acceptodds
Under review as a conference paper at ICLR 2027

Language-Forcing: Semantic Alignment for Whole-Body World Action Model

12 views

Abstract

World Action Models (WAMs) jointly predict robot actions and future observations. However, their visual prediction objectives do not explicitly distinguish task-relevant transitions from incidental scene changes. For whole-body policies, this distinction matters because task progress depends on coordinated changes in body pose, contact, and object state. We introduce LANGUAGE-FORCING, a training strategy that uses language to emphasize task-relevant changes in the future visual representations of a whole-body WAM. Given paired current and ground- truth short-horizon future frames, a frozen vision-language model assesses task progress and predicts plausible subsequent state changes. We encode these descriptions with a frozen text encoder and align their embeddings with intermediate, action-conditioned future-visual features, complementing the original visual and action prediction objectives. All auxiliary components are used only during training, introducing no additional inference overhead. On HumanoidArena, LANGUAGE-FORCING achieves an average success rate of 79.17%, exceeding the strongest competing policy by 9.46 percentage points and the same WAM without language alignment by 10.77 percentage points. Across five real-world tasks, language alignment raises average success from 53.33% to 71.33% over the same WAM baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotics;World Action Model;Humanoid;Loco-Manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6AFMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper builds on the pretrained GigaWorld-Policy-0.5 WAM [1], adapting its action interface for whole-body control and adding auxiliary language supervision. A frozen model describes task progress and anticipated changes from current and future training frames, and the resulting text embeddings supervise intermediate future-visual features. Evaluation includes simulation and real-robot tasks.

[1] GigaWorld-Policy-0.5 — GigaWorld Team et al.

Critical strengths

The experiments show improvements, including on real world robot. The ablations provide evidence that the correspondence and content of the language targets matter.

Critical weaknesses

Novelty is limited. The method builds on an existing pretrained WAM, and mainly adds an auxiliary language alignment loss. The contribution appears to be an incremental adaptation of established representation-alignment techniques [1–3].

[1] GigaWorld-Policy-0.5 — GigaWorld Team et al.

[2] WorldScape Policy 2.0 — Su et al.

[3] Spatial Forcing — Li et al.

Questions for authors

Under the same backbone and training budget, how does the method compare with other semantic-alignment methods?

Is Language Forcing compatible with other WAM improvements, and does combining them yield additional gains beyond either method alone?

Beyond using existing humanoid control and benchmarking frameworks, which aspects of Language Forcing itself are specifically designed for humanoid whole-body control, and how do they differ from related alignment methods in robot arm manipulation or VLN?

3 comments
AuthorsOpenReview ↗

Comment for Reviewer 6AFM

We thank Reviewer 6AFM for asking where the method differs from established alignment techniques and whether it transfers across WAM improvements. These are important questions for defining the contribution precisely.

“Under the same backbone and training budget, how does the method compare with other semantic-alignment methods?”

As also explained in our response to Reviewer 8hce, the matched visual-target control gives the following comparison. We include the result here for readability. only the alignment target changes.

modelLanguage-ForcingDINOv2 alignment
Giga-World-Policy-0.579.1776.19

We thank Reviewer 6AFM for pointing us to WorldScape Policy 2.0, an important related direction. The methods address different learning targets: WorldScape uses event-caption supervision for planning representations, whereas Language-Forcing uses paired current/future observations to provide transition-specific progress and next-state supervision for action-conditioned future-visual features. This difference in target, representation, and use of the action trajectory is fundamental, so the two methods should not be treated as interchangeable versions of the same method.

“Is Language Forcing compatible with other WAM improvements, and does combining them yield additional gains beyond either method alone?”

The compatibility results are:

SettingAverage SR (%)
Giga-World-Policy-0.5, W/O LF68.39
Giga-World-Policy-0.5, W/ LF79.17
fastwam, W/O LF61.13
fastwam + auxiliary language-prediction head67.50
fastwam, W/ LF70.00
lingbot-va, W/O LF65.36
lingbot-va, W/ LF72.98

The within-backbone comparisons show that LF adds a gain beyond the alternative WAM improvements, while the auxiliary head provides a separate compute-matched reference.

