acceptodds
Under review as a conference paper at ICLR 2027

ThoughtVLA: Evolving Thought Trajectories for Continual Vision-Language-Action Adaptation

Abstract

Continual adaptation of pretrained vision-language-action (VLA) models aims to acquire new manipulation skills from sequentially arriving demonstrations while retaining performance on previously learned tasks. However, in progressive LoRA-MoE methods, updating the router can change the expert mixtures used for previously learned tasks even when their experts remain frozen. Meanwhile, pretrained semantic similarity alone may not reliably distinguish learned skills that are semantically related but require different manipulations. To address these challenges, we propose ThoughtVLA, a continual adaptation framework built around Continually Evolving Thought Trajectory (CETT), which formalizes how pre-action reasoning for skill selection evolves as task knowledge accumulates. In ThoughtVLA, Evolving Skill Reasoning selects a learned skill by interpreting the current input with pretrained semantics and accumulated task knowledge, and resolving ambiguity with task attributes. Each acquired skill retains a LoRA composition, which is used throughout execution when that skill is selected. Beyond skill reasoning, Complementary Knowledge Transfer supports new-skill acquisition by selecting prior LoRA updates that reduce the residual error left beyond the limited rank of new adapter and incorporating them into the new skill’s LoRA composition. Across single-arm, bimanual, humanoid, mixed-embodiment, cross-backbone, and real-robot evaluations, ThoughtVLA achieves higher final success and lower forgetting than the compared methods.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual robot learning, Continual learning, Vision-language-action model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hSqqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ThoughtVLA, a framework for continual adaptation of pretrained vision-language-action models. It introduces Continually Evolving Thought Trajectories (CETT), implemented through Evolving Skill Reasoning (ESR) for pre-action skill identification and Complementary Knowledge Transfer (CKT) for reusing previous LoRA updates. The method separates skill selection from execution and attempts to prevent changes in expert routing from interfering with previously acquired skills. Evaluation covers multiple LIBERO streams, RoboTwin, humanoid manipulation, mixed embodiments, different VLA backbones, and real-robot experiments.

Critical strengths
  1. The paper addresses a meaningful failure mode in continual VLA adaptation: previously learned skill policies can remain frozen while their effective behavior changes because of routing decisions. Separating skill identification from execution is a well-motivated design. The strong results across different embodiments and the real-robot evaluation make this contribution consequential.

  2. The two principal components are supported by targeted ablations. On LIBERO-Long, ESR improves final success from 52.8% to 70.2% when conditional refinement is enabled, while CKT further improves it to 81.0%. The comparison with alternative knowledge-transfer criteria helps substantiate the proposed low-rank complementarity principle.

Critical weaknesses

W1. Limited conceptual novelty of Continually Evolving Thought Trajectories.

Although CETT is presented as an evolving reasoning mechanism, its implementation largely relies on frozen semantic embeddings, accumulated feature statistics, instruction matching, and predefined attribute-based score refinement. The reasoning rules themselves remain fixed throughout continual learning. Consequently, it is unclear whether CETT constitutes a substantially new reasoning paradigm or a structured, knowledge-augmented skill retrieval mechanism. A stronger comparison against prototype-based retrieval, discriminative skill classification, and attribute-aware routing would help isolate its conceptual contribution.

W2. Strong assumptions about task structure and skill availability.

The continual-learning protocol assumes known stage boundaries, explicitly grouped demonstrations, and an identifiable skill associated with each training task. At evaluation, the model selects among previously acquired skills rather than learning to recognize and compose arbitrary new behaviors. These assumptions substantially simplify continual adaptation compared with real-world scenarios involving unsegmented experience, overlapping skills, and evolving task definitions. The generality of the proposed framework under less structured task arrivals remains unclear.

W3. Limited support for within-episode skill switching and compositional execution.

ESR selects a skill before execution, and its corresponding LoRA composition is retained throughout the episode. This design may be insufficient for long-horizon tasks requiring different manipulation primitives at different execution stages. For example, a task involving opening a container, grasping an object, and performing a precise placement may benefit from dynamically switching between specialized skills. The paper does not establish whether episode-level skill selection can handle such compositional requirements as effectively as more flexible routing mechanisms.

W4. The additional contribution of the CKT criterion is relatively modest compared with a simpler residual baseline.

The CKT ablation on LIBERO-Long reports 81.0% final success, compared with 79.4% for full-residual transfer. Although the proposed rank-aware criterion is theoretically motivated, its incremental improvement over this simpler alternative is only 1.6 percentage points. The paper would benefit from uncertainty estimates, sensitivity to LoRA rank, and an analysis of when the rank-aware score selects a different source skill and whether those differences consistently improve execution.

W5. The theoretical approximation underlying CKT may not reflect actual nonlinear adaptation dynamics.

CKT estimates transfer complementarity using gradients and compressed activations evaluated at frozen base weights. Its low-rank approximation argument characterizes a local response rather than the full nonlinear optimization process. Moreover, the candidate-selection procedure validates only the highest-scoring proposal and does not systematically explore alternative candidates if confirmation fails. The robustness of this approximation across different adaptation regimes and increasing skill-library sizes remains insufficiently characterized.

W6. Limited evidence for scalability to substantially longer continual-learning streams.

The largest evaluated mixed-embodiment stream contains 55 skills. Both the retained knowledge and the LoRA composition library grow as additional skills are acquired. Although the reported storage and parameter overheads are favorable, the paper does not investigate substantially longer streams in which semantically overlapping skills accumulate. Such settings may introduce increasing ambiguity, retrieval errors, and interference between closely related skill representations.

Questions for authors
  1. How much of ESR's improvement depends on its evolving knowledge representation rather than simply replacing learned expert routing with a strong frozen-encoder prototype or attribute-aware skill classifier under a matched information budget?

  2. How robust is skill reasoning when task descriptions are ambiguous, contain contradictory attributes, or depart substantially from the training instruction distribution? Could the authors provide uncertainty estimates for the real-robot results?

  3. See the above weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer hSqq

W1: Conceptual contribution of CETT relative to skill retrieval.

