acceptodds
Under review as a conference paper at ICLR 2027

Multiple-Choice Supervision for Embodied Vision-Language Models: A Data-Centric Fine-Tuning Framework

Abstract

Supervised fine-tuning (SFT) can improve the perception and reasoning of embodied vision-language models (VLMs), yet how supervision should be formulated to promote transfer across embodied tasks remains unclear. In particular, comparisons between choice and open-ended training conflate the information supplied by candidate answers with the form of the supervised target. To disentangle these factors, we introduce Embodied Choice Tuning (ECT), a data-centric framework that jointly designs capability coverage, candidate inputs, distractors, and output targets. ECT retains candidate sets with confusable distractors in the input and supervises the correct option with a compact letter target. We instantiate ECT with a corpus of 10,000 examples spanning eight embodied capabilities and 102 question types. From the same visual observations, questions, and correct answers, we derive four supervision formats that isolate candidate availability, distractor semantics, and output targets. We evaluate the resulting models through Embodied Arena on nine established benchmarks in their native task formats. Across three model families, ECT yields the highest aggregate performance among the four formats. On Qwen3.5-4B, this advantage amounts to 3.89 points over the base and 2.27 points over open-ended supervision. Controlled comparisons further show that candidate inputs improve transfer when answer targets are fixed. A controlled distractor intervention shows that confusable distractors outperform clearly mismatched alternatives for all three choice targets. With candidates fixed, supervising answer text, either alone or alongside the letter, does not outperform letter-only supervision. Together, these results support ECT as a practical approach for data-limited embodied SFT: retain confusable semantic alternatives in the input and supervise the final decision with a compact letter target.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
embodied intelligence ; data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gjsUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Embodied Choice Tuning (ECT), a data-centric framework for improving embodied vision-language models through multiple-choice supervision. The authors construct a dataset of 10,000 examples covering eight embodied capabilities and compare four supervision formats: Letter, Content, Joint, and Open. The framework combines semantically confusable distractors with compact letter-based supervision. Experiments across three model families and nine Embodied Arena benchmarks show that ECT achieves the highest aggregate performance among the tested formats. Additional experiments investigate the contributions of candidate availability, distractor difficulty, and output target formulation.

Critical strengths
  1. Well-controlled comparison of supervision formats. The authors construct four supervision variants from identical visual observations, questions, and correct answers. In particular, comparing Content and Open while keeping target sequences identical helps isolate the contribution of candidate information. This experimental design provides useful evidence about how supervision formulation affects embodied VLM adaptation.

  2. Broad empirical evaluation with consistent improvements. The study evaluates three model families across nine benchmarks covering spatial understanding, visual reasoning, and embodied perception. ECT achieves the highest aggregate score among the four formats in all five reported model and seed comparisons. The additional Near versus Far experiments provide evidence that distractor construction affects transfer performance.

Critical weaknesses
  1. Limited methodological novelty and relatively modest improvements. The proposed framework primarily combines established multiple-choice supervision with semantically similar distractors and short answer targets. Both multiple-choice training and hard distractor construction have been extensively explored in prior work. The main contribution is their empirical comparison in embodied VLMs rather than a fundamentally new learning method. Moreover, the average improvement over Content supervision is only 1.16 points on Qwen3.5-4B, making the practical significance of the specific Letter formulation less clear.

  2. Incomplete isolation of the supervision mechanism. Although the authors carefully control training examples, the compared target formats differ substantially in sequence length and effective example weighting. Table 7 reports an average of 2 supervised tokens for Letter versus 8.76 for Content and 10.93 for Joint. Consequently, the observed advantage may partly reflect differences in loss weighting and optimization rather than the semantic benefits of compact choice supervision. Additionally, the Near versus Far comparison is conducted with only one training seed, and the evaluation focuses entirely on offline perception and reasoning rather than downstream robotic action performance.

Questions for authors
  1. Supervision length confounding. Have the authors considered normalizing the training loss independently for each example, so that longer answer targets do not receive greater weight? Would Letter supervision retain its advantage over Content and Joint under this control?

  2. Distractor difficulty. The Near versus Far experiments use only one training seed, and Appendix A reports that approximately 39% of examples retain identical distractors in both conditions. Can the authors provide additional evidence that the improvement is robust across seeds and directly attributable to distractor difficulty?

