acceptodds
Under review as a conference paper at ICLR 2027

Should the Action Expert Train the VLM? Action-Token Supervision and Gradient Routing in Mixture-of-Transformer VLAs

Abstract

Vision-language-action (VLA) models increasingly pair a pre-trained vision-language model (VLM) with a flow-matching expert that conditions its action predictions on the VLM's key-value representations. Those representations must carry information useful for control, yet the VLM was never trained for the action expert's flow-matching objective. Two signals can address this: gradients backpropagated from the expert's loss (coupling), or VLM supervision to predict discrete tokens encoding the same target actions. Prominent training recipes forgo the first, blocking expert-to-VLM gradients (insulation) to protect the VLM's existing capabilities, while also supervising it on vision-language data and action tokens. Because vision-language co-training also aims to preserve those capabilities, the additional protection provided by insulation is unclear. At the same time, insulation removes a direct signal about which features continuous action prediction requires. On the other hand, action-token supervision offers an alternative way to learn action-relevant representations, which may explain why insulation succeeds in practice. We disentangle these factors in autonomous driving and manipulation by independently varying action-token supervision and whether the expert's loss updates the VLM. In open-loop driving, insulation is competitive with coupling when action tokens are supervised, but increases trajectory-prediction error by roughly 25% when they are not. When both components are trained jointly from the outset, coupling increases closed-loop driving route progress by approximately 17% with action-token supervision and 40% without it, and similarly improves robot manipulation performance under both settings. These findings show that the value of insulation depends on the VLM's action supervision and the evaluation setting, challenging insulation as a default VLA training recipe.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, Mixture-of-Transformer

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

This paper investigates how action-token supervision affects the choice of whether gradients from a continuous action expert should update the VLM backbone. The authors independently vary action-token supervision and expert-to-VLM gradient flow, comparing coupled and insulated training across three training schedules in autonomous driving, with an additional evaluation in robot manipulation. The results show that action-token supervision reduces the open-loop performance gap between coupling and insulation, whereas coupling improves closed-loop driving and manipulation performance under concurrent training even when action tokens are supervised. The study provides empirical evidence that the benefits of insulation depend on the surrounding training recipe and evaluation setting.

Critical strengths
  1. The paper addresses a practically important training question: whether the action expert should update the VLM, and how this choice depends on action-token supervision. By varying these factors independently within a shared architecture, particularly in the concurrent-training setting, it helps disentangle design choices that are often bundled together in existing VLA recipes.

  2. The evaluation examines multiple training schedules, open- and closed-loop driving, and robot manipulation. Notably, a small open-loop difference with action-token supervision does not imply equivalent closed-loop performance: coupling still improves driving progress under concurrent training. The manipulation results provide additional evidence beyond driving, making the findings useful for selecting VLA training strategies.

Critical weaknesses
  1. The evaluation does not fully establish the trade-off between improved control and preservation of VLM capabilities. Lingo-Judge measures driving-rationale quality, but does not assess broader visual-language abilities that insulation is intended to protect. Moreover, coupling does not consistently improve safety alongside route progress: under VLM-first training with action-token supervision, progress is nearly unchanged, while the off-road-or-at-fault-collision rate increases from 0.203 to 0.282. Although these limitations are acknowledged or reported, they constrain broader recommendations about the benefits of coupling.

  2. The training-history comparisons do not isolate training order from differences in component-level optimization and data exposure. Despite matching total optimizer steps, the action expert receives approximately 145.5K, 197K, and 97K updates under concurrent, expert-first, and VLM-first training, respectively. In addition, the VLM-first warm-up uses different data mixtures depending on whether action-token supervision is enabled. These comparisons remain useful as evaluations of complete training recipes, but provide weaker evidence that training order itself explains the observed differences.

Questions for authors
  1. Preservation of VLM capabilities: Could you evaluate the coupled and insulated checkpoints on a held-out general vision-language benchmark, ideally including the pretrained VLM as a reference? Lingo-Judge measures driving-rationale quality, but a broader evaluation would clarify whether the control gains from coupling come at the expense of capabilities that insulation is intended to preserve.

  2. Progress–safety trade-off: How should the route-progress improvements be interpreted alongside the safety results in Tables 10–12? For example, under VLM-first training with action-token supervision, coupling leaves progress nearly unchanged while increasing the off-road-or-at-fault-collision rate from 0.203 to 0.282. Could you clarify which settings support a recommendation for coupling when both progress and safety are considered, and whether differences in rollout duration or termination affect these comparisons?

