly suffers from multi-hop error accumulation, while tool-augmented agents remain inaccurate at complete-route reasoning even after reducing candidate ordering bias, highlighting their difficulty in integrating visual and geometric evidence for sequential spatial decisions.
AuthorsOpenReview ↗

Author Response to Reviewer vtHg (Weakness 3)

Responce to W3:

We thank the reviewer for their attention to benchmark validity and external comparability. To address these concerns, we have added generalization experiments involving geometric perturbations, changes in visual representations, and external data in the revised manuscript. The corresponding results are presented in the Benchmark Validity and Generalization subsection on pages 13–14 of the revised appendix.

(1) Evaluation Significance of Answer ACC.

Our task requires the model not only to recover the complete route from sequential navigation instructions but also to answer questions about its approximate distance. Accordingly, we use Route ACC and Answer ACC to evaluate complete-route identification and distance interval prediction, respectively. In urban environments, multiple candidate routes may partially satisfy the same navigation instructions, making it difficult to identify the ground-truth route. Answer ACC is intended to evaluate whether the model can correctly answer coarse-grained distance questions under route uncertainty, rather than its ability to calculate distances for known routes.

It is important to clarify that a correct distance answer does not necessarily imply correct complete-route identification. For example, the ground-truth route and an incorrect route may pass through different locations but have lengths of 1.4 km and 1.6 km, respectively, both falling within the 1–2 km interval. Thus, selecting an incorrect route may still yield the correct distance answer. The early-decision intervention experiment in the appendix further illustrates this phenomenon: correcting early errors improves subsequent route matching accuracy by 22.9 percentage points, while Answer ACC increases by only 1.6 percentage points. This indicates that coarse-grained distance answers may mask route reasoning errors. Therefore, Route ACC serves as the primary evaluation metric in this paper, while Answer ACC provides a complementary measure of final distance question answering performance.

(2) Robustness to OSM Geometric Errors.

To assess RaCE's robustness to OSM geometric errors, we conducted perturbation experiments with Qwen3.5-9B. Unlike RQ5, which examines geometric weights, we fixed λgeo=16\lambda_{\mathrm{geo}}=16 and varied perturbations to candidate coordinates, bearings, and road distances to simulate real-world map inaccuracies.

\hline \mathbf{Perturbation\\;Level} & \mathbf{Location} & \mathbf{Heading} & \mathbf{Distance} \\\\ \hline \text{None} & 91.1 & 91.1 & 91.1 \\\\ 5\\,\text{m}\\;/\\;5^{\circ}\\;/\\;\pm5\\% & 85.6 & 89.2 & 85.0 \\\\ 10\\,\text{m}\\;/\\;10^{\circ}\\;/\\;\pm10\\% & 77.8 & 84.1 & 77.5 \\\\ 20\\,\text{m}\\;/\\;20^{\circ}\\;/\\;\pm20\\% & 65.7 & 70.8 & 68.9 \\\\ \hline \end{array}$$ The results show that RaCE's Route ACC gradually decreases as perturbation magnitude increases, with coordinate and distance errors having more pronounced effects. This is because coordinate deviations alter the relative spatial relationships among candidate locations and subsequent reference positions, while distance errors affect hop-level geometric scores and accumulate across hops, ultimately changing the ranking of complete routes. In contrast, bearing perturbations have a relatively smaller impact on candidate ranking. These findings are consistent with the observation in RQ5 that excessively large geometric weights may amplify map errors, indicating that RaCE exhibits some robustness to mild geometric perturbations, but larger spatial errors can degrade its route reasoning performance. (3) Generalization Across Visual Representations and External Data. To evaluate RaCE beyond OSM-rendered maps, we conducted two additional experiments. First, we replaced OSM basemaps with Esri World Imagery satellite imagery while keeping instructions, candidate locations, and road constraints unchanged. Second, we adapted a subset of the public Map2Seq dataset into candidate route inference tasks, retaining its human-written instructions and reference routes. $$\begin{array}{lccc} \hline \mathbf{Evaluation\\;Setting} & \mathbf{Greedy\\;Route\\;ACC} & \mathbf{RaCE\\;Route\\;ACC} & \mathbf{Answer\\;ACC} \\\\ \hline \text{Original OSM-Rendered Maps} & 89.6 & \mathbf{91.1} & 95.7 \\\\ \text{Esri Satellite Imagery} & 83.5 & \mathbf{86.0} & 91.2 \\\\ \text{Map2Seq (Adapted)} & 75.9 & \mathbf{79.0} & 85.4 \\\\ \hline \end{array}$$ The results show that although changes in visual representation and data source lead to some decrease in Route ACC, RaCE still outperforms Greedy by 2.5 and 3.1 percentage points on satellite imagery and the adapted Map2Seq data, respectively. These improvements mainly stem from the complementary roles of visual semantics and road geometry constraints, together with the joint reasoning over cross-hop spatial relationships, allowing subsequent evidence to help reassess ambiguous early choices.
AuthorsOpenReview ↗

Author Response to Reviewer vtHg (Question 1-3)

Responce to Q1:

We thank the reviewer again for the suggestion. Regarding the Greedy baseline suggested by the reviewer, we further clarify that the w/o Branch variant in Table 2 already uses the same visual and geometric scoring, spatial reference frame updates, and cross-hop evidence accumulation as the full RaCE model. The only difference is that it retains only the locally optimal branch at each hop.

