acceptodds
Under review as a conference paper at ICLR 2027

Silent Collapse: Diagnosing and Repairing Low-Precision Robot Policy Training at the Action Level

Abstract

Robot learning is now trained largely with low precision, bf16 and fp8 alike. Low precision removes information the policy needs during training, and this can be silent: every monitor reads normal, yet the robot no longer produces the right action. We study this failure at the level of the executed action. It is not one failure mode but three, concentrated where the signal is weak along the activation gradient chain. The two most dangerous ones are new: mantissa starvation and conservative collapse. Both rewrite the action output without warning, with striking unevenness across tasks and seeds. We build a seconds-long probe locating the weak-signal points where failures start and present CURE (Cheapest sUfficient REpair), a routing protocol assigning each failure mode its cheapest sufficient fix: a free analytic bypass where the weak signal has a closed-form gradient, exact backward fidelity from the UPH (Unified Precision Health) layer where it does not, and LockedAdamW, an optimizer locking the per-step noise-displacement budget. We test the routing across policy families from a 0.8M ACT to 7B vision-language-action models under the twin-run protocol. On the ACT collapse cell, the routed scheme brings the collapsed posterior’s KL back into the fp32 band; on the 7B CogACT, LockedAdamW breaks below the fp32 self-retraining floor on both seeds; on real training hardware, one tested Transformer Engine fp8 configuration trains normally—consistent with the same mechanism—while unscaled variants fail in simulation. The probe answers in 3.4 seconds at 0.8M and in minutes at 7B; the repair starts free.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Low-precision training, Robot learning, Numerical stability, Training reliability

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_VxJ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies "silent failures" in low-precision robot policy training: the loss curve and overflow monitors all read normal, yet the action output has already collapsed. The paper identifies three failure modes, localizes the root cause to a self-reinforcing loop in the backward activation-gradient chain, where weak signals are swallowed by rounding and then amplified by Adam, and proposes a 50-step twin probe for diagnosis and the CURE routing protocol for repair, validated on a matrix spanning 0.8M ACT to 7B VLA models.

Critical strengths
  1. Valuable motivation: silent action-level failure is a real and overlooked problem, and attributing the posterior collapse of CVAE policies to the numerical layer, with recovery by precision means alone, is a meaningful observation.
  2. Careful mechanism design: the item-by-item exclusions and the "zero-mean noise vs. equal-power toward-zero bias" control experiment are clean, and the rescue-band prediction is falsifiable and confirmed.
  3. Honest reporting: the limitations section explicitly acknowledges the absence of real-robot validation, probe false positives, the lock's domain boundary, and a single harness per 7B model.
Critical weaknesses
  1. Severely insufficient sample sizes: most conclusions rest on 1–3 seeds per arm, and the authors themselves acknowledge that lottery outcomes change with the harness; the statistical reliability of the key conclusions is inadequate.
  2. Weak probe validation: only 7 cells, with a bootstrap 95% CI of [0.38, 1.00] for the 0.86 rank correlation, nearly including zero; the prospective routing validation covers only 2 cells.
  3. Anecdotal evidence at scale: the full quantitative picture comes almost entirely from the 0.8M ACT; at 7B, CogACT has only 2 seeds on a single harness, the OpenVLA-OFT fp32 reference itself went NaN (making that cell uninterpretable), and the method misfires on GR00T.
  4. Limited behavioral evidence: rollout success is resolvable on only one cell, the primary metric µdiv admits only directional readings below the floor, and the real-hardware fp8 validation covers a single configuration, at anecdotal level.
  5. Weak necessity of the repairs: the authors themselves show that four one-line changes (Kahan, ε, β2, clipping) are equally effective, so the incremental advantage of LockedAdamW and UPH is unsupported.
Questions for authors
  1. Can the core cells be expanded to sufficient seeds (≥5–10) and re-verified on two harnesses?
  2. Can the conservative collapse on CogACT be reproduced on more large models and seeds?
  3. Can the confidence interval of the probe correlation be tightened with added samples?
  4. Can the incremental gains of LockedAdamW/UPH over one-line fixes and global stochastic rounding be quantified?

Note: the motivation and mechanism analysis of this paper are valuable, but the current experimental reliability is insufficient. The score may be raised if the authors provide validation with sufficient seeds and large-scale models across adequate samples, and tighten the probe statistics.

Reviewer_muLZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper identifies three numerical failure modes in low-precision robot policy training: overflow, mantissa starvation, and conservative collapse, which can degrade actions despite normal training metrics. It introduces CURE, a framework combining a short diagnostic probe with targeted repairs: analytic gradient bypass, selective high-precision backward computation (UPH), and LockedAdamW. Experiments across models from 0.8M to 7B parameters demonstrate these failure mechanisms and evaluate repair effectiveness and computational overhead. To be clear - I find most of this LLM generated and I'm only trying to summarize what authors have written as opposed to taking a stance on the summary

Critical strengths
  1. The idea of investigating effect of floating point errors is a decent hypothesis which has showed given some meaningful insights in the LLM space, and does beg a scientific, concrete investigation to robot policies.
Critical weaknesses
  1. This paper uses extremely heavy use of LLMs to the extent where it makes reading the paper almost impossible. Despite this, I still read the paper once. This is a major weakness which I believe cannot be addressed during the rebuttals -- and based on this reason alone, I am voting for a rejection of this paper. What makes me believe this has heavy LLM usage? To name a few examples:
    • the seed lottery

    • the hidden brake

    • the starvation loop

    • the posterior dies

    • the mantissa is eaten

    • the swallowed volume

    • the lottery pocket

The excessive usage of the word decisive: "A decisive test follows", "The decisive cell", ...

