acceptodds
Under review as a conference paper at ICLR 2027

The Exact Exchange Rate of Gradient-Free Credit

Abstract

Gradient-free training is back at LLM scale, and its cost is now the central question. Folklore holds that on a layer of parameters, scalar-broadcast evolution strategies need times more samples than backpropagation — directionally right, but the constant has never been measured. We derive it exactly: for antithetic rank-1 perturbations of an layer, the estimate from forward evaluations against the true gradient obeys with , parameter-free, exact in the linear regime at frozen weights; dynamic training adds landscape-dependent excess variance atop this floor. We verify it on every layer of the reference MLP (ratios 0.93–1.12) and of a frozen extreme-aspect-ratio chain (0.976–1.006), killing all four alternatives in that pre-registered experiment (maximum deviation 0.006 across a span of ). The law then prices, by extrapolation, the largest gradient-free training system to date: it operates at per-step and finances that signal with coherently integrated steps at the GPU-hours of backpropagation. Channel by channel, our measurements yield a design-space map (two primitives, four laws, six measured rows) and three impossibility statements, among them: no scalar-broadcast channel escapes the exchange rate, and approximate backpropagation is the wrong objective. This is a measurement study, not a new method.

Primary area:
optimization
Keywords:
gradient-free training, evolution strategies, credit assignment, sample complexity, zeroth-order optimization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CXSaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SUMMARY The article offers analysis of weight perturbation based gradient free training. In particular, it derives an expression for the cosine (square) angle between the perturbation based gradient and the backpropagation gradient, as a function of the dimensions oof hte weight matrix and the number of forward evaluation iterations per update. Then this expression used to evaluate different experimental scenarios including the large scale experiment of EGGROLL article. The authors also make some information theoretic claims regarding the updates, which are not rigorously based.

Critical strengths

Theorem 1 providing the cosine (square) alignment result, providing a clear expression in terms of weight matrix dimensions is valuable. The article also contains some numerical analysis based on this.

Critical weaknesses

Theorem 1 provides a useful result, however, it is not as major result to base a full article. The theorem itself and proof steps are relatively elementary. The rest of the numerical analysis and claims are not in a level to be offered as fundamental findings for a flagship ML conference article. In addition, the presentation is poor, obscured by financial analogies, convoluted narratives, undefined symbols, and imprecise assumptions.

COMMENTS ABOUT SECTIONS

Abstract

  • the sentence about the cosine(square) angle with backrop G is ambigious.
  • killing all four alternatives: ambigious and not precise statement.
  • that pre-registered experiment: not clear
  • not familiar with the use of prices and finances terms within this context.

Introduction

  • parallelizes trivially: weight perturbation, if handled in the same worker (by the claim: fits hardware smaller than backprop case), synchronous perturbations in multiple weights for paralllel updates cause performance issues? and if handled in multiple workers requires communication of update terms among workers and requires more hardware than backprop then. I am not sure if trivial is appropriate here.

  • The use of exchange-rate terminology (in title) and at the end of paragraph 1 is not precise?

  • the third paragraph: cos^2 is not clear, the parameters m and n are not defined yet. and what floor? the last sentence is also not clear (aspect-ratio chain...). what is folklore here, and what are natural alternatives (citations?)

  • Paragraph 4 is also not clear.

Section 2

This section could be clarified and potentially be placed in the Appendix.

Section 3

  • Before Theorem 1, subindex sσs_\sigma was used for perturbation scale, right after that it is used as forward evaluation index sis_i.

  • The expression for g^N\hat{g}_N assumes an even NN value. In, Theorem 1, g^1\hat{g}_1 with that definition does not make sense. subindex of \hat{g} is not really the number of forward evaluations but antithetic index. whitened-> white. also zero mean i.i.d. unit variance.

  • Proof of Theorem 1: The expression Eb[(c⊤b)22∣∣b∣∣2]=(n+2)∥c∥2E_b[(c^⊤b)^22||b||^2] =(n+2)\|c\|2 is valid for kurtosis of bb is equal to 33, i.e. Gaussian. However, this assumption is not stated.

Overall, Theorem 1 appears very basic, but useful. Its proof can be expanded for better clarity.

  • Table 1: A column showing normalized error in predicting the cos2cos^2 would be informative.

  • Section 3.4 what are the source of these hypotheses (references?) why do you need to compare against them if H0 is theoretically grounded?

Section 4 In my opinion this whole section can be reduced to one or two paragraphs, which is looking at the large scale experiments of EGGROLL (Sarkar et al., 2026) from the lens of the Theorem 1 of the current article.

