acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic-Symbolic Intent Disambiguation for Robust Embodied Control under Ambiguous Speech

Abstract

Speech provides a natural and flexible interface for embodied control, yet its inherent ambiguity challenges accurate task understanding and reliable execution. Meanwhile, embodied interaction continuously provides new environmental information that can further constrain plausible interpretations of the original spoken instruction. Robust speech-conditioned embodied control therefore requires an agent to continuously refine its intent understanding during execution and actively explore discriminative environmental information when competing intents lead to action disagreement. To this end, we propose a probabilistic-symbolic intent disambiguation framework, where a probabilistic measure quantifies how speech ambiguity affects action decisions, while symbolic reasoning localizes and resolves the critical intent differences underlying action disagreement through embodied interaction. Specifically, we introduce speech-induced action uncertainty to quantify how uncertainty in speech interpretation propagates to the current action decision, providing a criterion for when disambiguation is needed. We then develop a unified symbolic intent–world modeling method that maps candidate speech intents and observations into a shared symbolic space for action reasoning and intent disambiguation. Building on this representation, we introduce world-constraint alignment for speech reinterpretation, which localizes intent differences to guide targeted exploration and induces world constraints from interaction feedback to refine speech interpretations. We further introduce SPEAR-Bench, a benchmark that systematically varies perceptual, paralinguistic, and content ambiguity to evaluate task understanding and embodied control under ambiguous speech. Experiments on SPEAR-Bench and REI-Bench show that our method consistently improves task success while reducing interaction cost compared with strong baselines for embodied control under ambiguous speech.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Neuro-symbolic, Embodied Planning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_rH6FOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

this paper proposes a probabilistic-symbolic intent disambiguation framework, where a probabilistic measure quantifies how speech ambiguity affects action decisions, symbolic reasoning localizes and resolves the critical intent differences underlying action disagreement through embodied interaction. This paper also introduce a benchmark SPEAR-Bench. which systematically varies perceptual, paralinguistic, and content ambiguity to evaluate task understanding and embodied control under ambiguous speech.

Critical strengths

1.the paper is focusing on ambiguity at task intent level rather than improving speech recognition. It's practical in embodied device use cases. 2. This work recognize the action uncertainty from intent uncertainty.
3. this work take into consideration the environmental feedback with language understanding.

Critical weaknesses

The environmental evidence does not always reflect user intent. exploring the environment may not help with problems like if a persons is looking for an object but both objects are visible in the scene. Also, a certain object not visible in the seen doesn't mean it doesn't exist. she the constrains in figure 2 might not help with the disambiguation.

Questions for authors
  1. how are different levels of ambiguity defined and classified? does the proposed method perform consistently across different levels of ambiguity?
  2. Does the evaluation across different languages and Accents? Since the input is speech if I understand correctly and it's not text. the accents and pronunciation may affect the final accuracy. How robust is the proposed method?
  3. The proposed methods show it will use environment feedback, I am wondering if complexity of the environment affects the final results.
  4. How many different environments and scenarios have been tested? how does the method perform as environmental complexity increases.
4 comments
AuthorsOpenReview ↗

Response to Reviewer rH6F (Part 1 of 2)

Dear Reviewer rH6F,

We sincerely thank you for your thoughtful review and constructive feedback. We greatly appreciate your recognition of the practical importance of task-level speech ambiguity, our modeling of how intent uncertainty affects action decisions, and the integration of environmental feedback into speech understanding. We address your questions and clarify our contributions below.

Q1: Ambiguity Definitions and Performance Across Levels

  • how are different levels of ambiguity defined and classified? does the proposed method perform consistently across different levels of ambiguity?

Ambiguity Definition and Classification. We thank the reviewer for the question regarding ambiguity definitions and levels. SPEAR-Bench defines three ambiguity dimensions with three levels each: (1) Perceptual (P): P0–P2 correspond to zero, one, and at least two corrupted task-relevant speech segments; (2) Paralinguistic (T): T0–T2 contain zero, one, and at least two cue types, such as tone, fillers, and pauses; (3) Content (C): C0 uses direct instructions, C1 introduces interruptions or self-corrections, and C2 involves context-dependent references or omissions. All variants preserve the original task goals and environments and undergo automated checks and human verification.