A1: Thank you for this thoughtful question. As defined in Sec. 1 at l. 72–82, thought in CETT is a structured pre-action process that interprets the task against acquired skills and refines ambiguous interpretations before identifying a skill. It goes beyond skill retrieval in three ways:

  1. Conditional refinement. Attribute evidence is consulted only when related skills remain ambiguous after semantic and instruction matching, adding 8.7 SR acc. and 8.6 FSR points in Table 7.
  2. Evolving knowledge. Accumulated statistics recalibrate semantic comparison and instruction matching as skills accumulate, adding 4.8 / 5.2 FSR points over frozen statistics in Table 6.
  3. Decoupled execution. Later stages cannot alter earlier skills' compositions, which avoids the routing drift of progressive LoRA-MoE.

Table R9 compares ESR with prototype retrieval and discriminative classification on the same CKT-trained policies and frozen encoder; its first row, ESR's own semantic score, is plain prototype retrieval.

Table R9. RoboTwin, Table 7 protocol: SR acc. on RoboTwin's unseen evaluation instructions for all 50 tasks after the final stage; FSR is closed-loop.

CategorySelectorEvidenceSR acc. (%)FSR (%)
Prototype-based retrievalSemantic prototype c0per-skill mean VLM features58.228.5
Prototype-based retrievalHiDe-LLaVA [2]per-skill image and text anchors64.236.1
Discriminative classificationRanPAC [3]random projection, ridge classifier54.726.4
ESR ablationESR w/o attributes c1no attribute refinement82.344.2
ESR ablationESR w/o gatingrefinement always applied80.541.3
OursFull ESRfull evidence, updated statistics91.052.8

As shown in Table R9, ESR leads prototype retrieval and discriminative classification on the same policies by at least 26.8 pp in SR acc. and 16.7 pp in FSR. Removing attribute refinement or its gating lowers SR acc. to 82.3% and 80.5%, so the gain comes from conditional reasoning over attributes rather than attribute evidence alone. We have added Table R9 to Table 7 and to Appendix C.7 as Table A21.

[2] Guo et al. HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model. ACL 2025. (https://arxiv.org/abs/2503.12941)

[3] McDonnell et al. RanPAC: Random Projections and Pre-trained Models for Continual Learning. NeurIPS 2023. (https://arxiv.org/abs/2307.02251)

W2: Strong assumptions on task structure and skill availability.

A2: Thank you for this question.

Our setting. ThoughtVLA learns tasks over a sequence of stages from current-stage demonstrations only and keeps no raw trajectories. After each stage, it is evaluated on all tasks seen so far without oracle skill or stage IDs, so it must identify the required skill itself before acting.

Relation to prior work. Our protocol follows recent continual VLA work such as Liu et al. [1], CLARE [4] and Stellar VLA [5], which evaluate mainly on LIBERO; our validation extends it in three respects:

  • More embodiments. Beyond four single-arm LIBERO streams, we evaluate 50 bimanual RoboTwin and 24 GR-1 humanoid tasks.
  • Long cross-embodiment stream. One mixed stream learns 55 single-arm, bimanual and humanoid skills over 51 stages.
  • Second backbone. The four LIBERO streams are repeated with Qwen3.5-2B and an OpenVLA-OFT-style head, as detailed in Appendix C.1.

Less structured arrivals. We rerun LIBERO-Long with irregular stages of 3, 1, 4 and 2 tasks and no task labels within a stage: a grouping step clusters each stage's demonstrations into skills with ESR's instruction-matching features, while stage boundaries remain given. All methods use 110K formal steps; SkillsCrafter and DMPEL are the two strongest baselines on LIBERO-Long in Table 1.

Table R10. LIBERO-Long with irregular stages. Stage columns give final success in % on the tasks introduced at that stage; FSR, CSR and Forg. follow Table 1.

MethodStage 1, 3 tasksStage 2, 1 taskStage 3, 4 tasksStage 4, 2 tasksFSRCSRForg.
SkillsCrafter52.766.078.061.065.871.414.4
DMPEL57.360.070.042.059.668.917.3
ThoughtVLA86.782.083.565.080.681.71.3
ThoughtVLA with standard stages, Table 1————81.083.41.4

As shown in Table R10, ThoughtVLA reaches 80.6% FSR without task labels, within 0.4 pp of standard stages, with 1.3 pp forgetting, and leads SkillsCrafter and DMPEL at every stage and by 14.8 and 21.0 pp in FSR, while their forgetting reaches 14.4 and 17.3 pp. We have added this experiment to Appendix C.4 with Table A17.

AuthorsOpenReview ↗

Response to Reviewer hSqq

W3: Within-episode skill switching and compositional execution.

A3: Thank you for the insightful comment. Our streams already contain such multi-stage tasks, each learned from full demonstrations. Table R11 selects them by a fixed rule: all ten LIBERO-Long tasks; GR-1 tasks 1–6, which place an item into a cabinet, drawer or microwave and close it; RoboTwin tasks whose instructions contain "then"; and the real-robot multi-step tasks 17–20.

Table R11. Final success in % on multi-stage task groups, pooling in-domain and out-of-domain rollouts for the real robot; the baseline is the strongest by overall FSR in the corresponding table.

Task groupTasksThoughtVLAStrongest baselineBaseline
LIBERO-Long, all tasks1081.0SkillsCrafter, Table 167.6
GR-1 tasks 1–6640.8DMPEL, Table 328.0
RoboTwin 11, 26, 33, 42, 45548.4AtomicVLA, Table 325.0
Real robot 17–20423.8SkillsCrafter, Table 516.3

As shown in Table R11, ThoughtVLA outperforms the strongest baseline on every group, including AtomicVLA on RoboTwin, 48.4% vs. 25.0%, although AtomicVLA chains atomic skills within each episode. Since each skill is learned from full demonstrations and the action head responds to each observation, execution spans subtask boundaries within the selected policy. Switching between separately acquired skills within an episode is a natural extension, which we add to Appendix D. We have added Table R11 to Appendix C.4 as Table A18.

[1] Liu et al. Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning. arXiv:2603.03818, 2026. (https://arxiv.org/abs/2603.03818)

[4] Römer et al. CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion. RA-L 2026. (https://arxiv.org/abs/2601.09512)

[5] Wu et al. Continually Evolving Skill Knowledge in Vision Language Action Model. arXiv:2511.18085, 2026. (https://arxiv.org/abs/2511.18085)

W4: Uncertainty, rank sensitivity and source selection of CKT versus full-residual transfer.