“Beyond using existing humanoid control and benchmarking frameworks, which aspects of Language Forcing itself are specifically designed for humanoid whole-body control, and how do they differ from related alignment methods in robot arm manipulation or VLN?”

We thank the reviewer for asking us to make the whole-body design boundary explicit. Language-Forcing uses a 35D humanoid state/action interface and action-conditioned future-visual features. Its target describes posture, contact, object state, and likely next motion. This matters because humanoid loco-manipulation requires the robot to move and manipulate simultaneously. In many arm-manipulation scenes, the background and lighting are relatively stable and the main changes are the arm and object. A humanoid whole-body policy sees the background, lighting, texture, object, and the robot body change together; large low-value pixel changes can overwhelm the smaller but critical changes in contact, stance, and object state. Language provides a concise task-level regularizer for this imbalance.

VLN also encounters rapid background and lighting changes, so the same principle may transfer to navigation. Because the discussion phase was time-limited, we could not run a broader VLN extension experiment; we leave that study for future work. The tested contribution is a transition-specific language target for an action-conditioned WAM, demonstrated on coordinated humanoid whole-body control. It is not a claim that cosine feature alignment is new in isolation, and it adds no inference-time cost.

Thank you for your responses to my three questions, which have addressed my concerns, particularly the clarifications regarding the novelty and boundary of the work. I recommend that the authors strengthen the clarification of the boundary in the main text to better highlight the novelty.

In addition, the results of LF on different backbones demonstrate the generalizability of the proposed method.

I consider this a simple but meaningful direction for improving humanoid WAM, and I will raise my score accordingly.

AuthorsOpenReview ↗

Thank you for the response

We thank Reviewer 6AFM for your encouraging feedback and for recognizing the value of Language-Forcing for humanoid WAMs. We are glad that our responses addressed your concerns and clarified the contribution. Thank you again for your time, careful review, and helpful suggestions.

Reviewer_8hceOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Language-Forcing, a training strategy for whole-body humanoid World Action Models (WAMs). A frozen VLM looks at the current and future frames and describes the task progress and what may happen next, and the text embedding is aligned with the WAM's intermediate future-visual features. These parts are only used in training, so there is no extra inference cost. Experiments on HumanoidArena and five real-world tasks demonstrate the effectiveness of Language-Forcing.

Critical strengths
  • The idea is quite straightforward: the task instruction stays the same during the whole episode, while the VLM description changes with each transition, giving denser task-level supervision at no inference cost.
  • The comparison with the base WAM is clean, since the language loss is the only difference, and Language-Forcing improves all 7 simulated and all 5 real-world tasks. The Fig. 4 controls also show that the description has to match each sample.
Critical weaknesses
  • The biggest weakness and the main question for this paper is that the key claim, that language is what helps, is not well supported. All alignment targets in Fig. 4 are text, and there is no non-language target such as a frozen image embedding of the future frame, the usual choice in feature alignment works like REPA (Yu et al., ICLR 2025) and Spatial Forcing. Future-image captions already recover about half of the gain, and the prompts also ask for colors, textures and materials (App. A.5). So I feel the contribution beyond REPA-style alignment is limited.
  • The mechanism is not clear to me. The language loss is on future-visual tokens, which the action queries do not attend to and which are dropped at inference. There is no analysis of how this helps the action branch, and the Fig. 3 attribution method is not described.
  • The baselines are retrained by the authors on a modified 35D action interface (App. A.4) with different batch sizes, and Language-Forcing is not the best on P&PBox or Boxing, so the "state-of-the-art" claim should be toned down. There are also some typos (e.g., "humnanoid").
Questions for authors
  • Could you replace the text target with a frozen image embedding of the future frame (e.g., DINOv2), keeping everything else fixed? If it is clearly worse, I would consider raising my score.
  • The task-instruction target (Fig. 4) and the repeated-current-frame ablation (Sec. 4.3) both report exactly 65.18%. Are these separate runs?
  • For the real-world experiments, were the language targets generated from real-robot frames?
1 comment
AuthorsOpenReview ↗

Comment for Reviewer 8hce

We thank Reviewer 8hce for reading the paper carefully and for the constructive suggestions that help make the paper more complete. The matched visual-target control directly tests whether the gain comes from language-specific supervision or from adding any frozen representation target.