  1. The figure is very likely to be VLM generated -- and incredibly hard to understand. It has too much text that is not contextualized. The manuscript introduces numerous failure modes, schemes, metrics, abbreviations, and experimental regimes. Key empirical findings and exceptions are scattered across the main text and appendices. A consolidated comparison of diagnostic predictions, repair success, task returns, and computational costs would make the central conclusions substantially clearer.

  2. The 50-step probe achieves a Spearman correlation of 0.86, but on only seven cells with a wide confidence interval [0.38, 1.00]. Prospective routing is tested on just two cells, with a false positive on Octo, leaving its robustness across models and hardware unclear.

  3. Validation relies heavily on action divergence rather than task success, although similar divergence can yield very different returns (7.0 vs. 22.4). On MH-Lift, the combined repair still collapses in 1/6 runs versus 0/3 for fp32, raising concerns about actual behavioral reliability.

Questions for authors

See weaknesses.

Reviewer_JeN8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents silent numerical failures in low-precision robot policy training, their causes, and proposes repair protocols to ensure reliable behaviour. Low-precision training silently removes critical signals, causing action drifts without warning, and the paper identifies three failure modes that emerge at weak-signal points: overflow, mantissa starvation, and conservative collapse.

The paper shows that overflow is detectable via overflow counters; mantissa starvation and conservative collapse are silent. Proposed criteria run in seconds to predict failures by measuring activation and gradient signals during short twin runs. Validation ladders are ACT (0.8M), evaluated on Square, Can, and Kitchen; Octo (93M), π0.5, GR00T (3B), CogACT (7B), OpenVLA-OFT (7B). The paper’s broader census also includes RDT, SmolVLA, DreamerV3, BC/MDN, Diffusion Policy, Flow, BeT, and VQVAE.

Critical strengths

The relevant novelty is connecting numerical error with action-level behavioural failure and mechanism-specific repair. Prior mixed-precision practice primarily evaluates losses, overflow, or broad stability; this paper argues that a robot policy can be numerically healthy by those metrics while behaviour has already collapsed. The strongest causal evidence is the starvation analysis. Protecting the scalar loss or parameter gradients doesn't help, while preserving backward activation-gradient fidelity does; equal-power zero-mean noise is tolerated, whereas persistent near-zero bias can trigger collapse.

In particular:

  • Action Level definition and measurement of collapse: measuring divergence in executed actions against an fp32 twin directly addresses failures that normal loss, gradient norms, and overflow counters can miss. The paper also links action drift to degraded rollout return across 21 cells, though the relationship is weak near the fp32 noise floor.

  • Mantissa starvation as a mechanism for numerical posterior collapse. It localizes failure to the backward activation-gradient chain, shows a toward-zero bias is damaging, and rescues the posterior using precision-only interventions without changing the objective.

  • Conservative collapse via AdamW second-moment inflation. The authors observe that low-precision noise inflates AdamW’s second moment and leads to overly timid, behaviourally degraded policies. The CogACT 7B result, where LockedAdamW improves both seeds while UPH does not, is also relevant.

  • Three-mode taxonomy: overflow, starvation, conservative collapse: turning diagnosis into a seconds-to-minutes procedure is highly actionable. The reported rank correlation with full-run drift is 0.86 across seven cells, and the prospective ACT routing case succeeds.

  • The exponent-floor experiments, including healthy e6m2 behaviour and real Transformer Engine fp8 validation, provide a concrete numerical-design insight: range/floor placement can matter more than mantissa width in these cases.

Critical weaknesses
  • Limited task-level and real-robot validation. The central outcome is robot behaviour, yet the paper reports a clear task-success result mainly for MH-Lift. It explicitly states that real-robot transfer is unvalidated. Action divergence correlates with return across 21 cells, but that does not establish broad closed-loop success across tasks or hardware.

  • Incomplete repairs in the hardest ACT settings. Can and Kitchen are diagnosed as starvation cases, but neither bypass nor UPH reaches the stated fp32 floor at the tested budget. The paper’s honest conclusion is “no sufficient scheme,” which weakens the claim that CURE reliably closes diagnosed failures.

  • LockedAdamW is not robustly established as a general repair. It succeeds on CogACT and substantially extends Square training, but it can misfire on GR00T, stall on Lift at 51k steps, and produce NaNs after 150k steps on Square. Its global scalar control ignores per-layer variation, and its interaction with standard learning-rate schedules is untested.

-The evidence currently supports a strong diagnostic and mechanism paper more convincingly than a mature, general-purpose training solution.

Questions for authors

Can you provide evidence of the distinct practical advantage of LockedAdamW? Can you discuss Adafactor-style RMS clipping and broader precision-management baselines?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.