  3. Practical significance for embodied learning. The reported improvements are measured on offline VLM benchmarks. Is there evidence that multiple-choice supervision improves downstream robotic decision-making or manipulation performance? Without such evidence, how should the claimed benefits for embodied models be interpreted?

8 comments
AuthorsOpenReview ↗

Response to Reviewer gjsU (1/6) — Q1: Sample-normalized loss

Thank you for your review and suggestions. We address your questions below and believe that the additional evidence and clarifications resolve your concerns.

1. Does Letter remain ahead of Content and Joint after sample normalization?

Question: Different target lengths may give longer answers greater training weight. Can loss be normalized independently per example, and does Letter retain its advantage?

Response: We completed this control. Letter remains ahead of both Content and Joint in all three seeds, retaining the highest aggregate among the four supervision formats.

The submission reports average supervised lengths of 2, 8.76, and 10.93 tokens for Letter, Content, and Joint. To address length-dependent example weighting directly, we use:

Lsample=1N∑i=1N(1Ti∑t=1Tiℓi,t).\mathcal{L}_{\mathrm{sample}} =\frac{1}{N}\sum_{i=1}^{N} \left(\frac{1}{T_i}\sum_{t=1}^{T_i}\ell_{i,t}\right).

Here, TiT_i is the supervised-token count for example ii, and NN is the number of examples in the actual gradient-accumulation window. We average within each example before averaging across examples, preventing longer answers from receiving greater example weight solely because they contain more supervised tokens.

We trained each of Letter, Content, Joint, and Open with seeds 42, 43, and 44, for 12 runs. They share the base model, the same 10K questions, two training epochs, effective batch size 64, and evaluation of the final step-314 checkpoint. Letter, Content, and Joint share Near candidates; Open removes candidates and supervises the corresponding correct answer text.

Table 1. Qwen3.5-4B nine-benchmark aggregates with sample-normalized loss.

Supervision formatseed42seed43seed44Three-seed mean ± SD
Letter57.30857.47057.24357.340 ± 0.116
Content56.28254.97056.84356.032 ± 0.961
Joint57.24056.86656.82356.976 ± 0.229
Open55.80855.49054.41855.239 ± 0.728

Table 2. Paired advantages under the normalization control.

Paired comparisonseed42seed43seed44Mean differenceSeeds favoring Letter
Letter − Content+1.026+2.499+0.400+1.3093/3
Letter − Joint+0.068+0.604+0.420+0.3643/3
Letter − Open+1.500+1.979+2.825+2.1023/3

After normalization, Letter's average advantages are +1.309 points over Content, +0.364 over Joint, and +2.102 over Open. Both Letter–Content and Letter–Joint are positive in all three seeds. This directly answers your question: Letter's aggregate advantage remains after sample normalization. The original ordering withstands this additional control for example weighting.

AuthorsOpenReview ↗

Response to Reviewer gjsU (2/6) — Q2 (1/2): Cross-seed and cross-model robustness

2. Near–Far robustness, unchanged distractors, and distractor difficulty

Question: Near–Far uses only one seed, and approximately 39% of examples have unchanged distractors. Can robustness and the role of distractor difficulty be better supported?

Response: We added cross-seed, cross-model, and random-distractor comparisons, all supporting Near's training value, and audited candidate replacement across the 10K corpus.

2.1 Additional experiments reproduce Near's aggregate advantage

The original comparison uses Qwen3.5-4B seed 42 with Letter, Content, and Joint targets:

Table 3a. Near–Far results already reported in the submission, retaining its two-decimal precision.

Target (original Qwen, seed42)NearFarNear − Far
Letter57.2356.15+1.08
Content56.7254.68+2.04
Joint56.7154.95+1.76

We added Letter comparisons for Qwen seeds 43 and 44 and InternVL3.5-4B-Instruct seeds 42, 43, and 44. The original seed-42 result is reported separately rather than pooled with the new batch into a three-seed mean.

Table 3b. Additional paired Near–Far results with Letter supervision.

Model (additional Letter comparison)seedNearFarNear − Far
Qwen3.5-4B4357.47055.366+2.104
Qwen3.5-4B4457.24355.894+1.349
Qwen3.5-4BTwo-seed mean57.35655.630+1.726
InternVL3.5-4B-Instruct4244.22543.681+0.544
InternVL3.5-4B-Instruct4344.39543.518+0.877
InternVL3.5-4B-Instruct4444.24943.371+0.878
InternVL3.5-4B-InstructThree-seed mean44.29043.524+0.766

The new Qwen seeds improve by 2.104 and 1.349 points, and the InternVL seeds by 0.544, 0.877, and 0.878 points. Every additional paired comparison is positive across both families, providing repeated evidence for the finding.