\hline \mathbf{Method} & \mathbf{Route\\;ACC} \\\\ \hline \text{Greedy + Full RaCE Scoring} & 89.6\\% \\\\ \text{RaCE} & \mathbf{91.1\\%} \\\\ \hline \end{array}$$ The results show that under the same scoring mechanism, RaCE improves Route ACC by 1.5 percentage points over Greedy. Although Greedy also accumulates cross-hop evidence, its early elimination of alternative candidate branches prevents subsequent evidence from correcting earlier choices. In contrast, RaCE can reassess different route hypotheses. Therefore, geometric scoring and spatial reference frame updates provide an important foundation for overall performance improvements, while the main contribution of the multi-branch mechanism is most evident in high-ambiguity scenarios. This is further supported by the ambiguity-group analysis presented in our response to **Weakness 1**. **Responce to Q2:** We thank the reviewer again. As detailed in our response to **Weakness 2**, we report baseline parse failure rates and analyze the low accuracy of CoT and tool-augmented agents. **Responce to Q3:** We thank the reviewer again for their attention to the fairness of geometric information across baselines. The original experiments followed each baseline's existing information acquisition and reasoning strategies, resulting in differences in how geometric information was accessed and utilized. Specifically, CoT and Beam-4 received map images and candidate location information but did not directly receive precomputed candidate bearings, directional deviations, or distance deviations from the road graph. Although MSSR and Spatial-Agent could query relevant information through public GIS interfaces, their actual use of this information depended on the models autonomously invoking tools and integrating the returned results, without a unified explicit geometric scoring mechanism. In contrast, RaCE directly computes bearing and distance deviations from the road graph and incorporates them into candidate route scoring. To control for differences in geometric scoring, we added CoT+Geometry and Beam-4+Geometry experiments in the **Geometry-Matched Baseline Comparison subsection on page 14 of the revised appendix**. Both baselines retain their original reasoning and search strategies, with only the bearing and distance deviation scores from RaCE added, allowing us to evaluate performance under matched geometric scoring conditions. $$\begin{array}{lrrr} \hline \mathbf{Method} & \mathbf{Hop\text{-}1\\;ACC\\;(\\%)} & \mathbf{First\text{-}2\\;ACC\\;(\\%)} & \mathbf{Route\\;ACC\\;(\\%)} \\\\ \hline \text{CoT} & 60.3 & 28.6 & 14.9 \\\\ \text{CoT + Geometry} & 86.0 & 72.5 & 60.7 \\\\ \text{Beam-4} & 51.2 & 23.4 & 10.6 \\\\ \text{Beam-4 + Geometry} & 88.7 & 77.6 & 63.4 \\\\ \text{RaCE} & \mathbf{96.5} & \mathbf{95.1} & \mathbf{91.1} \\\\ \hline \end{array}$$ The results show that, on Qwen3.5-9B, adding the same geometric scoring improves the Route ACC of CoT and Beam-4 by an average of 49.3 percentage points, demonstrating that explicit geometric constraints substantially improve spatial discrimination among route candidates. However, the enhanced baselines still trail RaCE by 30.4 and 27.7 percentage points, respectively, indicating that geometric scoring alone is insufficient to prevent error accumulation across hops. Specifically, CoT makes sequential decisions and cannot revise early errors using later evidence. Although Beam-4 retains multiple candidate branches, it does not explicitly model dynamic spatial reference-frame updates or cross-hop evidence consistency, allowing routes that satisfy local geometric constraints but violate global spatial relationships to receive high scores. In contrast, RaCE dynamically updates spatial reference frames and accumulates visual and geometric evidence, enabling later observations to reassess earlier route hypotheses. Therefore, RaCE's performance gains arise not merely from geometric scoring, but from the combined effects of spatial reference-frame updating, cross-hop evidence integration, and route-level reasoning.
Reviewer_zAR1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi-hop route inference on a known urban road graph. Given sequential route instructions, a starting location, and visual observations associated with locations, the method selects a sequence of candidate places and predicts the interval containing the total road distance of the route. RaCE uses a vision-language model to assess semantic matches and map geometry to evaluate direction and distance constraints, then accumulates these scores along complete routes. It selects the highest-scoring route and separately predicts a distance interval by pooling evidence from routes that support the same answer. Score bounds allow visual observations to be queried selectively while preserving the outputs of exhaustive querying. The authors evaluate the method on OSM map scenes from five cities using several vision-language models.