Performance Across Ambiguity Levels. As shown in Table R2, our method achieves the highest success across most ambiguity levels and shows greater robustness as perceptual and content ambiguity increase, while baselines degrade more substantially. Some methods perform better at C1 than C0 with Qwen, whereas DeepSeek shows the opposite trend. This is because speech ambiguity introduces not only interference but also additional information that Qwen can exploit through speech cues, whereas DeepSeek relies on ASR transcripts and cannot fully utilize such information.

Table R2: Performance Across Ambiguity Levels on SPEAR-Bench (Success, %)

(a) DeepSeek-V4.1-Flash

AmbiguityLevelReActReflActWALL-ETOCCOurs
PerceptualP058.3369.7559.2671.6080.66
P158.0268.8357.1071.3078.19
P249.3861.4251.2361.7375.21
ParalinguisticT057.1072.2259.8864.8178.91
T154.3263.8954.0171.3078.81
T254.3263.8953.7068.5276.34
ContentC066.9878.7068.8381.7986.42
C163.2774.6962.3578.4080.76
C235.4946.6036.4244.4466.87

(b) Qwen3.8-Omni-Flash

AmbiguityLevelReActReflActWALL-ETOCCOurs
PerceptualP070.6871.9169.4479.0178.91
P170.3773.4670.9975.3179.84
P265.1269.1463.5871.9179.01
ParalinguisticT075.0078.4075.0075.9379.73
T164.8166.0562.6575.6278.40
T266.3670.0666.3674.6979.63
ContentC070.0672.8470.0684.2682.92
C187.3589.5186.1183.3383.33
C248.7752.1647.8458.6471.50
AuthorsOpenReview ↗

Response to Reviewer rH6F (Part 2 of 2)

Q2: Robustness Across Languages

  • Does the evaluation across different languages and Accents? Since the input is speech if I understand correctly and it's not text. the accents and pronunciation may affect the final accuracy. How robust is the proposed method?

We thank the reviewer for highlighting cross-language robustness. To further evaluate our method across languages, we conduct a supplementary French speech experiment on SPEAR-Bench under the P0–T0–C0 condition. As shown in Table R3, our method achieves 83.33% and 77.78% task success with English and French speech, respectively, a decrease of only 5.55 percentage points. These results provide preliminary evidence of our method's robustness across spoken languages.

Table R3: Task Success Across Languages on SPEAR-Bench (P0–T0–C0).

LanguageSuccess (%)
English83.33 ± 2.78
French77.78

Q3/Q4:Environmental Coverage and Complexity Analysis

  • How many different environments and scenarios have been tested? how does the method perform as environmental complexity increases.
  • The proposed methods show it will use environment feedback, I am wondering if complexity of the environment affects the final results.

Environment and Scenario Coverage. We thank the reviewer for highlighting environmental coverage and complexity. The 36 source tasks in SPEAR-Bench cover all six ALFWorld task types, all four scene types, and 30 distinct FloorPlans, spanning diverse task goals, environment layouts, and interaction requirements. Each task is further instantiated under 27 speech ambiguity conditions, yielding 972 evaluation instances. Detailed task and scene distributions are provided in Appendix C.1 (Table 7).

Performance Across Environmental Complexity. To analyze how environmental complexity affects performance, we divide the environments into Low, Medium, and High complexity groups based on the number of searchable locations. As shown in Table R4, task success decreases and interaction steps increase with environmental complexity. This trend is consistent with the greater exploration effort required to acquire discriminative evidence in larger search spaces, indicating that environmental complexity is an important factor affecting task success and interaction cost.

Table R4: Performance across environmental complexity levels on SPEAR-Bench (Qwen3.8-Omni-Flash).

ComplexityAll Search Loc.# TasksSuccess (%) ↑Steps ↓
Low≤171384.5219.13
Medium18–501977.7125.01
High≥51469.4429.92
Reviewer_rH6FOpenReview ↗

feedback

Thank you for the detailed responses and additional experiments. The clarification of ambiguity levels and the analysis of performance across different ambiguity conditions address my main concerns. While the evaluation of different accents and more diverse languages remains limited. I will maintain my original score.