A4: Thank you for this careful comment. As described in Appendix C.6, the two strategies differ only in the ranking score and the Δr > 0 gate.

Uncertainty. Table R12 adds training seeds 43 and 44 for full-residual transfer; all runs share the 500 initial states of evaluation seed 7, so the comparison is paired.

Table R12. Paired LIBERO-Long comparison at rank 16, with published values from Tables 8 and A15.

MethodSeed 42Seed 43Seed 44Mean
Full residual FSR (%)79.478.678.878.9
Rank-aware CKT FSR (%)81.081.080.680.9
Paired difference (pp)+1.6+2.4+1.8+1.9

As shown in Table R12, CKT outperforms full residual under every seed, by 1.9 pp on average, more than the seed-to-seed spread of either method, 0.4 and 0.8 pp.

When the criteria select different sources. Full residual ranks candidates by the total error after transfer, ‖R‖²_F, while Δr uses ε_r(R), the error left after the best rank-r new update. CKT can thus prefer a larger but low-rank residual, which the new update can absorb, over a smaller but spread-out one: in Eq. A8 of Appendix A, full residual prefers B since 4.5 < 25.25, but CKT prefers A since ε₁ = 0.25 < 2.25. CKT also proposes transfer only when the maximum Δr is positive, whereas full residual always proposes its top candidate.

Whether the differences improve execution. They improve it consistently: besides leading under every seed in Table R12, CKT raises CSR, which averages performance over all stages, by 2.2 pp in Table 8 and lowers forgetting from 1.7 to 1.4 pp.

Rank sensitivity. Full residual is the r = 0 case of Δr, since ε₀(R) = ‖R‖²_F by Appendix A, so the two ranking scores differ only through the rank budget. Table R13 repeats the Table 8 comparison with new-update ranks 8 and 32, sharing the stage-1 checkpoint and all other settings; Δr uses the adapter's rank and the scaling factor is 2r.

Table R13. LIBERO-Long results in % for different new-update ranks; rank 16 is from Table 8.

RankMethodFSR ↑CSR ↑
8Full residual69.470.9
8Rank-aware CKT72.873.7
16Full residual79.481.2
16Rank-aware CKT81.083.4
32Full residual80.081.8
32Rank-aware CKT81.483.5

As shown in Table R13, CKT improves on full residual at every rank, by 3.4, 1.6 and 1.4 pp in FSR and 2.8, 2.2 and 1.7 pp in CSR at ranks 8, 16 and 32, so the advantage is not specific to rank 16. The gain is largest at rank 8, where the rank budget is tightest, and narrows as a higher-rank update absorbs more of the residual itself. We have added Tables R12 and R13 to Appendix C.8 as Tables A25 and A26.

AuthorsOpenReview ↗

Response to Reviewer hSqq

W5: Local approximation versus nonlinear adaptation, and rejected proposals.

A5: Thank you for this insightful comment. We agree that the low-rank analysis characterizes the local response at the frozen base weights, not the full nonlinear optimization, so CKT uses it only to rank candidates and lets actual adaptation decide.

First, Δr ranks all candidates from gradients and compressed activations without training an adapter per candidate, keeping scoring affordable as Table R2 shows.

Second, the top-ranked proposal is tested by paired adaptation trials, which run the actual nonlinear fine-tuning with and without transfer, and is adopted only if the held-out action-loss reduction meets the threshold and more than half of the trajectories improve, following Eq. A11. We have added this clarification to Appendix A.1 at l. 837–840.

W6: Scalability to substantially longer streams with overlapping skills.

A6: Thank you for this constructive comment.

First, on stream length, our RoboTwin stream learns 50 bimanual tasks and our mixed-embodiment stream 55 single-arm, bimanual and humanoid skills over 51 stages, as listed in Table A1; for reference, the longest streams in Liu et al. [1], Stellar VLA [5] and CLARE [4] contain 10, 30 and 40 single-arm LIBERO tasks. On the 55-skill stream, ThoughtVLA reaches 67.3% FSR with 3.4 pp forgetting, versus 51.5% and 5.7 pp for AtomicVLA, the strongest baseline in Table 3. Memory also stays compact: unlike the replay of [1] and [5], ThoughtVLA stores no raw trajectories, and it has the lowest parameter growth and storage in Table A11, 4.8M/4.9M per stage and 53.0/53.5 MB. We agree that even longer streams merit further study, as noted in Appendix D.

Second, for ambiguity among closely related skills, the stress test in Table R3, repeated in Q2 below, evaluates ESR after all 50 bimanual skills are learned, under unseen instructions and their edits, including confusable pairs such as stacking three vs. two blocks in tasks 45/46. Moreover, previous LoRA updates are frozen and each skill keeps its own composition, so new skills do not overwrite earlier ones.

Q1: Contribution of evolving knowledge under a matched information budget.

A1: Thank you for the question. We assess the contribution of evolving knowledge from two angles.

First, the same information with and without evolution in Table 6. Base + SR (static) and Base + SR share the same records, demonstrations, LoRA updates and reasoning rules; they differ only in whether the shared statistics S_s are recalibrated as skills accumulate, which isolates the evolving part of the knowledge.

Table 6, excerpt. LIBERO results without oracle IDs; FSR and CSR are in % and forgetting in pp.

VariantGoal FSR ↑Goal CSR ↑Goal Forg. ↓Long FSR ↑Long CSR ↑Long Forg. ↓
Base + SR (static)75.478.918.947.650.522.4
Base + SR80.283.812.552.858.118.7

Recalibrating S_s improves FSR by 4.8 and 5.2 points and reduces forgetting by 6.4 and 3.7 points.

Second, alternative selectors on the same policies in Table R9. All selectors share the frozen encoder and CKT-trained policies, and the ESR ablations remove only attribute refinement or its gating.

Table R9, repeated from A1. RoboTwin, Table 7 protocol; FSR is closed-loop.