The comparison fixes the question, visual input, correct answer, correct position, supervision target, and training budget. Qwen Near uses sample normalization and Far uses token normalization; every Letter target contains two supervised tokens, including the terminator, so the two normalizations are mathematically equivalent in this comparison.

AuthorsOpenReview ↗

Response to Reviewer gjsU (3/6) — Q2 (2/2): Unchanged and random distractors

2.2 Unchanged distractors reflect constraints on valid alternatives

Near and Far contain the same 10K questions. We replace distractors where valid different alternatives can be constructed and retain the remaining examples:

Table 3c. Candidate-replacement coverage in the complete training corpus.

Candidate changesQuestionsShare of training set
At least one distractor replaced6,09860.98%
Distractors unchanged3,90239.02%
Total10,000100%

Of the 6,098 changed examples, 6,017 replace all three distractors, 71 replace two, and 10 replace one, totaling 18,203 replaced incorrect options. Among the 3,902 unchanged examples, 3,874 have finite answer domains; the remaining 28 involve symmetrically retained source-annotation disputes or candidate overlap after review. Construction records are described in Appendix A of the submission.

The unchanged records are identical on both sides, preserving question count, training content, and capability coverage, with differences confined to the incorrect candidates of the other 6,098 questions. This intervention holds the complete training corpus fixed: changing incorrect candidates while retaining the correct answers and training content produces the repeatable transfer gains in Table 3b.

2.3 Random distractors provide additional evidence for candidate construction

We additionally sample incorrect candidates from the full answer pool while retaining questions, visual inputs, correct answers, and their positions.

Table 4. Additional three-seed comparisons of Near and random distractors with Letter targets.

Model (three seeds)Near meanRandom meanMean differenceseed42 differenceseed43 differenceseed44 differencePositive seeds
Qwen3.5-4B57.34055.360+1.981+2.985+1.149+1.8093/3
InternVL3.5-4B-Instruct44.29041.538+2.751+2.896+2.646+2.7123/3

Near's mean advantages over Random are +1.981 points for Qwen and +2.751 for InternVL, with positive differences in all three seeds of each model. Together with Near–Far, these results show that candidate construction influences transfer.

Near–Far is constructed according to confusability relative to visual evidence. Both sides preserve the question's answer domain and comparable answer granularity; Far distractors must clearly conflict with visual evidence and undergo structural checks and independent model review. The paired design and additional experiments jointly support the training value of constructing distractors according to semantic confusability.

Paired intervention, repeated training, and the random-distractor control together strengthen the paper's conclusion that confusable distractors have training value.

Reviewer_u8skOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes Embodied Choice Tuning, a SFT training framework that combines confusable candidate answers with letter-only targets to improve embodied reasoning abilities of VLMs. They construct a 10K training corpus, and compare different supervision formats to investigate candidate availability, distractor construction, and output targets. Across three model families and several embodied benchmarks, ECT achieves the highest aggregate performance among the tested formats.

Strengths

  • The paper addresses a practical question about supervision design. Matched comparisons help separate the effects of candidate inputs from those of output targets. Evaluation covers several embodied capabilities, demonstrating model effectiveness beyond letter-based answering.
  • The overall paper is relatively clear, including per-benchmark results, different training seeds, and scene/video-level bootstrap confidence intervals.

Weaknesses

  • Limited methodological novelty and analysis of task diversity. Although the corpus covers eight capability families, the central comparisons reformulate the same underlying questions into different supervision formats. The contribution primarily concerns format design; the benefits of capability coverage and task diversity are not clear.
  • The study evaluates domain-specific SFT on embodied benchmarks but does not examine retention of general visual understanding, reasoning, or instruction-following abilities. Only leveraging SFT may lead to decrease of general reasoning ability.
  • Comparisons mainly involve the same base models under different formats, without including strong embodied VLMs (e.g., RynnBrain, RoboBrain, HY-Embodied, etc). Moreover, additional model families and distractor interventions use only one training seed, while the Letter–Joint advantage is relatively small. Related work also lacks sufficient discussion with recent embodied VLM work.