Section 5 I could not follow the discussion in this section, which put forwards some information theoretic statements such as mutual information, capacity without any rigorous treatment.

Section 6 What are these labels P-7.6x refering to? Regarding the discussion in this section: it is clear that gradient does not necessarily offer the best search direction, but it is informative to see alignment vs accuracy values together.

Section 7 I would expect better discussion here discussing the utility, contributions, conclusions and limitations of the article. Instead it contains specific information which is confusing and containing labels like R-7.55a which is not predefined.

Questions for authors
  • Regarding Theorem 1: Do you assume that a,b are Gaussian and and zero mean?

  • what are some symbols used in the article such as P7-6, R-7..

  • what are the hypotheses in Section 3, are they from any specific references?

  • Could you provide more rigorous statements about the information theoretic claims made in Section 5?

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

The paper presents itself as a measurement study of the cost of gradient-free (scalar-broadcast, ES-type) training relative to backpropagation. Its central question is: what is the "exchange rate" between forward evaluations and gradient information in a scalar-broadcast channel? It makes three contributions:

  1. For antithetic rank-1 perturbations of an m×n layer, the per-step alignment with the true gradient obeys cos²(ĝ_N, G) = N/(N+N₀) with N₀ = 2(m+2)(n+2) ≈ 2d. The law is parameter-free and exact in the linear regime at frozen weights (Theorem 1). It is verified on two small architectures (MLP-4, CNN-4) and on a pre-registered extreme-aspect-ratio chain that separates it from four alternative constants, including the folklore 2mn.

  2. Read through the law, the largest published scalar-broadcast system (EGGROLL) is inferred to operate at per-step cos ≈ 0.02. It compensates through ~10⁵ Adam-integrated steps, at ~180× the GPU-hours of backpropagation.

  3. A taxonomy of "credit channels": two primitives, four laws, six rows and three impossibility statements. It is complemented by an experiment showing that alignment with backpropagation is neither necessary nor predictive of accuracy.

Critical strengths

S1 – A precise, parameter-free constant, tested by a well-designed decisive experiment. Theorem 1 is derived cleanly (Appendix A, including the kurtosis generalization). The extreme-aspect-ratio chain (§3.4) is a smart design: it separates the exact constant from the 2mn folklore where they actually differ, with a maximum prediction deviation of 0.006 over a 340× span of N₀. This turns a ∼d scaling statement into a usable number, which is the paper's main value.

S2 – Careful scoping and a common instrument. The explicit "what we do not claim" list avoids confusion from the outset. Failed pre-registered verdicts are reported verbatim. A single cosine-to-BP probe is used to compare several backpropagation-free methods against backpropagation (§6). The resulting design map is a useful conceptual tool for situating future gradient-free methods.

Other strengths:

Sound formalism for the measurement instrument (single autograd entry point, explicit conventions), and a detailed reproduction protocol (Appendices G–H); I could not inspect the code itself as it is not available as supplementary materials despite the authors stating so.

Critical weaknesses

W1 – The paper reads as two works, and the second is weakly supported. §2–4 give a focused and mostly sound result on the noise floor of rank-1 antithetic ES, although it is obscure at times. §5–6 build a broader taxonomy of "credit channels" largely on the authors' earlier experiments (the T-channel, GFN and v5–v8 lineage of Appendices D–F), and the main text does not introduce those methods. Only the bandwidth and coherence laws connect to Theorem 1. The "two primitives generate four laws" framing and the impossibility statements are stronger than the heterogeneous evidence supports: one theorem, one information argument, and two measured regularities, one of which rests on a single-seed sweep while the other is "undecided" on CIFAR-100. The methods, motivation and key results behind §5 should be introduced explicitly in the main text.

W2 – Several claims oversell the evidence.

  • "Exact." On MLP-4, the linear-regime floor (ratios 0.93–1.12) comes from a single cell (r=1, N=512, σ=0.01; Tables 6–7). The ratios are reported without uncertainty, and the per-sample SD of cos at L1 alone (0.0085 ± 0.0010) corresponds to roughly ±25% on N₀eff.
  • Null control (§3.4). I appreciate the honest reporting of the failed null control, but the post-hoc explanation (a "miscalibrated gate") is partly circular. It leaves the sanity check unable to detect estimator contamination in this regime, and the revised registration that turns the FALSE verdict into a "decisive confirmation" was written after the result was known.
  • EGGROLL correspondence (Table 3). The table only shows that EGGROLL's choices (huge population, Adam integration, rank for conditioning) are coherent with the paper's picture of the channel. Most rows are qualitative and rest on an inferred cos ≈ 0.02 that is never measured at that scale. No quantitative prediction, such as the 180× cost, is derived from the law and then tested. Row 4 relies on rank flatness, which Appendix A.5 itself calls "expected, not diagnostic" on near-square layers. The ✓ marks and statements such as "the rate is paid, not evaded" or "extrapolation is arithmetic, not hope" therefore oversell what the evidence supports.