“Could you replace the text target with a frozen image embedding of the future frame (e.g., DINOv2), keeping everything else fixed? If it is clearly worse, I would consider raising my score.” Yes, and thank you for suggesting this control. The only changed component is the frozen target; the WAM, projector, block, loss weight, data, and 50K-update schedule are unchanged.

SettingTargetSuccess rate (%)
W/O LF--68.39
W/ DINOv2 alignmentFuture-image embedding76.19
W/ Language-ForcingTransition text79.17

The visual target improves the baseline, while transition text adds a further gain under the same alignment interface.

“The task-instruction target (Fig. 4) and the repeated-current-frame ablation (Sec. 4.3) both report exactly 65.18%. Are these separate runs?”

Yes. They are separate runs with different alignment targets or visual-frame inputs. Both contain 1,095 successes out of 1,680 rollouts, which explains the identical percentage.

“For the real-world experiments, were the language targets generated from real-robot current/future frames?”

Yes. The real-robot cache uses paired current and future frames, task context, and the 0.96-second interval, exactly as in simulation. The response is encoded by frozen UMT5 and cached before policy training. The teacher, encoder, and projector are removed at inference.

We appreciate the reviewer's careful attention to the mechanism, and we are glad to explain our intuition. In humanoid loco-manipulation, a large WAM video expert models simultaneous changes in robot motion, objects, background, lighting, and texture. Many pixel changes are weakly related to task completion, while contact, posture, object state, and the next body motion are decisive. The language target acts as a training-time regularizer that encourages the action-conditioned predictive representation to emphasize those task-relevant changes. The action branch does not need future-visual tokens or the teacher at inference.

We also thank the reviewer for the wording check. We will state that LF obtains the best average success rate under the matched HumanoidArena protocol, while avoiding a claim that it is best on every task, and we correct “humnanoid”.

Reviewer_NGfWOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
1(−1)
Read review
Paper summary

The paper proposes Language-Forcing (LF), a training-time auxiliary loss for a whole-body World Action Model (WAM; built on GigaWorld-Policy 0.5). A frozen VLM (Gemma-4-12B) receives the current frame and the ground-truth frame 0.96 s later and writes a two-paragraph text ("task progress" and "future prediction"). The text is encoded by the frozen UMT5 encoder and mean-pooled into one 4096-d vector. The mean of the future-visual tokens at block 22 of the visual expert is passed through a small MLP and aligned to this vector with a cosine loss (weight 1.0). The VLM, the text encoder and the projector are dropped at inference, so there is no extra inference cost. On HumanoidArena (7 simulated tasks, 4 evaluation modes, 200 demonstrations per task) LF reaches 79.17% average success, versus 68.39% for the same WAM without LF and 69.70% for the best baseline (GR00T N1.7). On five real Unitree G1 tasks (30 trials each) it raises average success from 53.33% to 71.33%. Ablations vary the target type (task instruction / future-image caption / progress+future text), shuffle the text within a task, change the visual-frame inputs, the alignment block, the loss weight, and the training-set size.