CategorySelectorEvidenceSR acc. (%)FSR (%)
Prototype-based retrievalSemantic prototype c0per-skill mean VLM features58.228.5
Prototype-based retrievalHiDe-LLaVA [2]per-skill image and text anchors64.236.1
Discriminative classificationRanPAC [3]random projection, ridge classifier54.726.4
ESR ablationESR w/o attributes c1no attribute refinement82.344.2
ESR ablationESR w/o gatingrefinement always applied80.541.3
OursFull ESRfull evidence, updated statistics91.052.8

As shown in Table R9, ESR leads prototype retrieval and discriminative classification on the same policies by at least 26.8 pp in SR acc. and 16.7 pp in FSR. Removing attribute refinement or its gating lowers SR acc. to 82.3% and 80.5%, so the gain comes from conditional reasoning over attributes rather than attribute evidence alone.

AuthorsOpenReview ↗

Response to Reviewer hSqq

Q2: Ambiguous, conflicting or out-of-distribution instructions, and real-robot uncertainty.

A2: Thank you for the question. For ambiguous and out-of-distribution instructions, we repeat the RoboTwin stress test from our response to Reviewer weVY on W3.

First, attribute refinement is conditional and falls back safely: it applies only when the initial correction was triggered, the two leading scores are within a margin and attribute evidence separates them, following Eq. 3 and Appendix B.2; otherwise, including when no supported attribute is parsed, the scores after instruction matching are kept. Contradictory attributes are averaged in the attribute agreement of Eq. A10, so if they no longer separate the leading candidates, the gate stays closed and the same fallback applies.

Second, we freeze the final RoboTwin checkpoint and test all 50 tasks with RoboTwin's own unseen instruction R0, five single edits of it and one random pair of edits, keeping scene, target skill and parsed attribute set fixed, so every G1–G6 instruction is an unseen variant of R0.

  • R0: RoboTwin's unseen evaluation instruction for each episode, unchanged.
  • G1, verb paraphrase: one verb in R0 is replaced by a synonym from a fixed lexicon, such as "rotate" → "spin".
  • G2, object alias: one object is renamed with another of RoboTwin's unseen descriptions of the same asset.
  • G3, syntactic restructuring: R0 is reordered by fixed rewriting rules without adding or removing content words.
  • G4, irrelevant background: the fixed sentence "The tabletop is visible." is prepended to R0.
  • G5, out-of-schema attribute: one object receives a real color, material or shape adjective from its asset descriptions, unseen in that task's training instructions and not read by the parser.
  • G6, combined edits: two edits from different categories in G1–G5 are applied together.

Table R3, repeated from our response to Reviewer weVY. Skill reasoning accuracy in % on all 50 RoboTwin tasks after the final stage; all selectors use the same policies and instructions.

IDInstruction conditionSemantic prototype c0HiDe-LLaVAESR
R0RoboTwin unseen instruction58.264.291.0
G1Verb paraphrase68.671.489.7
G2Object alias71.967.090.8
G3Syntactic restructuring62.468.090.4
G4Irrelevant background59.867.988.2
G5Out-of-schema attribute79.165.486.3
G6Combined edits, two of G1–G553.157.587.6

As shown in Table R3, ESR reaches 91.0% on RoboTwin's unseen instructions, stays at 86.3–90.8% under every single edit, including out-of-schema attributes, and at 87.6% under combined edits. It leads both alternatives under every condition by 7.2–30.1 pp over the stronger one, with the largest margin under combined edits, where they fall to 53.1% and 57.5%.

For real-robot uncertainty, we evaluate the final policy two more times under the in-domain and out-of-domain protocols of Table 5, with 10 rollouts per task per run.

Table R14. ThoughtVLA real-robot FSR in % over three evaluation runs.

ConditionRun 1, Table 5Run 2Run 3Mean ± std
In-domain53.049.552.551.7 ± 1.9
Out-of-domain40.543.042.542.0 ± 1.3

As shown in Table R14, FSR is 51.7 ± 1.9% in-domain and 42.0 ± 1.3% out-of-domain. Even the lowest runs, 49.5% and 40.5%, stay above the strongest baselines in Table 5, SkillsCrafter and AtomicVLA, both at 43.5% in-domain and 31.0% out-of-domain. We have added Table R14 to Appendix C.5 as Table A19.

Reviewer_JqUXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ThoughtVLA, a framework for continual adaptation of pretrained vision-language-action models. The method separates pre-action skill selection from skill execution. Evolving Skill Reasoning (ESR) uses accumulated task knowledge and conditional attribute-based refinement to select a learned skill, whose stored LoRA composition then remains fixed throughout execution. Complementary Knowledge Transfer (CKT) selects prior LoRA updates that complement the limited capacity of a new low-rank adapter. The paper evaluates the framework across simulated and real-robot continual learning settings and reports improved final success rates and reduced forgetting over the included baselines.

Critical strengths

The separation of skill selection from execution is well motivated and supported by a useful controlled ablation.

The paper identifies a concrete failure mode: freezing previous experts does not necessarily preserve old-task performance if subsequent changes to routing cause the wrong skills to be selected. The proposed separation addresses this issue directly. Table 6 provides supporting evidence by holding the CKT-trained policies fixed while varying the skill selector. In particular, adding attribute refinement increases RoboTwin final success from 44.2% to 52.8%. This comparison helps isolate the contribution of skill disambiguation from changes to the underlying execution policies.

Critical weaknesses
  1. Missing experimental comparison and insufficient engagement with a highly relevant finding on pretrained VLAs.

The paper cites Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning by Huihan Liu et al. [1], but does not adequately address its implications for the proposed framework. That work shows that simple experience replay can already provide strong retention and adaptation in pretrained VLAs. This finding is directly relevant to the motivation and empirical significance of ThoughtVLA.

A well-tuned LoRA or action-head fine-tuning baseline with simple experience replay, evaluated using the same pretrained initialization and task stream, is therefore an important missing comparison. ThoughtVLA avoids replaying previous demonstrations, which is a meaningful property, but its evaluation already includes replay-based methods such as DMPEL and MLR-IFA. It remains unclear how the proposed approach compares with a simpler replay strategy across success rate, forgetting, total retained storage, and training cost. The brief acknowledgment in the related work does not resolve this question. Without this comparison and a substantive discussion of [1], the evidence for the additional value of the evolving knowledge and transfer mechanisms remains incomplete.