Critical strengths
  1. Separating visual recognition from explicit map constraints is a sensible and effective design. The VLM handles the semantics of candidate places, while map computations handle direction, distance, and reachability, with spatial references updated for each candidate route. Results across several backbones and the ablations in Table 2 support this approach. Keeping alternatives when early choices are ambiguous also makes sense, and Table 3 and Figure 5 provide evidence that later observations can help resolve these choices.

  2. Adaptive visual querying provides a clear benefit. With fixed scores and the same tie-breaking rules, the score bounds determine whether further visual evidence could still change the outputs. In Table 2, adaptive querying reduces the average number of logical queries from 21.0 to 3.3 while preserving the outputs of exhaustive querying.

Critical weaknesses
  1. The additional benefit of global reasoning is not yet clear enough. Local selection with the same scoring function already comes close to the full method. I would like a clearer picture of when the more complex global reasoning is helpful.

  2. The practical motivation for the task setup needs more explanation. Its relationship to existing route-understanding tasks, and the practical needs behind fixed candidate sets and distance-interval prediction, are not yet clear to me.

Questions for authors

1. Relationship to existing tasks.

I would like to better understand how this task differs from existing map-based route understanding and vision-language navigation. These tasks also involve following a sequence of instructions using landmarks and directions. Which difficulty is this paper specifically targeting, and where do existing approaches fall short?

2. Practical motivation for the evaluation setup.

The main experiments use three steps with four candidate locations per step, and ask for the interval containing the total route distance. What practical use case does this setup represent? In particular, once the route is determined, its distance can be computed from the map. What ability is the separate distance-interval evaluation intended to measure? A concrete example showing what it tells us beyond route accuracy would help.

3. Benefits of keeping multiple routes.

In Table 2, the w/o Branch variant, which keeps only the best branch at each step, already achieves 89.6 Route ACC, close to 91.1 for the full method. Could you include this comparison across the ambiguity groups in Table 3 to show when keeping multiple routes is most helpful?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer zAR1 (Weakness 1-2)

Responce to W1:

We thank the reviewer for the suggestion. In RQ3 on page 8 of the revised manuscript, we have added a comparison between Greedy and RaCE under the same scoring mechanism to demonstrate the role of global reasoning across different ambiguity levels.

\hline & & \mathbf{Low\\;(\\%)} & & & \mathbf{Medium\\;(\\%)} & & & \mathbf{High\\;(\\%)} & \\\\ \hline \text{Evidence stage} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} & \text{Hop-1 acc} & \text{Greedy route acc} & \text{Ours Route acc} \\\\ \hline \text{Hop-1 only} & 100 & \text{-} & \text{-} & 100 & \text{-} & \text{-} & 88.7 & \text{-} & \text{-} \\\\ \text{Hop1-2} & 100 & \text{-} & \text{-} & 100 & \text{-} & \text{-} & 89.1 & \text{-} & \text{-} \\\\ \text{Full route} & \mathbf{100} & \mathbf{94.0} & \mathbf{94.0} & \mathbf{100} & \mathbf{94.1} & 92.9 & \mathbf{91.9} & 83.8 & \mathbf{88.0} \\\\ \hline \end{array}$$ The results show that the benefits of global reasoning depend on candidate ambiguity. In low- and medium-ambiguity groups, local evidence already distinguishes most candidates reliably, leaving limited room for correction. Meanwhile, misleading later evidence may cause incorrect branches to overtake correct routes. Consequently, RaCE matches Greedy in the low-ambiguity group but trails by 1.2 percentage points in the medium-ambiguity group, where five new errors arise from incorrect second-hop branches receiving higher third-hop visual scores. In contrast, high ambiguity makes early greedy decisions more prone to discarding correct routes. By preserving competing hypotheses for reassessment with later evidence, RaCE improves Route ACC by 4.2 percentage points and first-hop accuracy by 3.2 percentage points. Among 29 corrected first-hop decisions, 28 lead to correct complete routes, with five additional routes recovered through later-hop corrections across the test set. These findings highlight RaCE's value in urban environments with visually similar landmarks and road structures, where cross-hop evidence helps correct early decisions that local evidence alone cannot reliably resolve. **Responce to W2:** We thank the reviewer again for the valuable suggestion. Our task focuses on map localization and distance question answering from natural-language route descriptions. For example, a user may describe traveling east from a subway station to a café, then turning right past a hospital to a park, and ask for the route distance. Since multiple locations may match these descriptions, the system must combine visual semantics and road geometry to recover the intended route. Unlike conventional route planning with known endpoints, early location choices affect subsequent reference positions, directions, and candidate sets. Consequently, multiple routes may satisfy local constraints, making accurate identification of the complete route challenging. Regarding the fixed candidate set, this paper adopts a controlled setting with three hops and four candidates per hop to separate candidate location retrieval from multi-hop route reasoning, enabling a fair evaluation of different methods' reasoning capabilities under consistent conditions. In practical applications, candidate locations can be obtained through map search or POI retrieval, without being restricted to a fixed number. Regarding distance interval prediction, we agree that once a route is identified, its distance can be directly calculated using the road network graph. However, when multiple candidate routes receive high support, the model may struggle to determine which route the user actually described, and the highest-scoring individual route and its corresponding distance interval may still be incorrect. Therefore, we further investigate whether aggregating evidence across different routes can help infer the approximate distance corresponding to the original route description. For example, three candidate routes have relative support weights of 0.40, 0.35, and 0.25, with distances of 0.8, 1.4, and 1.6 km, respectively. Although the first route receives the highest individual score, the latter two routes jointly support the 1–2 km interval with a cumulative weight of 0.60. This illustrates that the most likely individual route and the most strongly supported distance interval do not necessarily coincide. Therefore, Route ACC measures the accuracy of complete route identification, while Answer ACC additionally evaluates the model's ability to aggregate multiple route hypotheses and infer a coarse-grained distance answer under route ambiguity.
AuthorsOpenReview ↗