Questions

  1. How sensitive are the findings to training-set size, capability composition, and task diversity?
  2. Does ECT preserve general reasoning and instruction-following performance after embodied SFT?
Critical strengths

see above

Critical weaknesses

see above

Questions for authors

see above

16 comments
AuthorsOpenReview ↗

Response to Reviewer u8sk (1/6) — Q1 (1/2): Training-set size and capability composition

Thank you for your review and suggestions. We address your questions below and believe that the additional evidence and clarifications resolve your concerns.

1. Q1: Training-set size, capability composition, and task diversity

Q1: How sensitive are the findings to training-set size, capability composition, and task diversity?

Response: We provide complementary evidence from different corpus configurations, matched supervision comparisons, and transfer to diverse tasks, including additional closed-loop execution experiments.

1. Training-set size: gains on an already capable embodied model

We evaluate ECT with approximately 3.8K and 10K examples on a strong proprietary embodied foundation model used within our company. The scale study shares a common backbone and evaluation protocol; the aggregate equally averages the nine native benchmark scores.

MetricProprietary BaseECT: 3.8KECT: 10K
Cosmos65.68665.49067.451
EmbSpatial84.50584.06682.940
VSIBench62.86062.21862.151
Where2Place74.76274.04876.905
ERQA41.50745.77947.348
VABenchPoint70.33370.33370.667
OpenEQA61.33362.19361.831
RoboVQA71.10971.29271.729
UniEQA43.80344.54344.977
Aggregate63.98964.44065.111
Gain over Base—+0.452+1.122

Means and differences are computed before rounding.

Only 3,821 adaptation examples, with more limited fine-grained coverage, raise the aggregate from 63.989 to 64.440 (+0.452). This aggregate gain is obtained on top of an already capable embodied model, demonstrating practical incremental value with limited additional data. The 10K configuration reaches 65.111 (+1.122). Thus, ECT is already effective with the smaller configuration, while the larger, more comprehensive corpus achieves a higher aggregate: 10K > 3.8K > Base.

2. Capability composition: different corpus configurations and matched comparisons

Both corpora cover eight broad capability families. The 3.8K subset has 82 question types and 89 capability–type combinations; the 10K corpus has 102 types and 109 combinations, with different capability proportions. Positive aggregate gains under both configurations provide supplementary evidence for ECT's applicability across corpus sizes, compositions, and fine-grained coverage.

In the main study, Letter, Content, Joint, and Open share the same 10K examples, visual observations, questions, correct answers, capability composition, and type coverage. Each capability family accounts for approximately 10%–15% of the corpus. Letter's relative advantage therefore cannot be attributed to receiving more data, a more favorable capability mixture, or different task coverage. Corpus organization supplies broad capabilities; matched comparisons identify the roles of candidate inputs and output targets under shared content. These experiments support the practical applicability and supervision-design principles of ECT from complementary perspectives.

AuthorsOpenReview ↗

Response to Reviewer u8sk (2/6) — Q1 (2/2): Task diversity and additional embodied execution

3. Task diversity: training coverage, native interfaces, and additional embodied execution

Matched questions control content differences within a diverse corpus. The 10K corpus draws from 19 sources and spans eight embodied capability families and their fine-grained task types. Reformulating each example into four supervision formats preserves this shared task coverage while making supervision factors comparable.

All nine benchmarks retain their native instructions, response requirements, and scoring. Models must select answers, estimate numbers, predict coordinates, and generate open-ended text. Examples include numerical questions in VSIBench, coordinate outputs in Where2Place and VABenchPoint, and text generation in OpenEQA, RoboVQA, and UniEQA. Evaluation inputs and scoring are fixed across model conditions, directly testing transfer across tasks and response interfaces beyond the letter target used in training.

RoboVQA provides concrete robot-related downstream QA evidence. In separate public-backbone experiments with sample-normalized loss, Letter's three-seed mean is 64.914, compared with 62.027 for Open (+2.887 points). Planning QA, affordance generation, and future prediction improve by 5.500, 7.365, and 5.934 points, respectively, with positive gains in all three seeds of each task. These tasks require generated text, demonstrating transfer to different robot-related answering tasks.

We additionally conducted closed-loop embodied task-execution experiments in simulation: navigation in EB-Navigation and multi-step manipulation in EB-ALFRED. Using a public backbone, the model selects actions or skills from visual observations; the executor acts in the environment, changes its state, and returns new visual feedback for subsequent decisions. We measure actual task-completion success, directly evaluating embodied execution benefits.