  1. The absolute performance and practical value of the evaluation setting are insufficiently contextualized.

ThoughtVLA achieves a mean final success rate of 90.4% across the four LIBERO streams, with 87.0% on Spatial and 81.0% on Long. Although these results improve over the reported baselines, the absolute performance is not sufficiently compelling on its own to establish the practical value of the additional skill-selection, knowledge-maintenance, and transfer machinery. This is particularly relevant given that the protocol assumes known training-stage boundaries and task-grouped demonstrations, while the retained skill library grows over time.

The paper lacks joint multi-task and independently trained per-task reference results using the same backbone and data. Such references would help establish the performance achievable in this setup and clarify how much of the remaining gap comes from continual adaptation, skill selection, or incomplete skill acquisition. The oracle-skill comparison in Table 6 is informative, but it evaluates the same learned execution policies with correct skill identities and therefore does not provide this reference.

Consequently, I find the practical significance of the reported performance–resource trade-off insufficiently established. A clearer contextualization of the final success rates would strengthen the case that this setting demonstrates a consequential benefit from the proposed framework.

Questions for authors

Can the authors provide a comparison with a well-tuned simple experience replay baseline motivated by Liu et al. [1], using the same StarVLA initialization, task sequence, and evaluation protocol? Evaluating several replay budgets and reporting total retained storage and training cost would help clarify when ThoughtVLA offers a meaningful advantage. How do the resulting observations relate to the conclusions of [1]?

Can the authors provide joint multi-task and independently trained per-task reference results on LIBERO under the same backbone and data, with training budgets clearly specified? These references would help contextualize the 90.4% average final success rate and identify the main source of the remaining performance gap.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JqUX

W1 & Q1: Comparison with simple experience replay across budgets, storage and training cost, and relation to Liu et al. [1].

A1: Thank you for raising this important point. We agree that [1] is directly relevant, and we now compare against it under our protocol.

Setup. We follow the replay rule of [1]: a fixed number of transitions per past task, sampled once and reused; we evaluate 100, 500 and 1,000 transitions, of which 100 and 1,000 are also used in [1]. The replay baselines use the same configuration as ThoughtVLA, namely the backbone, rank-16 action-head LoRA, demonstrations and 110K formal steps per suite, and Sequential FT is taken from Table 1; for each budget, the learning rate and current:past ratio are tuned on held-out action loss.

Table R7. FSR: final average success; CSR: current-stage average success; Forg.: forgetting as defined in Sec. 5.1. Stored replay is the size of the past transitions kept for replay at the final stage.

SuiteMethodTrans./past taskFSR (%)CSR (%)Forg. (pp)Stored replay (MB)
GoalSequential FT09.826.094.70
Goal+ replay10083.686.16.7393
Goal+ replay50092.293.83.61,965
Goal+ replay1,00094.895.51.93,930
GoalThoughtVLA095.095.51.20
LongSequential FT010.023.883.30
Long+ replay10064.475.219.6393
Long+ replay50077.080.45.91,965
Long+ replay1,00081.082.01.63,930
LongThoughtVLA081.083.41.40

Results. As shown in Table R7 and consistent with [1], replay cuts forgetting on Long from 83.3 pp to 1.6 pp with 1,000 transitions per past task. ThoughtVLA matches it without storing any past transitions, at 95.0% vs. 94.8% FSR on Goal and 81.0% for both on Long, whereas replay keeps 3,930 MB. With 100 transitions, replay already stores 393 MB, about seven times ThoughtVLA's 53 MB state, yet trails by 11.4 and 16.6 pp on Goal and Long. We have revised Sec. 2 at l. 124–126 and added this comparison to Table 4 and Appendix C.2:

"Liu et al. (2026) show that pretrained VLAs trained with simple experience replay forget little when enough past transitions are replayed; ThoughtVLA instead targets settings where raw trajectories are not retained."

[1] Liu et al. Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning. arXiv:2603.03818, 2026. (https://arxiv.org/abs/2603.03818)

W2 & Q2: Contextualizing absolute performance with same-backbone joint and per-task references and their training budgets.

A2: Thank you for this constructive suggestion. Table R8 reports joint and per-task references on all four LIBERO suites with the same backbone, LoRA, demonstrations and 110K steps per suite. Joint training sees all ten tasks from the start, each per-task policy trains for 11K steps, and ThoughtVLA uses 80K steps for tasks 1–5 and 6K per later task.

Table R8(a). Final success in % per task on LIBERO-Object and LIBERO-Long; task IDs follow Tables A4–A5, and ThoughtVLA averages are from Table 1.

SuiteMethod12345678910Avg
ObjectThoughtVLA100.098.0100.0100.0100.0100.0100.094.094.0100.098.6
ObjectJoint100.098.0100.0100.0100.098.0100.098.096.0100.099.0
ObjectPer-task88.0100.098.072.078.096.092.098.0100.092.091.4
LongThoughtVLA88.074.0100.084.098.076.070.092.034.094.081.0
LongJoint92.076.0100.086.098.080.076.092.056.098.085.4
LongPer-task86.074.0100.080.084.080.072.086.026.082.077.0

Table R8(b). Average final success in % per suite, with 110K formal steps per suite for every method. ThoughtVLA values are from Table 1.

MethodObjectGoalSpatialLongMean
ThoughtVLA98.695.087.081.090.4
Joint99.097.290.885.493.1
Per-task91.489.088.277.086.4

As shown in Table R8, ThoughtVLA comes within 0.4 pp of joint training on Object and 2.7 pp on average, and exceeds per-task training by 4.0 pp on average. This lead comes from transfer: on Long, ThoughtVLA without transfer reaches only 70.2% in Table 8, below per-task training at 77.0%, and CKT raises it to 81.0%. The gap to joint training reflects the small per-stage budget rather than forgetting: on Long, it is 1.6 pp on tasks 1–5 but 7.2 pp on tasks 6–10, with 1.4 pp forgetting. We have added these references to Table 4 and Appendix C.2.

Thank you for the response and additional experiments. While Table R8 largely contextualizes the LIBERO results (W2/Q2), my core concern regarding the practical necessity of the framework (W1/Q1) remains unresolved:

  1. Unconvincing trade-off: Table R7 confirms that simple experience replay with 1,000 transitions matches ThoughtVLA's performance and low forgetting. On modern robot platforms running a 4B VLM, ~3.9 GB of disk storage is negligible. Saving ~3.9 GB does not justify the substantial pipeline complexity (attribute parsing, gated ESR, SVD-based CKT, and discarded paired trial fine-tuning).
  2. Missing training cost: Despite my explicit request in W1/Q1, training cost was omitted from Table R7, leaving the overhead of CKT scoring and paired trials unaccounted for against simple replay.