Thank you for your positive feedback. We are pleased that our clarifications and additional experiments have addressed your main concerns. We sincerely appreciate the time and effort you have devoted to reviewing our work.

We agree that broader evaluation across diverse languages and accents would further strengthen our work. We will acknowledge this limitation and consider extending SPEAR-Bench for systematic cross-language and cross-accent evaluation in future work.

Reviewer_L3tzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies embodied control under ambiguous spoken instructions and proposes a probabilistic-symbolic intent disambiguation framework. The authors define the mutual information between the task intent and the next action as speech-induced action uncertainty, and use it to judge whether speech ambiguity already affects the current decision. Above a threshold, the agent compares competing candidate intents in a shared symbolic space, localizes the key differences, verifies them through environment exploration, and converts the feedback into world constraints for reinterpreting the original speech. The authors also build SPEAR-Bench from 36 ALFWorld tasks, controlling speech ambiguity along perceptual, paralinguistic and content dimensions, for 27 conditions and 972 instances. They compare against ReAct, ReflAct and WALL-E on SPEAR-Bench and REI-Bench, and report that their method improves task success and reduces interaction steps.

Critical strengths
  1. The paper clearly separates "whether the intent is uncertain" from "whether this uncertainty affects the current action", and uses the mutual information between the task intent and the next action as the criterion for triggering disambiguation. This is closer to decision needs than measuring intent or action uncertainty alone, and it gives a clear basis for the view that speech ambiguity need not be fully resolved before acting. Deciding whether to trigger, localizing the disagreement, exploring the environment and using the feedback to reinterpret the instruction form a closed loop, with intermediate results that can be inspected at each step. This idea is a useful reference for how embodied agents handle ambiguous instructions.

  2. SPEAR-Bench systematically varies the source and level of speech ambiguity while holding the task and environment fixed. This controlled design makes it possible to examine separately how different ambiguities affect intent understanding and task execution, whereas existing text-instruction benchmarks do not cover speech-level factors. The paper also provides condition-level results and ablations, which give a basis for further analysis. This is a valuable contribution beyond the method itself.

Critical weaknesses
  1. The main experimental results need more complete evaluation information to support them. Beyond the overall benchmark size, the paper does not report the number of valid evaluations and repetitions for each method under each condition, and it does not state the sample size used on REI-Bench. Some aggregate results do not indicate the corresponding backbone. This affects the interpretation of the results and the comparison between methods. The 972 instances of SPEAR-Bench come from 36 source tasks, so the conditions are not independent, yet the paper also reports no confidence intervals, which makes it difficult to judge whether the performance gaps are stable. In addition, none of the compared methods is designed specifically for ambiguous instructions, and it is also unclear whether SPEAR-Bench adequately covers cases in which the candidate intents are all feasible and the available environmental evidence cannot distinguish between them. Until this information is provided, it is difficult to judge accurately the size and stability of the performance gaps.

  2. The current experiments give preliminary support to the two core mechanisms, but they do not adequately explain where the performance improvement comes from. The condition-level results in the appendix show that the way the improvement varies with the type of content ambiguity differs between the two backbones, yet the paper neither performs module-level ablations by ambiguity type nor reports statistics on disambiguation triggers and exploration outcomes. It therefore cannot yet show how each mechanism works under different conditions. Speech information is handled mainly by the underlying speech-capable model; the framework itself does not explicitly model speech-specific factors, and the incremental value of speech is unclear. The existing ablation only replaces the mutual-information criterion with direct LLM judgment, without comparing against simple criteria such as intent entropy or action entropy; symbolic disagreement localization and world-constraint alignment are also replaced as a whole, so the contribution of each component cannot be separated. These results therefore support the full system over the ablation variants in task success under the tested configuration, but they are not yet sufficient to validate the specific advantage of speech-induced action uncertainty or the role of each component of the symbolic disambiguation procedure.

Questions for authors
  1. Please state the number of tasks actually evaluated and the number of repetitions for each method under each condition, and provide the raw success counts. Please also explain how the aggregate tables are computed, the success criterion and baseline configuration used for each method, and how REI-Bench was sampled and integrated. In addition, please explain how the method handles cases in which the candidate intents are all feasible and the available environmental evidence cannot distinguish between them, and report the share of such samples in SPEAR-Bench and their success rate.