  3. Training-history comparison: Can you provide a targeted comparison that better separates training order from component-level update budgets and data exposure, for example by extending VLM-first training to match the concurrent model’s expert-update count? Please also clarify how the different warm-up data mixtures in the VLM-first action-token-on/off conditions affect the interpretation. If such controls are unavailable, which conclusions should be understood as comparisons of complete training recipes rather than effects of training order itself?

1 comment

We thank the reviewer for recognizing the practical importance and soundness of our controlled study. We address each point below.

1. Capability retention and safety

Weakness (summarized): Rationale scores do not establish broader retention, and progress gains do not consistently improve safety.

We agree that driving-rationale quality alone cannot establish broader retention. We added a seven-benchmark VQA evaluation of all twelve driving checkpoints and pre-trained Qwen (Section 4.4; Appendix F; general comment, “VQA Retention Across Training Recipes”). This complements Actions as Language [1], cited in our original Related Work, which reported poor VQA retention despite insulation and vision–language co-training.

We also agree that VLM-first training with action-token supervision does not favor coupling: progress is essentially unchanged, while off-road-or-at-fault collisions increase from 0.203 to 0.282. We challenge insulation as a default; we do not recommend coupling across all settings.

2. Interpretation of training histories

Weakness (summarized): Component-update budgets and warm-up data differ across histories, so training order is not isolated.

Our purpose is to test routing across commonly used MoT-VLA training strategies, not to isolate the causal effect of training order. The highest observed route progress under concurrent training is a result among evaluated recipes, not evidence that order alone explains the ranking.

As reported, expert updates differ across concurrent, expert-first, and VLM-first training (approximately 145.5k, 197k, and 97k). Within each history and token setting, coupling-versus-insulation comparisons match initialization, training data, update opportunities, and optimization schedules.

VLM-first token-on/off warm-ups change both supervision and data exposure, so that comparison cannot isolate supervision alone. The supervision–routing interaction also appears under concurrent and expert-first training with fixed training data; the VLM-first warm-up difference cannot explain those results.

3. General VQA results

Question (summarized): Can you evaluate general VL capabilities with the pre-trained VLM as a reference?

Both coupled and insulated checkpoints score substantially below pre-trained Qwen on all seven benchmarks across all three training histories. Relative performance varies by benchmark and recipe; insulation scores higher in some comparisons but does not prevent substantial degradation. Many scores are near the floor, limiting sensitivity to smaller routing differences. These results do not establish equivalent retention. Full results appear in the general VQA comment and Appendix F.

4. Progress–safety tradeoffs and termination

Question (summarized): Which settings favor coupling on progress and safety, and do rollout durations or termination affect comparisons?

Without action-token supervision, coupling increases progress and lowers observed off-road-or-at-fault-collision rates across all three histories. This is our clearest evidence favoring coupling on both measures.

With supervision, outcomes depend on the recipe. Concurrent coupling increases progress from 0.600 to 0.703 and lowers off-road-or-at-fault collisions from 0.248 to 0.226. Overall collision and off-road rates also decrease, although at-fault collisions are slightly higher and some tracking metrics favor insulation. Expert-first coupling increases progress but also at-fault collisions. VLM-first coupling leaves progress essentially unchanged and worsens the combined safety outcome. These results support recipe-specific decisions.

All models use the same 827 scenes and five rollouts per scene. Scoring stops at the first collision, off-road event, or 4 m deviation from the reference trajectory. Stopping rules are identical, but scored durations can differ. Safety failures affect progress; safety rates should not be interpreted as comparisons over equal driving time.

5. Matching expert-update budgets

Question (summarized): Can you extend VLM-first training to match concurrent expert updates? Which conclusions concern complete recipes?

We have not run this extension due to compute limits; each full driving run costs approximately 15,000 H100 GPU-hours. Matching concurrent expert updates would require approximately 100k additional joint steps, bringing VLM-first training to 400k total steps. This would help assess the smaller expert-update budget, but also increase VLM updates and would not isolate order alone.

Rankings across training histories compare complete recipes. Routing comparisons remain controlled within each recipe and token setting. Likewise, VLM-first token-on/off differences cannot be attributed solely to supervision because warm-up data exposure changes.