Critical strengths
  1. Controlled, matched comparison and useful diagnostics. LF and no-LF models share initialization, data, steps, and evaluation protocol (Appendix A.3), so the headline gain of LF over its own backbone is measured in a clean setting. The shuffled-text control (63.21% vs. 79.17%) shows the benefit depends on the text matching the observation pair, and λ=0 recovers the baseline exactly. This makes the claim "an observation-dependent auxiliary target helps this WAM" credible, and every LF variant except λ=2.0 beats the baseline by at least 5 points.
  2. Zero inference overhead with real-robot evidence. The method adds a training-only projector and offline targets, so deployment cost is unchanged (91 ms on an RTX 4090). Real-robot results on five whole-body tasks (53.33% → 71.33%) point in the same direction as simulation, and the appendix documents the protocol, success criteria, and hyperparameter-selection process unusually clearly.
Critical weaknesses
  1. The gain is not attributed to "language semantics", and the headline number is likely inflated by a single-seed peak.
    • The target is a deterministic function of the same two frames that the WAM's visual loss already supervises, mean-pooled into one vector. Controls that would separate "language" from "any extra feature-level regularizer" are missing: aligning to Gemma's own hidden state of the same frame pair (no decoding), aligning to a visual encoder (REPA-style), a non-language progress target (normalized time-step regression or phase classification), a different text encoder, and separate ablations of the "progress" and "future" paragraphs (the intro claims they are complementary but they are never tested separately).
    • All results use one training seed. With 240 rollouts per task the binomial standard error is about 2.6 points, and the key comparisons are of similar size. The default configuration is an outlier in both independent sweeps: block 22 gives 79.17 while blocks 1/8/15/29 give 73.99/74.23/75.48/74.46; λ=1.0 gives 79.17 while 0.5/1.5 give 74.40/73.39. The claim "progress+future text > image caption" rests on a 5.6-point gap (79.17 vs 73.57) of the same magnitude, with confounded prompts. The robust gain across LF variants looks like about +5 points, not +10.77. Only five layers are tested, and no neighboring layers (e.g., 20–24).
    • Reporting: no per-mode (base / visual / semantic / execution) results are given, although semantic robustness is the natural place to test the claim. Two different experiments report exactly 65.18% (task-instruction target in Fig. 4; "current frame repeated in all five positions" in Sec. 4.3), i.e., the same 1095/1680 successes, which should be checked.
  2. Limited novelty, a motivation that is not validated, and a benchmark that barely tests language.
    • The mechanism is intermediate-feature cosine alignment (REPA, Spatial Forcing) with the target replaced by pooled VLM-text embeddings. Related work already aligns WAM/VLA features to language or semantic signals (Voltron, LIV, WorldScape Policy 2.0, and concurrently SG-WAM). The difference to these is not shown experimentally. Nothing in the method is specific to whole-body control, so the title and contribution 2 are overclaims, and only one benchmark is used.
    • The central motivation (the visual loss is dominated by task-irrelevant changes such as viewpoint shifts) is never measured; the only evidence is four qualitative attribution maps with no described method and no quantitative score (simulation ground-truth masks would allow one). Pooling to a single vector gives no direct mechanism for spatial focus, and the VLM prompt asks for colors, textures and materials rather than task-relevant changes.
    • HumanoidArena uses a fixed task-specific instruction, and its semantic shift only changes distractors under the same instruction, so it tests visual/control robustness rather than language grounding.
    • Compute/knowledge fairness: LF distills a 12B VLM annotated over ~789k windows (stride-1, highly redundant), whose cost and effect on baselines are not reported (no GPU hours, no VLM-size ablation, no baseline given an equivalent auxiliary language/annotation signal). Baselines are re-trained by the authors under a new 35-D interface, and real-robot differences per task (5–6 trials out of 30) are within binomial noise.
Questions for authors

Clarification requests:

  1. Fig. 4 (task-instruction target) and Sec. 4.3 (current frame repeated in all five positions) both report 65.18% (1095/1680). Are these distinct runs? Please provide the raw counts.

  2. How was the attribution in Fig. 3 computed, on how many samples, and were the examples selected? Can you report the fraction of attribution mass on task-relevant regions using simulator masks?