Base, Open, Content, Joint, and ECT-Letter share task sets, prompts, visual budgets, generation settings, and action or skill executors. Existing weights connect directly to the execution system without simulation-specific training. Each round covers 60 Navigation tasks and 50 ALFRED tasks. Fine-tuned formats use training seeds 42, 43, and 44; Base uses identical weights in three evaluation rounds. The five settings and three rounds yield 1,650 task executions:

ModelNavigation mean success rateALFRED mean success rate
Base58.89%30.67%
Open60.56%32.00%
Content63.89%36.67%
Joint62.78%35.33%
ECT-Letter67.78%40.00%

ECT-Letter achieves 67.78% navigation success and 40.00% manipulation success, gaining 8.89/9.33 percentage points over Base and 7.22/8.00 over same-data Open. Its mean also exceeds Content and Joint in both environments, and it ranks first among the five settings in all six environment-by-seed comparisons.

These additional executed tasks strengthen the task-diversity evidence: navigation and multi-step manipulation involve different action or skill interfaces, goals, and completion criteria, requiring continuous perception, decision-making, and execution. Together with broad training coverage and native-format evaluation, they demonstrate transfer across diverse embodied tasks. ECT's supervision design has also been adopted in our company's embodied brain system, providing a concrete application example.

AuthorsOpenReview ↗

Response to Reviewer u8sk (3/6) — Q2: General-capability retention

2. Q2: Retention of general visual understanding, reasoning, and instruction following

Comment: Domain SFT may affect general capabilities, particularly general reasoning and instruction following.

Response: We completed independent evaluations. The three evaluated categories of general capability are broadly retained: mathematical reasoning remains close to the base, instruction following improves, and Letter outperforms same-data Open tuning on all three evaluations in every seed.

We evaluated the Qwen3.5-4B base and sample-normalized Letter, Content, Joint, and Open models with seeds 42, 43, and 44: 13 models on MMMU validation (900 questions), GSM8K test (1,319 questions), and IFEval (541 prompts), totaling 35,880 responses. Corresponding evaluation protocols are shared across models.

Table 1. General-capability evaluations. Fine-tuned rows report three-seed means ± sample standard deviations.

Model / targetMMMUGSM8K strict EMIFEval strict promptIFEval strict instruction
Base62.3389.0881.3386.93
Letter56.22 ± 0.6288.30 ± 0.2982.99 ± 0.6788.13 ± 0.42
Content45.74 ± 4.0687.29 ± 0.8882.38 ± 0.5387.57 ± 0.35
Joint56.11 ± 0.5887.74 ± 0.6881.70 ± 1.1587.29 ± 0.98
Open54.19 ± 0.6386.81 ± 0.8679.98 ± 0.5686.01 ± 0.48
Letter − Base-6.11-0.78+1.66+1.20

GSM8K uses 8-shot chain-of-thought prompting and strict exact match. IFEval reports strict prompt-level and instruction-level scores. Bold Letter values indicate the highest means among the four fine-tuning formats.

General reasoning. Letter scores 88.30 on GSM8K, close to the base's 89.08 (a −0.78-point difference), while exceeding Open by 1.49 points. Mathematical reasoning remains strong after embodied Letter tuning.

Instruction following. Letter improves strict prompt-level IFEval from 81.33 to 82.99 and strict instruction-level IFEval from 86.93 to 88.13, gains of 1.66 and 1.20 points, respectively. Both metrics exceed the base in all three seeds. Training on letters does not remove the ability to generate responses under general instructions; these direct evaluations show improved instruction following.

General visual understanding. Letter scores 56.22 on MMMU versus 62.33 for the base. Although this is a 6.11-point decrease, it remains above Open's 54.19 and is the highest mean among the four fine-tuning formats. Together with the reasoning and instruction-following results, ECT shows useful general-capability retention while obtaining embodied improvements.

For a direct comparison of supervision designs, we report per-seed differences against same-data Open tuning:

Table 2. Letter advantages over Open on general-capability evaluations.

Metricseed42: L−Oseed43: L−Oseed44: L−OMean differenceSeeds favoring Letter
MMMU+1.22+2.89+2.00+2.043/3
GSM8K strict EM+2.81+0.91+0.76+1.493/3
IFEval strict prompt+2.96+1.85+4.25+3.023/3
IFEval strict instruction+1.80+1.44+3.12+2.123/3

Letter exceeds Open on visual understanding, mathematical reasoning, and instruction following in all three seeds, by mean margins of +2.04, +1.49, and +3.02 points; the instruction-level margin is +2.12 points. These results show that compact choice supervision improves aggregate embodied performance while also favoring retention on the evaluated general tasks.