[1] Hancock et al. Actions as Language. https://arxiv.org/abs/2509.22195

Reviewer_G9HJOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes how gradients from the action expert affect the VLM backbone in VLA models. Specifically, it studies whether backpropagating the expert’s flow loss into the VLM helps which they call coupling, compared with blocking it, which they call insulation. The authors claim that insulation appeared effective in prior work because those works also supervised the VLM with discrete action tokens, which partially substitutes for the expert’s gradient and masks the cost of insulation. They show the results by crossing action-token supervision with coupling/insulation and evaluating driving models under three training schedules with open-loop ADE, closed-loop AlpaSim, and rationale quality as well as robot manipulation models in the PolaRiS simulator.

Critical strengths
  • The paper helps to analyze practical question of how to train a VLA model, and it’s relationship with the VLM backbone and the action expert model. By evaluating action-token supervision and gradient routing while holding all other conditions constant, the paper isolate the individual effects and interactions of both factors for the first time.
  • This paper addresses two embodiments, autonomous driving and robot manipulation, and evaluates both open-loop and closed-loop setups. For the training schedule as well, it specifically utilizes π0.5\pi_{0.5}/MolmoAct2 for the VLM-first order and APT for the expert-first order, demonstrating that the results are not restricted to a specific training scheme.
Critical weaknesses
  • The original purpose of insulation was to maintain the performance of VLM backbone model, but this is only experimented by a Lingo-Judge evaluation, where the confidence intervals largely overlap. The paper claims to prove that insulation provides little protective effect, which is too strong of a claim compared to the experiment. Also, the staged Lingo-Judge results in Table 2 sometimes favor insulation without tokens.
  • In co-training with action tokens, insulation yields better ADE, distance to ground truth (2.38 vs. 2.70), and wrong-lane rates (0.20 vs. 0.23). In contrast, coupling achieves higher route progress and driven distance (139.6 m vs. 116.7 m). Route progress may reward more aggressive driving behaviors, and concluding that "coupling is superior in closed-loop" without an adjusted composite metric overlooks this trade-off.
  • The token-off interventions differ by domain. Driving keeps teacher-forced future-value tokens in the sequence, while manipulation removes the FAST postfix. The VLM-first tokens-off warm-up also drops trajectory data entirely. How do you rule out that these differences, rather than supervision itself, drive the cross-schedule and cross-domain effects?
Questions for authors
  1. Is the benefit of coupling specific to the flow gradient, or does it simply reflect more total action-relevant gradient reaching the VLM? A control that increases the action-token CE weight under insulation, or reports relative gradient norms at the K/V interface, would distinguish “the expert’s gradient carries unique information” from “the VLM just needs a stronger action signal.”
  2. Can you report a general VL capability measure, such as a held-out VQA/captioning benchmark or LingoQA accuracy, for all conditions, and a language-following evaluation in manipulation? Without these, the claim that insulation offers “little additional protection” should be softened.
  3. Table 5 lists a 400k-step cosine schedule, but driving checkpoints are taken at 300k, so the learning rate has not annealed. Was this intended? Could it interact with routing, for example if coupled runs are more sensitive to a high late-stage learning rate?
  4. Figure 4 shows MoveLatteCup, but the evaluated environments are BlockStackKitchen, FoodBussing and PanClean. Which environments were actually used for Table 4?
1 comment

We thank the reviewer for their feedback.

1. VLM capability retention

Weakness (summarized): Insufficient retention evaluation; staged rationale scores sometimes favor insulation.

Our original evaluation scope was deliberate: as discussed in Related Work, Actions as Language [1] reported poor VQA retention in π₀.₅ despite insulation and vision–language co-training. The manuscript explicitly recognized that insulation need not preserve general VLM capabilities. We asked whether its benefits justify blocking a direct action-learning signal.

We now evaluate all twelve driving checkpoints and pre-trained Qwen on 9,185 identical examples across seven VQA benchmarks (Section 4.4; Appendix F; general VQA comment). Both routing choices score substantially below pre-trained Qwen on all seven benchmarks across all three histories.

Some staged rationale comparisons favor insulation. Overlapping intervals do not establish equivalence, and near-floor VQA scores limit sensitivity to smaller differences. Insulation does not prevent substantial degradation in our setting; this does not establish that it provides no protection.