  3. Were block 22 and lambda=1.0 truly fixed before any evaluation? Were other configurations tried during development? Please report the GPU hours for annotation (Gemma generation) and for training, with GPU type.

  4. Please give results per evaluation mode (base / visual / semantic / execution) for LF vs. no-LF.

Suggestions (experiments that would affect my score):

  1. Add controls that isolate "language": (a) align to Gemma's pooled hidden state of the same frame pair (no decoding), (b) a REPA-style visual-encoder target, (c) a non-language progress target (normalized time / phase), (d) a different text encoder, (e) progress-only vs. future-only paragraphs, (f) a stronger or weaker VLM teacher.

  2. Report mean +/- std over several training seeds at least for the main comparison (LF vs. no-LF) and for the layer / lambda sweeps, including neighboring layers (e.g., 20-24).

  3. Compare against the closest methods under the same setting (Spatial Forcing-style alignment, WorldScape Policy 2.0-style event-caption supervision), and give baselines an equivalent compute/annotation budget (e.g., an auxiliary language-prediction head on GR00T/pi0.5).

  4. Show generality beyond HumanoidArena (another benchmark or embodiment) and, in real-world experiments, include OOD / language-varied tests and more trials, or otherwise tone down "Whole-Body" in the title.

6 comments
AuthorsOpenReview ↗

(1/2) Comment for Reviewer NGfW

We thank Reviewer NGfW for identifying the attribution and reliability questions. These checks are important for interpreting what the auxiliary target contributes.

“Fig. 4 (task-instruction target) and Sec. 4.3 (current frame repeated in all five positions) both report 65.18% (1095/1680). Are these distinct runs? Please provide the raw counts.”

Yes. They are distinct runs. The task-instruction target has 1,095/1,680 successes, and the repeated-current-frame setting also has 1,095/1,680 successes.

“Were block 22 and lambda=1.0 truly fixed before any evaluation? Were other configurations tried during development? Please report the GPU hours for annotation (Gemma generation) and for training, with GPU type.”

Block 22 and λ=1.0\lambda=1.0 were fixed before final evaluation. Block 22 is a common representation alignment choice: early blocks mainly encode local visual structure, while later blocks increasingly aggregate semantic information. We also swept other blocks after the main experiment to test this choice. λ=1.0\lambda=1.0 keeps the language-loss scale close to the visual-loss scale, which regularizes the video expert without allowing the auxiliary term to dominate the action and visual objectives. Gemma annotation used 4xA800 GPUs for about three wall-clock hours (12 GPU-hours). Policy training used 8xA800 GPUs for about 17 wall-clock hours (136 GPU-hours). Cached targets are computed before policy optimization.

For the requested modes, LF versus W/O LF gives 78.57%/67.86% in base, 79.76%/69.05% in visual, 80.95%/70.24% in semantic, and 77.38%/66.43% in execution, using 420 rollouts per mode.

“Add controls that isolate "language": (a) align to Gemma's pooled hidden state of the same frame pair (no decoding), (b) a REPA-style visual-encoder target, (c) a non-language progress target (normalized time / phase), (d) a different text encoder, (e) progress-only vs. future-only paragraphs, (f) a stronger or weaker VLM teacher.”

We appreciate this request because it separates language semantics from generic representation alignment.

Model / controlmodelParamsSR (%)
W/O LFGiga-World-Policy-0.56.0B68.39
Gemma hidden state, no decodinggemma-4-12B12B75.00
DINOv2 future-image targetdinov286M76.19
Normalized progress target----72.62
Different text encodert5770M77.86
Progress-only textgemma-4-12B + umt5-xxl12B + 11B77.38
Future-only textgemma-4-12B-it + umt5-xxl12B + 11B76.43
Alternative VLMQwen2.5-VL-7B-Instruct7B76.90

Together, these controls test decoding, visual features, progress, paragraph content, encoder choice, and teacher capacity rather than treating every gain as a language-semantic effect.