Author Response to Reviewer zAR1 (Question 1-3)

Responce to Q1:

We thank the reviewer for the suggestion. Our work shares similar goals with existing map-based route understanding and vision-language navigation studies, as both involve multi-step route reasoning based on sequential landmark and directional instructions. Existing approaches typically combine map constraints and visual observations to make step-by-step route decisions, using route verification, memory, or backtracking mechanisms to correct errors. However, in urban environments, similar landmarks and road structures may allow multiple candidate locations to satisfy the same local visual and geometric constraints, making it difficult to identify the correct location based on local evidence alone.

Importantly, an incorrect early location choice not only changes the spatial reference positions, travel directions, and candidate sets for subsequent instructions, but may also lead to an incorrect route that continues to satisfy subsequent local constraints. Consequently, such errors are difficult to detect through local constraint checks, and even when backtracking is triggered, similar local evidence may lead to repeated incorrect selections. This is a key limitation that existing step-by-step decision-making and local error correction mechanisms cannot fully address.

To address this challenge, RaCE maintains multiple candidate routes during reasoning, updates their spatial reference states independently, and accumulates subsequent visual and geometric evidence to reassess earlier choices at the complete-route level. Therefore, the key distinction of our work lies in using subsequent evidence to compare alternative route hypotheses, rather than relying solely on local decisions or backtracking after errors are detected.

Responce to Q2:

We thank the reviewer again for their attention to this issue. As discussed in our response to Weakness 2, we adopt a controlled setting with three hops and four candidates per hop to separate candidate retrieval from multi-hop route reasoning, enabling a fair evaluation of different methods under consistent conditions. In practical applications, candidate locations can be obtained through map search or POI retrieval without being restricted to a fixed number.

Regarding distance interval prediction, we agree that once the correct route is identified, its distance can be directly calculated from the map. We introduce this auxiliary task not to evaluate distance calculation ability, but to investigate how uncertainty in route identification affects downstream spatial question answering. For example, when a user asks for the approximate distance of a route, two candidate routes may pass through different landmarks and have distances of 1.4 and 1.6 km, respectively, yet both fall within the 1–2 km interval. Thus, even if the model fails to identify the correct route, it may still predict the correct distance interval. In our evaluation, among 89 samples with incorrect route predictions, 58 still have correct distance interval predictions.

Therefore, in this paper, Route ACC serves as the primary metric for evaluating complete-route reasoning ability, while Answer ACC serves as an auxiliary metric reflecting coarse-grained spatial question answering performance under route ambiguity.

Responce to Q3:

We thank the reviewer again for the valuable suggestion. As discussed in our response to Weakness 1, the benefits of maintaining multiple candidate routes are most pronounced in high-ambiguity scenarios. To this end, we have added a comparison between Greedy and RaCE across different ambiguity groups under the same scoring mechanism in Table 3 of the revised manuscript. The results show that both methods achieve the same Route ACC in the low-ambiguity group. In contrast, in the high-ambiguity group, multiple early candidate locations exhibit similar visual semantics and geometric constraints, making it difficult for the local greedy strategy to distinguish among them accurately. By maintaining multiple candidate routes and using subsequent evidence to reassess earlier choices, RaCE improves Route ACC by 4.2 percentage points, further demonstrating the effectiveness of global reasoning when early decisions are ambiguous.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.