2. Progress, safety, and tracking

Weakness (summarized): Progress gains do not establish closed-loop superiority.

The original manuscript claims improved route progress under concurrent training, not superiority across all closed-loop metrics. Tracking outcomes favoring insulation are explicitly reported; these tradeoffs are part of our evidence.

With action-token supervision, coupling increases progress from 0.600 to 0.703. Observed overall collision rates decrease from 0.225 to 0.164, off-road rates from 0.161 to 0.136, and off-road-or-at-fault-collision rates from 0.248 to 0.226. At-fault collisions rise slightly (0.087 to 0.090), while ADE, distance to ground truth, and wrong-lane rate favor insulation.

AlpaSim stops scoring after the first collision or off-road event, but progress does not replace safety metrics. Separate metrics expose tradeoffs without imposing composite weights. We challenge default insulation; we do not prescribe universal coupling.

3. Token-off interventions

Weakness (summarized): Token handling and warm-up data differ across settings.

Token handling differed deliberately to follow each method’s native implementation described in the literature. Within each setting, coupled and insulated conditions use identical token handling. Driving masks future-action CE while retaining teacher-forced tokens; manipulation removes the FAST postfix. The driving expert’s K/V cache ends before future-action values, preventing target access. The pre-mask CE denominator stays fixed.

VLM-first token-off warm-up also removes trajectory data, so it cannot isolate supervision alone. The interaction also appears in concurrent and expert-first driving, where data and token handling are fixed and only future-action CE is disabled. The warm-up difference cannot explain those results. Cross-domain agreement supports a recurring pattern under the respective implementations, not an identical intervention.

4. Gradient strength versus information

Weakness (summarized): Does coupling simply strengthen action gradients?

That is consistent with expert gradients and action-token supervision partly substituting for one another. Our conclusion does not require unique information in flow gradients. Our objectives and loss weights are fixed. Higher action-token CE weights and K/V-interface gradient norms remain untested; magnitude versus information content remains unresolved.

5. Manipulation language following

Question (summarized): Evaluate general VL capabilities and manipulation language following.

The VQA evaluation addresses general VL capabilities. Manipulation language-following retention remains unmeasured. Fixed tasks do not readily support alternative instructions requiring independently scored behaviors in the same scene. Paraphrase success alone is insufficient: an instruction-ignoring policy could complete the task. We do not interpret task progress as language-following retention.

6. Learning-rate schedule

Question (summarized): Could the 300k endpoint with a 400k cosine horizon affect routing?

The endpoint was intentional; validation performance had stabilized. Learning rates had decayed but had not reached their minimum. Routing conditions share schedules within each history, but routing–annealing interactions remain possible. Alternative schedules remain untested; each driving run costs approximately 15,000 H100 GPU-hours.

7. Evaluated tasks

Question (summarized): Why does Figure 4 show MoveLatteCup?

Table 4 equally averages progress across BlockStackKitchen, FoodBussing, and PanClean. MoveLatteCup is qualitative and excluded from that average, as the revised caption states.

[1] Hancock et al. Actions as Language. https://arxiv.org/abs/2509.22195

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

This paper asks whether the action module should also train the vision-language model it relies on. Some popular methods block that feedback to protect the model’s existing abilities, but they also teach it to predict action tokens, giving it another way to learn about actions.

The authors separate these two choices and test them in driving and robot manipulation. They find that blocking feedback hurts most when action-token training is also absent, while allowing feedback often improves closed-loop performance even when action tokens are used. The takeaway is that blocking action gradients shouldn’t be the default: its value depends on the rest of the training setup.

Critical strengths
  • It identifies a real gap in previous comparisons: blocking action gradients was often bundled with action-token training, so it was unclear which part actually helped. Separating those choices makes the paper’s main argument convincing.

  • The experiments are well controlled. Within each training schedule, the authors keep the model, data, initialization, and optimization setup fixed, making it easier to attribute differences to the two choices they study.

  • It evaluates actual closed-loop behavior, not just trajectory prediction. The finding that similar prediction errors can lead to quite different driving performance is useful for choosing training recipes, and the manipulation experiments show the issue extends beyond driving.