I will maintain my current score.

AuthorsOpenReview ↗

Response to Reviewer JqUX

Thank you very much for your follow-up and for confirming that Table R8 contextualizes the LIBERO results. We address your two remaining points below.

1. Trade-off between replay storage and pipeline complexity.

We agree that replay is a strong baseline when past data can be kept, and that LIBERO, with ten tasks per suite, is a small setting in which its storage stays modest. We address the trade-off from two angles.

(a) Past data is often not retained in practice. Continual learning distinguishes rehearsal-free methods precisely because buffering past data raises memory costs and may be impossible under data-privacy constraints, as motivated by L2P [6] and CODA-Prompt [7] and followed by RanPAC [3]. The same holds for robots: demonstrations are often collected on customer sites or under data-protection and licensing terms that limit how long they may be kept, so the past data that replay relies on may be unavailable when later skills arrive. ThoughtVLA targets this regime: it retains no raw trajectories and keeps every learned skill frozen, so no later stage needs to revisit earlier data. We will state this positioning explicitly in the paper.

(b) Replay storage grows with longer streams. Under the same rule of 1,000 transitions per past task, storage grows with the number of tasks and with richer observations, as Table R15 shows for our two longest streams.

Table R15. Replay storage at the final stage with 1,000 transitions per past task. Storage costs are empirically measured for all task streams.

StreamPast tasksCamera viewsPast transitionsReplay storage (MB)
LIBERO, 10 tasks929,0003,930
RoboTwin, 50 tasks45345,00029,475
Mixed, 55 skills542 / 3 / 154,00024,890
ThoughtVLA, any stream——00

As shown in Table R15, replay storage reaches 29,475 MB on RoboTwin, about 7.5 times the 3,930 MB on LIBERO, and grows by 3,275 MB per RoboTwin stage, whereas ThoughtVLA keeps 53.5 MB and no past data. Reducing the budget to save storage costs accuracy: with 500 transitions per past task, replay trails ThoughtVLA by 2.8 and 4.0 pp on Goal and Long, and with 100 by 11.4 and 16.6 pp. Because ThoughtVLA does not rely on stored data, with its retained state growing by only one small LoRA per stage, it extends naturally to more application scenarios and to longer task streams.

Each component earns its place. Attribute parsing and gated ESR add 8.7 SR acc. points in Table 7 at 8–33 ms per episode. CKT, whose low-rank analysis uses only 32 compressed activation columns per layer, adds 10.8 FSR points over no transfer in Table 8, and evolving statistics add 4.8 and 5.2 FSR points in Table 6. The cost of paired trials is addressed below.

2. Training cost of CKT scoring and paired trials.

We apologize for omitting the training cost from Table R7 and add it in Table R16.

Table R16. Table R7 with training time per incremental stage, empirically measured on two A100 GPUs. Replay adds a full batch of past samples per step, and ThoughtVLA includes CKT scoring and paired trials.

SuiteMethodTrans./past taskFSR (%)Forg. (pp)Stored replay (MB)Train (h/stage)
GoalSequential FT09.894.702.0
Goal+ replay10083.66.73932.9
Goal+ replay50092.23.61,9652.9
Goal+ replay1,00094.81.93,9302.9
GoalThoughtVLA095.01.202.3
LongSequential FT010.083.302.0
Long+ replay10064.419.63932.9
Long+ replay50077.05.91,9652.9
Long+ replay1,00081.01.63,9302.9
LongThoughtVLA081.01.402.3

As shown in Table R16, ThoughtVLA trains about 0.6 h faster per stage than replay at every budget, since replay's extra batch of past samples outweighs CKT scoring and paired trials. ThoughtVLA stores no past data, and its training time is lower than that of four of the five continual baselines in Table A11, including MLR-IFA, which replays 100 samples per task. Paired trials run only for the top candidate, and their LoRAs are discarded so that all methods share the same formal steps. We will add Tables R15 and R16 to Appendix C.2.

We hope this addresses your concerns and would be glad to clarify further.

[3] McDonnell et al. RanPAC: Random Projections and Pre-trained Models for Continual Learning. NeurIPS 2023. (https://arxiv.org/abs/2307.02251)

[6] Wang et al. Learning to Prompt for Continual Learning. CVPR 2022. (https://arxiv.org/abs/2112.08654)

[7] Smith et al. CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning. CVPR 2023. (https://arxiv.org/abs/2211.13218)

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

This paper proposes ThoughtVLA, a continual adaptation framework that formulates continual VLA adaptation into two stages. It first interprets the current input using accumulated knowledge to select a learned skill for new-skill learning through Evolving Skill Reasoning (ESR). Then Complementary Knowledge Transfer (CKT) examines and determines which prior LoRA updates are incorporated into the new Skill’s LoRA adapter to construct the new skill’s LoRA composition. Extensive experiments demonstrate that ThoughtVLA could achieve higher final success and lower forgetting across multiple settings.

Critical strengths
  1. The Continually Evolving Thought Trajectory (CETT) provides an interesting perspective on continual VLA adaptation that introduces structured, thought-based reasoning skills that link skill acquisition with evolving task knowledge.
  2. The CKT mechanism offers a reasonable approach for learned knowledge reuse by assessing whether learned skills provide complementary knowledge through local approximation analysis, paired adaptation trials, and held-out validation.
  3. The broad empirical evaluation across single-arm, bimanual, humanoid, mixed-embodiment, cross-backbone, and real-robot demonstrates the effectiveness of ThoughtVLA. Ablation studies further substantiate the contributions of ESR and CKT.
Critical weaknesses
  1. All tasks introduced within a stage share the same newly trained LoRA adapter. The coefficients of any selected historical composition are directly inherited without being re-estimated for the new task. How to quantify the transfer contributions of different tasks and balance beneficial sharing against interference among tasks within a stage remains unclear.
  2. The scalability of the framework under a fixed resource budget remains unclear. Each stage permanently adds a LoRA update, while CKT and ESR operate over an expanding skill library, which may lead to high candidate-scoring cost and selection latency over long task streams.
  3. ESR’s explicit attribute refinement relies on a predefined parser and fixed rules for correcting candidate scores. The robustness of this mechanism to attributes outside the predefined schema and previously unseen attribute combinations remains underexplored.
Questions for authors
  1. As the framework could select no transfer when learning new tasks, to what extent does CKT reuse previously learned skills? Could the authors quantify the reuse relationships between tasks?
  2. Beyond the task-order and training-seed analyses in Table A12, how sensitive is the framework to the composition and size of the initial skill set? Do the alternative task orders also change which tasks are learned in the initial stage, or only reorder subsequent tasks?
  3. With the learned policy library and LoRA compositions held fixed, how do text and visual input contribute to skill selection?
3 comments
AuthorsOpenReview ↗