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

This paper investigates how different supervised fine-tuning formats affect the transfer of embodied vision-language capabilities. The authors propose Embodied Choice Tuning (ECT), which presents multiple-choice questions with semantically confusable distractors and supervises the correct option using a compact letter target.

The paper constructs a dataset of 10,000 examples covering eight embodied capability categories and 102 question types. Four supervision formats are compared under matched training data, with additional experiments investigating the importance of distractor design. Evaluation across nine embodied benchmarks and three model families shows that the proposed approach consistently achieves the highest aggregate scores among the tested supervision formats. For Qwen3.5-4B, ECT improves the aggregate score by 3.89 points over the base model and 2.27 points over open-ended supervision.

Critical strengths
  1. A well-controlled study of supervision formats. The paper carefully separates several factors that are often conflated when comparing multiple-choice and open-ended training, including candidate availability, distractor semantics, and target representation. In particular, comparing Content with Open while keeping the supervised answer text fixed provides useful evidence about the contribution of candidate inputs. This is a meaningful empirical contribution to understanding embodied VLM fine-tuning.

  2. Broad evaluation with reasonably consistent results. The experiments cover nine benchmarks with different evaluation formats and include three model families. The results show that the benefits of multiple-choice supervision are not restricted to multiple-choice evaluation. The additional distractor experiments also provide evidence that the semantic quality of incorrect alternatives matters during training. These observations are useful for designing data-efficient embodied VLM training pipelines.

Critical weaknesses
  1. The methodological novelty is relatively limited, and the explanation of the letter-target advantage is not fully isolated. The proposed training approach mainly combines existing multiple-choice supervision with carefully designed distractors. While the controlled experiments are useful, the performance advantage of letter-only supervision may partly arise from differences in target length, token-level loss weighting, or optimization behavior rather than the semantic properties of the target format itself. The paper acknowledges these differences but does not fully disentangle them. Moreover, the improvements over other choice-based formats are relatively modest.

  2. The relationship between the reported benchmark improvements and practical embodied intelligence remains unclear. The evaluation focuses on offline perception, reasoning, and question answering rather than actual robot action generation or closed-loop task execution. It is therefore difficult to determine whether the improvements translate into better task planning, action selection, or manipulation performance. Additionally, the Near-versus-Far comparison is primarily conducted with one model and one training seed, and the aggregate improvement for Letter is relatively small, making the robustness of this particular conclusion less certain.

Questions for authors
  1. Can the authors perform a more controlled comparison between letter-only and answer-text supervision by normalizing the loss contribution per training example or matching the effective supervised-token budget? This would help determine whether the improvement comes from the target representation itself or from optimization and weighting differences.

  2. How robust are the Near-versus-Far results across training seeds and model families? The reported Letter improvement is relatively modest, so additional evidence would strengthen the conclusion regarding distractor quality.

  3. Do the improvements in embodied reasoning benchmarks translate to downstream robot action prediction or manipulation tasks? Even a limited evaluation of action-related performance would help establish the practical significance of the proposed fine-tuning strategy.

7 comments
AuthorsOpenReview ↗

Response to Reviewer vtjK (1/6) — Q1: Sample-normalized loss

Thank you for your review and suggestions. We address your questions below and believe that the additional evidence and clarifications resolve your concerns.

1. Does the Letter advantage remain with sample-normalized loss?

Question: Can sample-normalized loss, or a matched effective supervised-token budget, provide a more controlled comparison of Letter and answer-text supervision and examine the effects of loss weighting and optimization?

Response: We implemented the proposed sample-normalization control, and Letter retains its overall advantage.

We first average the loss over the supervised tokens of each example, then average across the examples in the actual gradient-accumulation window:

Lsample=1N∑i=1N(1Ti∑t=1Tiℓi,t).\mathcal{L}_{\mathrm{sample}} =\frac{1}{N}\sum_{i=1}^{N} \left(\frac{1}{T_i}\sum_{t=1}^{T_i}\ell_{i,t}\right).

Here, TiT_i is the number of supervised tokens in example ii, and NN is the number of examples in that accumulation window. Longer answers therefore do not receive greater example weight simply because they contain more supervised tokens.