Other weaknesses:

  • Some terms are used but never defined. "Scalar-broadcast family" and "antithetic rank-1 estimator" appear from the abstract onwards and are assumed to be understood without definition.
  • The authors use financial metaphors "Exchange rate", "financing", "price list" and "spending the law" that sometimes feel imprecise. The quantity is a sample-complexity constant. Defining each metaphor once in standard terms (N₀ as the evaluation count for cos² = ½; "financing" as multi-step averaging) would ease comparison with prior work such as Nesterov & Spokoiny (2017) and Malladi et al. (2023).
  • Labels such as A-02b, A-06, G1–G3, T1′, P-7.6a and "sealed figure" appear in the main text without explanation. The concept of gates is only defined in Appendix C and needs one sentence in §3.4. Some wording is informal (e.g. "deliberately boring").
  • Tables 3 and 4 are hard to read and look draft-like: dense cells, no horizontal separators.
  • The experimental framework remains narrow. Two small architectures on MNIST/CIFAR, d ≤ 3.15×10⁶, three seeds for the scale experiments and a single consumer GPU, extrapolated to d ~ 10⁹–10¹⁰.
Questions for authors

Q1 - Appendix A.3 shows that ⟨ĝ_N, G⟩ concentrates at rate N̂^(−1/2), and asserts the same order for ‖ĝ_N‖². From which number of pairs N̂, relative to the dimensions m, n, does the limit describe a single draw rather than an average? How should the law be read at small N (e.g. N = 64)?

Q2 - Why do the σ = 0.01 cells, where the floor is claimed to be exact, not appear in Table 1, and why is σ = 0.01 measured at a single (r, N) cell on MLP-4 and not at all on CNN-4? Can you report N₀eff/N₀ with uncertainty, and the σ = 0.01 N-constancy on the trained networks?

Q3 - The decisive aspect-ratio experiment uses a frozen, randomly initialized chain (σ ∈ {0.01, 0.05}). Does the H0-vs-H4 separation survive on a trained network in the linear regime?

Q4 - Can the expected null cos(G, G′) for the chain be derived a priori and compared with the measured 0.1287, so that the null control becomes informative again?

Q5 - Can the law yield a quantitative prediction for EGGROLL (e.g. the GPU-hour ratio or the number of steps) that is compared with the reported value, rather than a consistency statement?

Reviewer_mAd8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Because this paper reads like LLM output, I'm going to make the assumption that it is more or less entirely written by AI and review with the kind of disdain that a person who expects others to volunteer their valuable time reading, when in fact they have themselves spent no actual time producing it - merits. The "authors" purport to have originated all scientific ideas by themselves, but when it is clear they have not even taken the time necessary to ensure that their conference paper looks like a conference paper , and not some hybrid chatGPT generated press-release, and a summary of some claude session, I think I am within my rights as a reader to take the human originality of this work with a grain of salt. Right down to the abstact, there are these awful financial metaphors for what they claim to have accomplished that clearly do not come from a human being.

Critical strengths

I don't see any.

Critical weaknesses

I feel the fact that volunteers are expected to take this work seriously despite it being obvious and unverifiable LLM output, right down to the title and abstract, speaks to a weakness of the review system in general. One should not be expected to spend ones time carefully evaluating the scientific merit of something that required no real time from an actual "scientist." Whether or not we take the claim at face value that only the writing was produced by LLM, and the core ideas and proofs are the result of some human effort, it seems to me that part of the work of scientific production is scientific communication. If the authors can't be bothered to do the minimal work to ensure their conference paper sounds like a conference paper, why should reviewers be expected to engage with the material as though it were a conference paper? In either case the writing here clearly places this work below the criterion for acceptance in a serious scientific conference such as this one. And writing aside, this contribution is not in my opinion sufficient for ICLR.

Questions for authors

Just wondering if you'll confirm my suspicion that this paper was end-to-end LLM generated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.