Critical weaknesses
  • The paper doesn’t fully test the main reason people block action gradients: protecting the vision-language model’s existing abilities. It mainly checks whether the model produces reasonable driving explanations, but that’s much narrower than testing its general visual understanding, language skills, or reasoning. Better control performance could still come with losses elsewhere. The key question is: does allowing action gradients preserve those broader abilities, or does it trade some of them for better control?

  • Each condition is trained only once. The confidence intervals account for differences across evaluation scenes, but not for variation from training randomness. That matters especially when two methods score close to each other, or when the paper argues that training both modules together is the best starting point. Would those rankings hold across several training runs? The larger gains are encouraging, but the smaller differences need stronger evidence.

  • The practical evidence is still limited to simulation, with one model family per domain and only three manipulation tasks. Also, making more progress along a driving route doesn’t necessarily mean driving more safely. The main question is whether allowing gradients still helps on real robots and across broader tasks, without increasing collisions or other unsafe behavior. The current results make a good case against treating gradient blocking as the default, but not yet for treating unrestricted gradient flow as the default either.

Questions for authors

see weakness

1 comment

We thank the reviewer for their feedback and for recognizing that our results challenge gradient insulation as a default. We address each concern below.

1. Retention of pre-trained VLM capabilities

The paper doesn’t fully test the main reason people block action gradients: protecting the vision-language model’s existing abilities.

As discussed in Related Work, Actions as Language [1] reported poor VQA retention in π₀.₅ despite insulation and vision–language co-training. Our original manuscript therefore recognized that insulation need not preserve general VLM capabilities and focused on whether its benefits justify removing a direct action-learning signal. We agree that measuring retention in our own checkpoints strengthens this analysis.

We now evaluate all twelve driving checkpoints and pre-trained Qwen on seven general VQA benchmarks. Both coupled and insulated models score substantially below pre-trained Qwen on all seven benchmarks across all three training histories. See Section 4.4, Appendix F, and our general comment, “VQA Retention Across Training Recipes.” Insulation does not prevent substantial degradation in these benchmark scores; many scores are near the floor, limiting sensitivity to smaller routing differences.

Under concurrent training with action-token supervision, coupling increases route progress from 0.600 to 0.703, with similar observed driving-rationale scores. Insulation scores higher on some VQA benchmarks, indicating benchmark-specific tradeoffs, but neither approach maintains pre-trained VQA performance. These findings support evaluating insulation within the specific training recipe.

2. Training-seed variation

Each condition is trained only once. The confidence intervals account for differences across evaluation scenes, but not for variation from training randomness.

We agree that our experiments do not establish whether small differences or recipe rankings persist across seeds. Each driving run uses 64 H100 GPUs and approximately 15,000 GPU-hours on average (Appendix A). Repeating the full factorial study across seeds is beyond our rebuttal compute resources.

Our main conclusion rests on the recurring interaction between action-token supervision and routing. Without action-token supervision, coupling substantially reduces open-loop error across both driving evaluation sets and all three training histories; with supervision, routing differences are much smaller. Both manipulation model families likewise show larger gains from coupling without action-token supervision. This recurring pattern supports questioning insulation as a default, but does not establish robustness to training randomness.

Our bootstrap intervals quantify variation across evaluation scenes for fixed checkpoints, not training runs. We interpret small differences cautiously and describe concurrent training as achieving the highest observed route progress among the evaluated recipes, rather than the best recipe across seeds. Our limitations explicitly acknowledge unmeasured training-seed variation.

3. Evaluation scope and safety

Reviewer concern (summarized): simulation-only evidence, limited architectures and tasks, and whether progress gains imply safer behavior or justify coupling as a default.

Our evidence remains simulation-only, but the revision adds a second manipulation model family, MolmoAct2 [2]. Using its released implementation, we vary action-token supervision and routing under concurrent training. Coupling increases task progress from 22.0% to 24.3% with action tokens and from 14.5% to 36.5% without them. Both MolmoAct2 and π₀.₅ show larger coupling gains without action-token supervision. The task set remains unchanged; physical-robot validation remains future work.

We agree that route progress alone does not establish safety and report safety and tracking metrics separately. Without action-token supervision, coupling increases progress and lowers the observed off-road-or-at-fault-collision rate across all three training histories. With supervision, concurrent coupling improves progress and lowers observed overall collision and off-road rates, whereas staged coupling increases at-fault collision rates. AlpaSim truncates scoring after the first collision or off-road event, so failures affect progress, but progress does not replace dedicated safety metrics.