  2. Please report the trigger rate and exploration outcomes for each type of ambiguity. Please also add controls using only intent entropy, only action entropy and random triggering, with a threshold determined separately on the calibration set for each criterion and the same trigger rate for random triggering, and give results with confidence intervals. While keeping the symbolic representation, please ablate disagreement localization, the symbolic world constraints and the language-level interpretation constraints separately, report the accuracy of the symbolic parser and of the derived constraints, and run the same model with audio and with transcripts.

5 comments
AuthorsOpenReview ↗

Response to Reviewer L3tz (Part 1 of 5)

Dear Reviewer L3tz,

Thank you for your careful review and constructive feedback. We appreciate your recognition of our uncertainty modeling and interactive disambiguation framework. We address your concerns regarding experimental reliability, baseline comparisons, and component effectiveness below.

Q1/W1: Evaluation Protocol and Reporting Details

  • Please state the number of tasks actually evaluated and the number of repetitions for each method under each condition, and provide the raw success counts. Please also explain how the aggregate tables are computed, the success criterion and baseline configuration used for each method, and how REI-Bench was sampled and integrated.
  • The main experimental results need more complete evaluation information to support them. Beyond the overall benchmark size, the paper does not report the number of valid evaluations and repetitions for each method under each condition, and it does not state the sample size used on REI-Bench. Some aggregate results do not indicate the corresponding backbone. This affects the interpretation of the results and the comparison between methods.

Response to Q1/W1: We thank the reviewer for the constructive suggestions on evaluation protocols and reporting completeness. In response, we have clarified the evaluation sizes, repetitions, success criteria, and backbone configurations for SPEAR-Bench and REI-Bench. We have also carefully checked and corrected the experimental reporting, providing per-condition evaluation counts, raw success counts, and aggregation procedures. The corresponding details and results have been updated in the revised manuscript and appendix.

  1. Evaluation Protocol Details

SPEAR-Bench. For fair comparison, we consistently define Success across all 972 instances as achieving the environment-specified task goal within at most five attempts, with all methods following the same 50-step interaction budget, while additionally reporting Final Intent Accuracy for our method. Our method is evaluated over three independent runs, reporting mean and standard deviation, while each baseline is evaluated once. Raw success counts and per-condition results for both backbones are provided in Appendix D.2 (Tables 8–9), with the Qwen3.8-Omni-Flash results reproduced in Part 5 of this response.

REI-Bench. We select 30 source tasks from REI-Bench, each evaluated under nine combinations of context conditions and referring-expression types, yielding 270 instances covering six task types and four scene types. All methods use the same DeepSeek-V4.1-Flash backbone and success criterion as SPEAR-Bench, with results reported from a single evaluation run. Detailed results are provided in Appendix D.4 (Table 11).

  1. Experimental Results and Raw Success Counts

We have carefully revisited and updated the experimental reporting, providing task success rates, raw success counts, and interaction steps on SPEAR-Bench and REI-Bench, with the corresponding backbones explicitly specified. As shown in Tables R1–R2, our method demonstrates stable performance across three independent runs on SPEAR-Bench and achieves higher task success than ReAct, ReflAct, and WALL-E with both backbones. On REI-Bench, our method also achieves the highest task success and the lowest interaction cost. Per-condition evaluation counts and raw success counts are provided in Appendix D.2 (Tables 8–9); the Qwen3.8-Omni-Flash results are also included in Part 5.

Table R1: SPEAR-Bench Results and Success Counts.

(a) DeepSeek-V4.1-Flash

MethodSuccess (%) ↑Steps ↓Intent Acc. (%) ↑
ReAct55.25 (537/972)27.37—
ReflAct66.67 (648/972)25.14—
WALL-E55.86 (543/972)27.23—
Ours78.02 ± 1.29 (2275/2916)24.33 ± 1.5882.44 ± 1.56 (2404/2916)

(b) Qwen3.8-Omni-Flash

MethodSuccess (%) ↑Steps ↓Intent Acc. (%) ↑
ReAct68.72 (668/972)24.54—
ReflAct71.50 (695/972)23.30—
WALL-E68.00 (661/972)24.65—
Ours79.25 ± 2.03 (2311/2916)23.43 ± 0.4982.34 ± 3.20 (2401/2916)

Table R2: REI-Bench Results and Success Counts with DeepSeek-V4.1-Flash.