We trained Letter, Content, Joint, and Open with seeds 42, 43, and 44, for 12 runs. All runs use the same 10K questions, a shared base model, two training epochs, an effective batch size of 64, and the final checkpoint at step 314. The three choice formats share the same Near candidates; Open removes the candidates while retaining the corresponding correct answer text.

Table 1. Qwen3.5-4B nine-benchmark aggregate with sample-normalized loss.

Supervision formatseed42seed43seed44Three-seed mean ± SD
Letter57.30857.47057.24357.340 ± 0.116
Content56.28254.97056.84356.032 ± 0.961
Joint57.24056.86656.82356.976 ± 0.229
Open55.80855.49054.41855.239 ± 0.728

Table 2. Paired differences within each seed.

Comparisonseed42seed43seed44Mean differenceSeeds favoring Letter
Letter − Content+1.026+2.499+0.400+1.3093/3
Letter − Joint+0.068+0.604+0.420+0.3643/3
Letter − Open+1.500+1.979+2.825+2.1023/3

Letter achieves the highest aggregate among the four formats in all three seeds. Its average advantages over Content, Joint, and Open are +1.309, +0.364, and +2.102 points, respectively. Thus, the paper's finding that Letter provides the strongest overall performance among the tested supervision formats remains supported under the additional normalization control.

This directly addresses target-length-dependent example weighting: after averaging loss per example, Letter remains ahead of Content, Joint, and Open in all three seeds, further supporting the supervision-format comparison in the submission.

AuthorsOpenReview ↗

Response to Reviewer vtjK (2/6) — Q2: Cross-seed and cross-model robustness

2. Robustness of Near–Far across training seeds and model families

Question: The original Near–Far comparison mainly uses one model and one training seed. Can the Letter improvement be reproduced?

Response: The additional experiments reproduce Near's overall advantage in every new matched seed across two model families.

The original comparison in Section 5.4 uses Qwen3.5-4B with seed 42 and compares Letter, Content, and Joint targets:

Table 3. Near–Far results already reported in the submission.

Target (original Qwen, seed42)NearFarNear − Far
Letter57.2356.15+1.08
Content56.7254.68+2.04
Joint56.7154.95+1.76

To examine training randomness and model-family dependence, we added Letter comparisons for Qwen seeds 43 and 44 and InternVL3.5-4B-Instruct seeds 42, 43, and 44. We report the original seed-42 result separately from the new batch rather than combining them into a three-seed mean.

Table 4. Additional paired Near–Far results with Letter supervision.

Model (additional Letter comparison)seedNearFarNear − Far
Qwen3.5-4B4357.47055.366+2.104
Qwen3.5-4B4457.24355.894+1.349
Qwen3.5-4BTwo-seed mean57.35655.630+1.726
InternVL3.5-4B-Instruct4244.22543.681+0.544
InternVL3.5-4B-Instruct4344.39543.518+0.877
InternVL3.5-4B-Instruct4444.24943.371+0.878
InternVL3.5-4B-InstructThree-seed mean44.29043.524+0.766

Near improves over Far by 2.104 and 1.349 points in the two additional Qwen seeds and by 0.544, 0.877, and 0.878 points in the three InternVL seeds. These results extend the original observation to repeated comparisons across two model families, supporting the reproducible training value of ECT's confusable candidates.

The comparisons hold questions, visual inputs, correct answers, and correct answer positions fixed while intervening on incorrect candidates. Qwen Near uses sample-normalized loss and Far uses token-normalized loss; for this Letter comparison, every target contains two supervised tokens, including the terminator, making the two normalizations mathematically equivalent.

The repeated aggregate advantage across seeds and families directly strengthens the paper's evidence for the role of distractor construction.

AuthorsOpenReview ↗

Response to Reviewer vtjK (3/6) — Q3 (1/2): Closed-loop simulation

3. Embodied benefits: closed-loop simulation, practical adoption, and Arena downstream tasks

Question: Do gains on question-answering benchmarks transfer to downstream action prediction or manipulation? Even limited action-related evaluation would help establish practical significance.

Response: New simulation results directly show improved navigation and multi-step manipulation success. Adoption in our company's embodied brain system and action-related downstream tasks in Arena further establish practical value.