Response to Reviewer weVY

W1: How are per-skill transfer contributions quantified, and how is within-stage sharing balanced against interference?

A1: Thank you for raising this important point.

  1. Coefficients. Inherited updates are reused as stored in line 12 of Alg. 1, and in line 18 the new stage LoRA is trained from fresh initialization with each skill's inherited updates active on half of each minibatch, following Eq. A12, so it learns the residual the new skill still needs on top of them. The inherited updates thus stay exactly as CKT validated them, and no extra per-skill parameters are fitted on limited demonstrations.
  2. Balancing sharing against interference. Each skill combines a shared part, the new stage LoRA trained on all skills of the stage, with a skill-specific part, the previous updates CKT selects for that skill alone as the best complement to its new rank-r update, following Eq. 8 and Appendix B.3. Skill-specific differences are thus carried by each skill's own updates rather than forced into the shared LoRA. A transfer is kept only if that skill's paired trials and held-out validation confirm it under Eq. A11; otherwise the skill uses no transfer, so a transfer that does not help a skill is not imposed on it.
  3. Quantifying contributions. On RoboTwin tasks 26–35 from stages 2–3, we compare ThoughtVLA with two variants: No transfer keeps the same shared stage LoRA without CKT and isolates the transfer gain, while PS-LoRA+CKT trains one new LoRA per skill with the same transfer and isolates the effect of sharing. Both are matched in rank and per-skill data and evaluated with oracle skill selection.

Table R1. Success rate in % on RoboTwin tasks 26–35 from stages 2–3 after each skill's stage, with 100 rollouts per task; task IDs follow Table A6. Transfer gain = ThoughtVLA − No transfer; sharing effect = ThoughtVLA − PS-LoRA+CKT; Params = new LoRA parameters per stage in millions.

Method26272829303132333435MeanParams (M)
No transfer64522143164487212530.14.9
PS-LoRA+CKT7476317126125615325344.624.5
ThoughtVLA727827652265119344842.24.9
Transfer gain+8+26+6+22+6+2+3+12+13+23+12.1–
Sharing effect−2+2−4−6−4−6−5+4+2−5−2.4–

First, beneficial sharing. As shown in Table R1, CKT transfer improves all ten skills over No transfer, raising mean success from 30.1% to 42.2%, a gain of 12.1 pp that reaches 26 pp on task 27. The gain holds in both stages, at 13.6 pp in stage 2 and 10.6 pp in stage 3.

Second, interference. The shared stage LoRA reaches 42.2% versus 44.6% for per-skill LoRAs, retaining about 95% of their success, and even surpasses them on three skills, while adding five times fewer new parameters, 4.9M versus 24.5M per stage. Interference within a stage is thus small relative to the transfer gain. We have added this clarification to Sec. 4.2 at l. 301–305, and Table R1 to Sec. 5.4 as Table 9 and to Appendix C.8.

W2: Candidate-scoring cost and selection latency over long streams.

A2: Thank you for this question.

First, per-stage growth is small: each stage adds only 4.8M/4.9M action-head parameters, and the final persistent state is 53.0/53.5 MB on LIBERO-Object/RoboTwin, the lowest in Table A11.

Second, scoring and selection costs grow at most linearly: CKT scoring is linear in the number of prior compositions by line 7 of Alg. 1, ESR semantic scoring is linear in the number of acquired skills by Eq. 2, refinement stays within K_c = 5 candidates, ESR runs once per episode, and paired trials run only for the top candidate. Table R2 measures these costs on the mixed-embodiment stream, with batch size 1 on one A100 after warm-up, and compares the per-episode selection cost with AtomicVLA, the strongest baseline on this stream, and CLARE, whose routing grows with the library.

Table R2. CKT cost per new skill and per-episode selection cost of ESR, AtomicVLA and CLARE on the 55-skill stream.

SkillsCKT scoring (min/skill)ESR (ms)AtomicVLA (ms)CLARE (ms)
51.8825163
144.312251627
246.817251651
347.022251673
447.3272516101
557.5332516135

As shown in Table R2, CKT scoring runs once per new task and takes at most 7.5 min at 55 skills, a small fraction of the 2.3/6.5 h per-stage training time on LIBERO-Object/RoboTwin in Table A11. ESR takes 33 ms per episode at 55 skills, 76× less than AtomicVLA and 4.1× less than CLARE. This is under one tenth of a single action-chunk prediction, 380/600 ms on LIBERO-Object/RoboTwin in Table A11. We have added this analysis to Appendix C.1 with Table A12.

AuthorsOpenReview ↗

Response to Reviewer weVY

W3: Robustness to out-of-schema attributes and unseen attribute combinations.

A3: Thank you for this insightful comment.

First, attribute refinement is conditional and falls back safely: it applies only when the initial correction was triggered, the two leading scores are within a margin and attribute evidence separates them, following Eq. 3 and Appendix B.2; otherwise, including when no supported attribute is parsed, the scores after instruction matching are kept.

Second, we freeze the final RoboTwin checkpoint and test all 50 tasks with RoboTwin's own unseen evaluation instruction R0, five single edits of it and one random pair of edits, keeping scene, target skill and parsed attribute set fixed. Every G1–G6 instruction is therefore an unseen variant of R0: G5 pairs an object with an attribute word that never appears in that task's training instructions, and G6 further combines two edits.