MethodSuccess (%) ↑Steps ↓Intent Acc. (%) ↑
ReAct50.74 (137/270)31.19—
ReflAct60.00 (162/270)27.38—
WALL-E47.04 (127/270)32.38—
Ours82.96 (224/270)21.2385.19 (230/270)
AuthorsOpenReview ↗

Response to Reviewer L3tz (Part 2 of 5)

W1: Statistical Reliability and Confidence Intervals

  • The 972 instances of SPEAR-Bench come from 36 source tasks, so the conditions are not independent, yet the paper also reports no confidence intervals, which makes it difficult to judge whether the performance gaps are stable.

We thank the reviewer for the constructive suggestion regarding statistical reliability. To further validate the consistency and robustness of our performance gains across source tasks, we conduct a paired bootstrap analysis over the 36 independent SPEAR-Bench source tasks, using 10,000 resamples to estimate success-rate gains over the baselines and their 95% confidence intervals, without treating ambiguity conditions derived from the same source task as independent samples.

As shown in Table 10 (Appendix D.3), our method achieves positive average success gains over ReAct, ReflAct, and WALL-E with both DeepSeek-V4.1-Flash and Qwen3.8-Omni-Flash, with all 95% confidence interval lower bounds above zero. These results further demonstrate consistent and statistically supported average performance gains across variations in source-task composition. The analysis is provided in Appendix D.3 of the revised manuscript.

Table 10 (Appendix D.3): Success Gains and 95% Confidence Intervals on SPEAR-Bench.

BackboneBaselineΔ Success (pp) ↑95% CI
DeepSeek-V4.1-FlashReAct+22.77[+10.84, +34.91]
ReflAct+11.35[+2.43, +19.96]
WALL-E+22.15[+10.29, +34.09]
Qwen3.8-Omni-FlashReAct+10.53[+3.12, +18.14]
ReflAct+7.75[+0.48, +15.29]
WALL-E+11.25[+3.12, +19.96]

Note: Δ Success = Ours − Baseline, measured in percentage points. The 95% CIs are estimated using 10,000 paired bootstrap resamples of the 36 source tasks, with success rates averaged across 27 conditions per task.

W1: Comparison with an Ambiguity-Aware Baseline

  • In addition, none of the compared methods is designed specifically for ambiguous instructions

We thank the reviewer for suggesting comparisons with methods specifically designed for ambiguous instructions. We additionally include TOCC[1], an ambiguity-aware method that addresses ambiguous referring expressions through contextual interpretation and instruction rewriting, and evaluate it on SPEAR-Bench and REI-Bench under the same evaluation protocol. On SPEAR-Bench, our method is evaluated over three independent runs, yielding 2,916 evaluations (972 instances × 3), while TOCC is evaluated once.

As shown in Tables R3–R5, our method achieves higher task success across both benchmarks and backbones, further demonstrating its effectiveness in embodied task execution under ambiguous instructions. The corresponding results have been included in the revised manuscript.

Table R3: Comparison with TOCC on SPEAR-Bench (DeepSeek-V4.1-Flash).

MethodSuccess (%) ↑Success CountSteps ↓
TOCC68.21663/97222.16
Ours78.02 ± 1.292275/291624.33 ± 1.58

Table R4: Comparison with TOCC on SPEAR-Bench (Qwen3.8-Omni-Flash).

MethodSuccess (%) ↑Success CountSteps ↓
TOCC75.41733/97221.08
Ours79.25 ± 2.032311/291623.43 ± 0.49

Table R5: Comparison with TOCC on REI-Bench (DeepSeek-V4.1-Flash).

MethodSuccess (%) ↑Success CountSteps ↓
TOCC63.33171/27027.78
Ours82.96224/27021.23

[1] Jiang, C., Zhou, C., and Yang, J. REI-Bench: Can Embodied Agents Understand Vague Human Instructions in Task Planning? International Conference on Learning Representations (ICLR), 2026.