First, new closed-loop simulation experiments directly evaluate navigation and multi-step manipulation benefits. On EB-Navigation and EB-ALFRED from EmbodiedBench, we evaluate Base, Open, Content, Joint, and ECT-Letter with a common Qwen3.5-4B backbone. We directly connect the existing checkpoints from training seeds 42, 43, and 44 for each fine-tuning format to the same action or skill execution system, without additional training on the simulation tasks. All groups share the task sets, evaluation prompts, visual budget, and generation settings. Navigation retains its eight native actions, 20-step limit, and success criterion of reaching within one meter of the target. ALFRED retains its fixed skill executor, 30-skill-step limit, and native goal-completion criterion.

Each evaluation covers all 60 Navigation tasks and all 50 ALFRED tasks, yielding 1,650 task executions across five settings and three evaluation rounds. Base loads no fine-tuning adapter and uses the same backbone weights in all three repeated evaluations. The table reports mean success rates across the three rounds.

Table 5a. Mean task success rates in the two closed-loop simulation environments.

ModelNavigation mean success rateALFRED mean success rate
Base58.89%30.67%
Open60.56%32.00%
Content63.89%36.67%
Joint62.78%35.33%
ECT-Letter67.78%40.00%

Table 5b. Mean success-rate gains of ECT-Letter over each control.

ComparisonNavigation gain (percentage points)ALFRED gain (percentage points)
Letter − Base+8.89+9.33
Letter − Open+7.22+8.00
Letter − Content+3.89+3.33
Letter − Joint+5.00+4.67

ECT-Letter reaches 67.78% and 40.00% on Navigation and ALFRED, improving over Base by 8.89 and 9.33 percentage points, and over same-data Open supervision by 7.22 and 8.00 percentage points. Its mean also exceeds Content and Joint in both environments. Letter achieves the highest success rate among all five settings in each of the six environment-by-seed comparisons. These results show that the capabilities learned through ECT transfer to action selection under continuous visual feedback and to multi-step task execution, further supporting the practical training value of Letter supervision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all three reviewers for their constructive feedback. In response, we provide additional controlled experiments, robustness analyses, and closed-loop simulation results, and clarify ECT’s contribution and practical value. The main additions and clarifications are summarized below:

  1. Research positioning and original contribution. We clarify ECT as a capability-oriented training paradigm that separates candidate inputs, distractor construction, and supervision targets under shared training content. Objective-level analysis and cross-task evaluation establish actionable training principles. We also explain its complementary relationship to existing embodied foundation-model efforts and how matched comparisons across Qwen, InternVL, and other backbones directly test supervision-design effectiveness.

  2. Supervision controls and robustness. We added sample-normalized training for all four formats with three seeds. Letter retains the highest nine-benchmark aggregate in every seed, with mean advantages of +1.309 over Content and +0.364 over Joint. We also extended Near–Far comparisons to additional Qwen seeds and three InternVL seeds, and added random-distractor controls. These results strengthen the evidence for compact targets and confusable candidate construction.

  3. Data scale, capability coverage, and practical adoption. ECT is already used in our company to train embodied models for real business applications. The results below are obtained on a proprietary backbone developed for these applications, which already has strong embodied capabilities. ECT provides consistent aggregate gains at both evaluated data scales: +0.452 with 3.8K examples and +1.122 with 10K examples. Both corpora cover eight capability families, while question-type coverage expands from 82 to 102 types and capability proportions also differ. These results demonstrate the practical value of further improving an already capable business model with limited adaptation data, alongside applicability across corpus configurations.

  4. General-capability evaluation. We added MMMU, GSM8K, and IFEval evaluations covering 13 models and 35,880 responses. Mathematical reasoning remains close to the base, while instruction following improves. Letter outperforms same-data Open tuning on general visual understanding, mathematical reasoning, and instruction following in every seed, providing direct evidence for comparing general-capability retention across supervision designs.

  5. Additional closed-loop embodied execution. We added five-setting, three-round evaluations in EB-Navigation and EB-ALFRED, totaling 1,650 simulated task executions. ECT-Letter achieves 67.78% navigation success and 40.00% multi-step manipulation success, improving over same-data Open supervision by 7.22 and 8.00 percentage points. These experiments directly evaluate decisions and execution under continuous visual feedback. Robot-related downstream tasks in Arena and adoption in our company’s embodied brain system provide complementary evidence.

We appreciate the reviewers’ suggestions. We believe these additions, together with our individual responses, address the main concerns and strengthen the empirical basis for ECT as a practical embodied training paradigm. We welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.