These results support evaluating routing jointly with action-token supervision, the training recipe, and the intended performance and safety metrics. We challenge insulation as a default; we do not propose coupling as a universal replacement.

References

[1] Hancock et al. Actions as Language. 2025. https://arxiv.org/abs/2509.22195

[2] Fang et al. MolmoAct2: Action Reasoning Models for Real-World Deployment. 2026. https://arxiv.org/abs/2605.02881

Comments on the paper

1 comment
AuthorsOpenReview ↗

VQA Retention Across Training Recipes

We thank all reviewers for their thoughtful feedback. We have revised the manuscript, with changes and additions marked in blue in the updated PDF.

All three reviewers requested a broader evaluation of pre-trained VLM capability retention. We therefore evaluate all twelve fine-tuned driving checkpoints, spanning different training histories, alongside the original pre-trained Qwen-3-VL model on the same 9,185 examples across seven benchmarks: VQAv2, ScienceQA, TextVQA, ChartQA, DocVQA, AI2D, and TabMWP. This lets us compare general VQA performance before and after driving fine-tuning.

We have added this evaluation to Section 4.4 of the revised manuscript. Appendix F includes the full results, scoring details, and examples of saved model answers.

In the tables below, Ins. denotes insulated routing, Coup. denotes coupled routing, and On/Off indicates whether future-action-token supervision is used. Qwen denotes the original pre-trained Qwen-3-VL model.

General VQA — Concurrent Training

DatasetQwenIns. / OnCoup. / OnIns. / OffCoup. / Off
VQAv278.350.000.510.000.05
ScienceQA87.0136.0435.8935.3535.50
TextVQA79.320.000.000.000.00
ChartQA79.880.000.000.000.00
DocVQA90.810.000.000.000.00
AI2D76.0727.3426.1724.7126.86
TabMWP53.2212.409.869.1811.52

General VQA — Expert-First Training

DatasetQwenIns. / OnCoup. / OnIns. / OffCoup. / Off
VQAv278.350.820.140.988.58
ScienceQA87.0135.1036.6435.3535.85
TextVQA79.320.210.000.001.15
ChartQA79.880.000.000.000.00
DocVQA90.810.000.000.000.14
AI2D76.0724.8026.8624.7126.66
TabMWP53.229.6710.6410.1612.30

General VQA — VLM-First Training

DatasetQwenIns. / OnCoup. / OnIns. / OffCoup. / Off
VQAv278.350.440.085.410.06
ScienceQA87.0135.4535.6535.4536.19
TextVQA79.320.000.200.000.00
ChartQA79.880.000.000.000.00
DocVQA90.810.070.000.000.00
AI2D76.0724.6126.7627.0526.37
TabMWP53.229.2810.8410.2511.62

Scores are percentages, with higher values indicating better performance. All models are evaluated on identical examples under the same settings. ScienceQA and AI2D are scored using option-label likelihood; TabMWP uses option likelihood for multiple-choice questions and generated answers otherwise. All generated answers have a ceiling of 64 new tokens; pre-trained Qwen reached this limit on 205 of the 1,024 TabMWP examples.

We also ran a post-hoc check of answer formatting. For example, this check scores an answer beginning with “Yes, ...” as “yes,” rather than scoring the whole explanation. We then apply the same benchmark scoring rule to the extracted answer. This raises the driving checkpoints’ VQAv2 scores to 1.28–22.24%, compared with 78.35% for pre-trained Qwen. Answer formatting therefore contributes to the low scores, but a substantial gap remains. The original benchmark scores remain our primary results; Appendix F reports the extraction check separately. We also include qualitative examples of VQA ability in Appendix F.

In summary, both coupled and insulated models perform substantially worse than pre-trained Qwen on all seven benchmarks and under all three training histories. Knowledge (gradient) insulation therefore does not prevent substantial degradation in benchmark performance under our evaluated training recipes. Relative performance varies across benchmarks and training recipes, with neither routing choice consistently outperforming the other. Many scores are near the floor, limiting sensitivity to smaller differences between routing choices. Together with our control results, these findings reinforce our conclusion that gradient insulation should be evaluated in the context of the training recipe and evaluation setting, rather than adopted as a default.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.