AuthorsOpenReview ↗

(2/2) Comment for Reviewer NGfW

We thank Reviewer NGfW for focusing on stability and comparison fairness. These points determine whether the gain is reproducible and how broadly it should be interpreted.

“Report mean +/- std over several training seeds at least for the main comparison (LF vs. no-LF) and for the layer / lambda sweeps, including neighboring layers (e.g., 20-24).”

CheckValuesSummary
Seeds 42/43/44, W/O LF68.39 / 68.75 / 67.9268.35 ±\pm 0.42
Seeds 42/43/44, W/ LF79.17 / 78.75 / 80.0079.31 ±\pm 0.64
Gain per seed10.78 / 10.00 / 12.0810.95 ±\pm 1.05 pp
Blocks 20--2477.50 / 78.33 / 79.17 / 78.75 / 78.33range 77.50--79.17
λ=0/0.5/1/1.5/2\lambda=0/0.5/1/1.5/268.39 / 74.40 / 79.17 / 73.39 / 68.75peak at 1.0

The LF standard deviation is 0.64 points and the gain remains positive for every seed. Neighboring blocks vary by 1.67 points, so the result is not confined to a single isolated layer.

“Compare against the closest methods under the same setting (Spatial Forcing-style alignment, WorldScape Policy 2.0-style event-caption supervision), and give baselines an equivalent compute/annotation budget (e.g., an auxiliary language-prediction head on GR00T/pi0.5).”

The Spatial-Forcing-style comparison is the DINOv2 control reported in our response to Reviewer 8hce: it reaches 76.19% on open-gigaai/Giga-World-Policy-0.5 with the same projector, block, loss weight, data, and update budget. We also tested an auxiliary language-prediction head on yuanty/fastwam; its average score is 67.50%, compared with 61.13% for yuanty/fastwam without the head and 70.00% with Language-Forcing.

WorldScape Policy 2.0 is fundamentally different from Language-Forcing, so we do not present a misleading drop-in numerical comparison. WorldScape supervises planning representations with event-level captions. Language-Forcing uses paired current and ground-truth future observations to produce transition-specific descriptions of task progress and likely next state, then aligns those descriptions with action-conditioned future-visual features. The target, supervised representation, and role of the action trajectory are different; this is a conceptual difference rather than a change of teacher or prompt.

“Show generality beyond HumanoidArena (another benchmark or embodiment) and, in real-world experiments, include OOD / language-varied tests and more trials, or otherwise tone down "Whole-Body" in the title.”

We thank Reviewer NGfW for raising the benchmark and scope question. It is important for judging how far the evidence should be generalized. Humanoid loco-manipulation benchmarks are scarce. HumanoidArena already combines seven tasks with four evaluation modes: base conditions, lighting and viewpoint changes, semantic asset and distractor changes, and expanded execution pose ranges. These modes provide structured OOD tests within the benchmark, including randomized lighting, texture/object variants, distractors, and expanded object poses. Our setting is humanoid loco-manipulation: the robot must move and manipulate, so success depends on coordinated legs, torso, arms, hands, contact, and object state.

The language-grounding claim is therefore scoped to transition-specific supervision under fixed task instructions. The term “whole-body” describes the humanoid policy and the coupled loco-manipulation requirement, not a claim of broad instruction generalization. We report the average-score claim precisely and avoid implying that LF is best on every individual task.

Thank you for the detailed response and the additional experiments. The annotation cost is smaller than I expected, and the multi-seed and neighboring-layer results are helpful. Two concerns remain.

  1. Novelty and the role of language. In the new controls, the DINOv2 future-image target (no language) already reaches 76.19%, i.e., about 7.8 of the 10.8 points gained by LF, so the language-specific gain is around 3 points. The per-mode results show almost the same gain in all four modes, with no extra benefit in the semantic mode, where semantic alignment should help most. Together with the fixed task instructions in HumanoidArena, I still see the method mainly as a feature-alignment regularizer with a language-derived target, close to existing representation-alignment work.