AuthorsOpenReview ↗

Response to Reviewer L3tz (Part 3 of 5)

Q1/W1:Handling Indistinguishable Candidate Intents

  • In addition, please explain how the method handles cases in which the candidate intents are all feasible and the available environmental evidence cannot distinguish between them, and report the share of such samples in SPEAR-Bench and their success rate.
  • it is also unclear whether SPEAR-Bench adequately covers cases in which the candidate intents are all feasible and the available environmental evidence cannot distinguish between them.

Handling Indistinguishable Intents. When multiple candidate intents have similar probabilities and remain compatible with the available environmental evidence, our method uses speech-induced action uncertainty to determine whether their differences affect the current action decision. If no substantial action disagreement exists, the agent proceeds with task execution; otherwise, it formulates verification hypotheses based on decision-relevant differences and performs targeted exploration. If the candidates remain indistinguishable after exploration, the agent selects the action with the highest probability under the belief-weighted policy: $ a_i^* = \arg\max_a \sum_k b_i^k \pi_i^k(a). $ The agent subsequently reinterprets the original speech instruction as new environmental observations become available.

Benchmark Coverage and Analysis. SPEAR-Bench introduces different sources of speech ambiguity while preserving the original ALFWorld task intents and environments, with human verification ensuring that the intended tasks remain unchanged. Whether environmental evidence can distinguish candidate intents depends on the interpretations generated by the model, rather than solely on static benchmark properties. Therefore, the current benchmark does not explicitly annotate cases in which candidate intents are completely indistinguishable.

To further examine intent recognition and task completion after environmental interaction, we analyze their joint outcomes. As shown in Table R6, task success reaches 91.42% when the final intent is correctly identified, but drops substantially to 22.52% when intent recognition is incorrect. These results indicate a strong association between accurate intent recognition and task success, further highlighting the importance of effective intent disambiguation in embodied task execution.

Table R6: Intent Recognition and Task Success on SPEAR-Bench (Qwen3.8-Omni-Flash).

Final Intent RecognitionGoal Completed (%)Goal Not Completed (%)Success Rate (%)
Correct75.27 ± 1.74(2,195)7.06 ± 1.50(206)91.42
Incorrect3.98 ± 0.31(116)13.68 ± 3.50(399)22.52

Q2:Exploration Trigger Rates and Verification Outcomes

  • Please report the trigger rate and exploration outcomes for each type of ambiguity.

We thank the reviewer for the constructive suggestion. We report exploration trigger rates and positive verification rates across ambiguity levels on SPEAR-Bench with Qwen3.8-Omni-Flash, covering all evaluation runs and executed attempts, including failures. The metrics are defined as: $ \mathrm{Trigger\ Rate}=\frac{N_{\mathrm{trigger}}}{N_{\mathrm{steps}}}\times100%$, $\mathrm{Positive\ Verification\ Rate}=\frac{N_{\mathrm{positive}}}{N_{\mathrm{trigger}}}\times100%. $ Here, NtriggerN_{\mathrm{trigger}}, NstepsN_{\mathrm{steps}}, and NpositiveN_{\mathrm{positive}} denote exploration triggers, environment steps, and positive verification outcomes, respectively. Each ambiguity level aggregates the nine combinations of the other two dimensions.

As shown in Table 5 (Appendix A.3), the overall trigger rate is 14.55%, with a positive verification rate of 52.25%. Exploration is generally triggered more frequently at higher ambiguity levels, particularly for content ambiguity (11.52% to 18.76%), while positive verification rates remain between 50% and 54% across conditions. These results suggest adaptive exploration triggering with comparable positive verification rates across ambiguity levels. Results with DeepSeek-V4.1-Flash are provided in Table 6 (Appendix A.3).

Table 5 (Appendix A.3): Exploration Trigger Rates and Verification Outcomes on SPEAR-Bench (Qwen3.8-Omni-Flash).

ConditionStepsTriggersTrigger rate (%)Verified Positive (%)
P0P_023.193.3014.2353.62
P1P_123.323.1313.4250.00
P2P_223.793.8116.0253.10
T0T_023.593.4714.7152.58
T1T_123.273.3114.2252.48
T2T_223.443.4614.7651.71
C0C_020.312.3411.5252.35
C1C_120.942.4611.7553.32
C2C_229.055.4518.7651.05
Overall23.433.4114.5552.25

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.