  • R0: the instruction that RoboTwin's own unseen evaluation draws for each episode, used unchanged.
  • G1, verb paraphrase: one verb in R0 is replaced by a synonym from a fixed lexicon, such as "rotate" → "spin".
  • G2, object alias: one object is renamed with another of RoboTwin's unseen descriptions of the same asset.
  • G3, syntactic restructuring: R0 is reordered by fixed rewriting rules without adding or removing content words.
  • G4, irrelevant background: the fixed sentence "The tabletop is visible." is prepended to R0.
  • G5, out-of-schema attribute: one object receives a real color, material or shape adjective from its asset descriptions, which the parser does not read.
  • G6, combined edits: two edits from different categories among G1–G5 are applied to the same instruction.

Table R3. Skill reasoning accuracy in % on all 50 RoboTwin tasks after the final stage; all selectors, as in Table R9, use the same policies and instructions.

IDInstruction conditionSemantic prototype c0HiDe-LLaVAESR
R0RoboTwin unseen instruction58.264.291.0
G1Verb paraphrase68.671.489.7
G2Object alias71.967.090.8
G3Syntactic restructuring62.468.090.4
G4Irrelevant background59.867.988.2
G5Out-of-schema attribute79.165.486.3
G6Combined edits, two of G1–G553.157.587.6

As shown in Table R3, ESR reaches 91.0% on RoboTwin's unseen instructions and stays within 4.7 pp of this under every single edit, at 86.3–90.8%, and within 3.4 pp under combined edits, at 87.6%. Out-of-schema attributes, which the parser does not read, keep ESR at 86.3%, ahead of both alternatives. ESR leads the semantic prototype and HiDe-LLaVA under every condition, by 7.2–30.1 pp over the stronger of the two, and the margin is largest under combined edits, where they fall to 53.1% and 57.5%. We have added this test to Appendix C.7 with Table A22.

Q1: To what extent does CKT reuse learned skills?

A1: Thank you for the question.

Extent of reuse. CKT proposes a transfer only when the best candidate has Δr > 0 and keeps it only if paired trials and held-out validation confirm it; otherwise the target learns without transfer. Accepted transfers are effective: on LIBERO-Long, transfer raises FSR by 10.8 pp in Table 8, and on RoboTwin tasks 26–35 it improves all ten skills, by 12.1 pp on average, in Table R1.

Reuse relationships. Skills learned in the same stage share that stage's update ω_s, so reuse is defined over stage updates: a target inherits the full composition of its source, and one accepted transfer can carry several earlier updates. Table R4 lists eight RoboTwin targets across the five incremental stages; full instructions are in Table A6.

Table R4. Stage updates that eight RoboTwin targets inherit through CKT; ω_s is the update learned in stage s, and — denotes no transfer.

TargetStageTaskInherited stage updates
262place_can_basketω1
333place_object_basketω1, ω2
374place_shoeω2
404put_object_cabinet—
445shake_bottle_horizontally—
455stack_blocks_threeω1
466stack_blocks_twoω1, ω5
486stack_bowls_two—

As shown in Table R4, five of the eight targets accept a transfer and three learn without one, so CKT reuses earlier updates selectively rather than by default. Accepted transfers also chain across stages: place_object_basket in stage 3 inherits both ω1 and ω2, and stack_blocks_two in stage 6 inherits ω1 and ω5, so the stage-1 update still contributes five stages later. We have added Table R4 to Appendix C.8 as Table A24.

AuthorsOpenReview ↗

Response to Reviewer weVY

Q2: Sensitivity to the composition and size of the initial skill set.

A2: Thank you for this question.

Composition. Yes, the alternative orders O1 and O2 in Table A15 also change the initial skill set, as listed in Table R5. Across the three initial sets, FSR stays within 0.6 pp, at 81.0 / 81.4 / 80.8 on LIBERO-Long and 95.0 / 94.8 / 95.4 on LIBERO-Goal.

Size. We further vary the number of initial tasks on LIBERO-Long among 3, 4 and 5 with the same 110K formal steps, using 68K / 74K / 80K steps in stage 1 and then 6K per incremental task.

Table R5. Initial-set sensitivity on LIBERO-Long with seed 42; O0–O2 are from Table A15, and task IDs follow Table A5.

FactorRunInitial task IDsFSR (%)Forgetting (pp)Δ FSR vs. O0 (pp)
CompositionO01, 2, 3, 4, 581.01.4—
CompositionO18, 4, 5, 7, 181.41.3+0.4
CompositionO23, 10, 6, 2, 480.81.6−0.2
Size3 initial tasks1, 2, 379.81.6−1.2
Size4 initial tasks1, 2, 3, 480.61.4−0.4

As shown in Table R5, with 3 or 4 initial tasks FSR reaches 79.8% and 80.6%, within 1.2 pp of the default of 5 initial tasks in O0, even though more tasks are then learned incrementally; forgetting stays at 1.4–1.6 pp. Across all five initial sets, FSR remains within 79.8–81.4%, so ThoughtVLA is insensitive to the composition and size of the initial skill set within the tested range. We have added this analysis to Appendix C.3 with Table A16.

Q3: Contribution of text and visual input to skill selection.

A3: Thank you for the question. We mask one modality at the ESR input only, keeping execution policies and LoRA compositions unchanged, and re-encode skill prototypes and shared statistics under the same masking without any training; text-only keeps lexical and attribute refinement, and image-only disables them. We report selection accuracy on RoboTwin's unseen evaluation instructions for all 50 tasks after the final stage, as in Table R3, and closed-loop FSR.

Table R6. ESR with one input modality masked on RoboTwin. SR acc.: selection accuracy on all 50 tasks after the final stage; FSR: closed-loop final success.

ESR inputSR acc. (%)FSR (%)
Image and text91.052.8
Text only82.138.6
Image only35.110.3

As shown in Table R6, ESR with both modalities reaches 91.0% selection accuracy, versus 82.1% with text only and 35.1% with image only: the instruction carries most of the task identity, since many RoboTwin tasks share similar scenes and objects, while visual input adds a further 8.9 pp on top of text. Closed-loop FSR follows the same order, 52.8% vs. 38.6% and 10.3%, so the two modalities are complementary and combining them gives the most reliable selection. We have added this analysis to Appendix C.7 with Table A23.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.