  2. Raw results and evaluation budget. The per-mode success counts (out of 420) are 330 / 335 / 340 / 325 for LF and 285 / 290 / 295 / 279 for W/O LF, i.e., the gap is 45 / 45 / 45 / 46 successes across the four modes. This is unusually regular given the binomial variance of 420-rollout evaluations. In addition, it is not stated how many rollouts were used for the new experiments: all new seed and layer results correspond to an integer number of successes out of 240, and all new control results correspond to an integer number out of 420, whereas the paper's protocol uses 1,680 rollouts per method. Could you please: (a) state the number of rollouts used for each new experiment (seeds 43/44, blocks 20/21/23/24, and the controls); (b) provide per-task, per-mode success counts for LF and W/O LF for each seed; (c) provide the episode-level evaluation logs (e.g., as supplementary material or an anonymous link).

Minor: the response describes the visual mode as including viewpoint changes, but Appendix A.1 and Table 7 describe only randomized directional lighting. Please clarify.

Comments on the paper

1 comment

We thank the Area Chair and all three reviewers for their careful evaluation. Language-Forcing adds transition-specific task-progress supervision to a humanoid World Action Model (WAM), improving whole-body control without increasing inference cost. We summarize the reviewers' assessments and our responses below.

Reviewer 6AFM: a meaningful direction with transfer across backbones. We thank Reviewer 6AFM for recognizing the improvements in simulation and on real robots, and the evidence that language-target content and correspondence matter. After our response, the reviewer explicitly stated that their concerns were addressed, described LF as a “simple but meaningful direction for improving humanoid WAM,” and indicated an intention to raise their score.

Their main questions concerned novelty, compatibility, and the whole-body setting. We clarified the fundamental distinction between WorldScape's event-caption supervision of planning representations and LF's transition-specific progress and next-state supervision of action-conditioned future-visual features. We also showed that LF improves fastwam and lingbot-va, which the reviewer recognized as evidence of generalizability. We explained why coordinating locomotion, stance, contact, and manipulation makes this supervision useful for humanoids, and will strengthen this explanation in the main text as requested.

Reviewer 8hce: dense supervision, a clean comparison, and consistent task improvements. We thank Reviewer 8hce for highlighting that LF supplies descriptions that change with each transition, giving denser task-level supervision at no inference cost. The reviewer also recognized the clean LF/no-LF comparison and improvements across all evaluated simulation and real-world tasks.

For their main concern about the language target, we added the requested DINOv2 control with the backbone, alignment interface, data, and training schedule held fixed. LF performs better than the matched visual target, supporting the value of transition-specific text supervision. We also explained how the training objective can guide predictive features toward task-relevant changes, while keeping deployment unchanged.

Reviewer NGfW: controlled evidence and reliability. We thank Reviewer NGfW for recognizing the matched comparison, zero inference overhead, and real-robot evidence, and for finding the multi-seed and neighboring-layer results helpful. We responded to the language-specific and reliability questions with additional target controls, stability checks, and detailed evaluation records for verification.

Language-Forcing gives a predictive robot model explicit supervision about what advances the task. This is especially useful for whole-body humanoids: walking, maintaining stance, making contact, and moving an object must work together, while the robot's motion also changes much of the visible scene. Descriptions based on current and future observations summarize progress and likely next-state changes at each transition. Aligning these descriptions with action-conditioned future-visual features provides a practical link between predicting the scene and learning coordinated locomotion and manipulation.

The practical value is that this richer supervision is used entirely during training. The deployed policy gains no additional teacher, text encoder, or inference steps. Improvements in simulation, on real robots, and across multiple WAM backbones support LF as a useful training strategy for humanoid control. Together, the supervision design and unchanged deployment cost make it a simple, meaningful way to improve existing WAMs.

We sincerely thank the Area Chair and all reviewers again for their time, careful evaluation, and constructive feedback, which have helped us strengthen the paper and communicate its contribution more clearly

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.