Pith. sign in

REVIEW 3 major objections 4 minor 12 references

Why the Third Axis Is Freedom

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Best-of-K training is really a lever on freedom, not on mode count.

desk verdict Solid, checkable theorems about best-of-K training and support size; the headline 29/30 empirical win is weaker than it looks because the freedom selector also reads the test-side contexts. read the letter →

arxiv 2608.05423 v1 pith:XYLEFJPB submitted 2026-08-05 cs.LG cs.AI

Michael Timothy Bennett This is my paper · ORCID
classification cs.LGcs.AI
keywords explorativemodelingfreedomoffunctionbest-of-Ktraininggenerativeexpressivitygeneralisationcoverageidentitymodelselectionpermissionprofiles
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Explorative Modeling (XM) draws $K$ candidate generations per training step and updates on the closest one; this paper argues that what such training actually optimises for is freedom, the weakness of the constraint implied by a model's behaviour, measured by how many further compatible commitments remain open. The paper proves that hard best-of-$K$ risk equals the expected $K$-th power of the probability that one candidate misses an acceptable region, which turns exploration into an amplification of whatever outputs the deployed policy already permits. In a finite permission language where freedom counts the nonempty subsets of a model's permitted output set, best-of-$K$ hit probability rises strictly with local freedom for every $K>1$ under uniform targets and balanced mass, and future compatibility under independent nonempty demands is exactly proportional to freedom. The paper concludes that generative expressivity is only a mode-count proxy for freedom, and presents experiments where larger $K$ increases measured freedom and a freedom-based model selector beats child-validation selection in 29 of 30 distribution-shifted worlds.

What carries the argument

The central objects are the permission profile $F(c)$, the set of outputs a generator can emit with positive probability at context $c$, and the finite embodied language in which a policy is the set of exclusions it enforces; freedom is the product over contexts of $2^{|F(c)|} - 1$, the number of nonempty subsets of permitted outputs. The central identity is the coverage identity $R_K(\pi) = \int_0^\infty \mathbb{E}_{c,Y}[(1 - q_{\pi}(c,Y,t))^K] \, dt$, which turns best-of-$K$ risk into the $K$-th power of a miss probability; under identification loss and a uniform target it becomes expected distinct-output coverage divided by the number of valid outputs. Theorems 5 and 6 show that sampled coverage rises strictly with local freedom for $K>1$ and that the marginal gain from another candidate grows with freedom, while Theorem 8 connects the same product of freedom factors to future compatibility.

What would settle it

On a finite output set with uniform targets and a predeclared admissibility rule, train two stochastic Forward XM models with $K=4$ and $K=32$ on identical data and audit their deployed permission profiles with fresh draws; if the larger-$K$ run ever shows strictly lower measured freedom while attaining no better best-of-$K$ loss than the $K=4$ run, the claim that exploration increases freedom falls. Equally, enumerating Theorem 5's formula for some $m$, $a$, $K \geq 2$ and finding $H^*_K(a)$ non-increasing in $a$ refutes the central monotonicity result.

Watch

Extended reading notes

Core claim

The paper's central claim is that the third pretraining axis found in Explorative Modeling is not generative expressivity, the greatest number of modes a training objective can retain, but freedom, defined on the trained policy as the volume of compatible completions it leaves open. The load-bearing formal result is Theorem 5: for uniform targets on $m$ valid outputs, a generator that places uniform mass on a permitted subset of size $a$ has best-of-$K$ hit probability $H^*_K(a) = (a/m)[1 - (1 - 1/a)^K]$, which is strictly increasing in $a$ for every $K \geq 2$, hence strictly increasing in the local freedom factor $2^a - 1$. Theorem 8 shows that under independent nonempty demands on unseen contexts, the probability that a child-correct profile remains compatible equals its total freedom divided by a constant, so freedom and future compatibility induce the same ordering. Together these results say that exploration rewards sampled access to a policy's permitted outputs, and what the policy permits determines how well it generalises to new requirements.

Load-bearing premise

The argument assumes that a model's semantics is exactly the finite set of outputs it can produce with positive probability, so freedom can be counted as the nonempty subsets of that set; if real outputs carry structure that distinguishes two equal-size supports, or dense generators assign positive mass everywhere, this counting cannot rank models and the theorems' ordering claims do not apply.

Editorial extensions

If this is right

  • At $K=1$ the objective cannot rank two sound policies by repertoire, since every distribution supported on the targets has the same risk; from $K=2$ onward the unique optimum is uniform mass over the full target support, so exploration is what makes freedom visible to training.
  • For fixed target-support size, the gain from one more candidate is $(1/m)(1 - 1/a)^K$, which rises strictly with permitted count $a$, so larger-$K$ gains should concentrate in models that can retain many correct outputs.
  • Under nonuniform targets, finite $K$ concentrates mass on frequent outputs, but as $K$ grows the optimum broadens toward uniform mass over full target support; large $K$ shifts training pressure from frequency-matching to coverage.
  • Deterministic $K$-head models are the only case with a support ceiling of $K$; stochastic Forward XM is governed by the coverage identity, so its deployed policy can have support larger than $K$.
  • Among child-correct policies, future compatibility probability is exactly proportional to total freedom, so selecting models for freedom is a direct route to robustness when output frequencies shift.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if $K$'s value is mediated by freedom, exploration scaling should interact with data size and architecture changes that alter a model's support; $K$ is not an independent capacity axis but a lever on another quantity.
  • Beyond the paper: Theorem 8's demand law draws each unseen requirement uniformly from nonempty output sets; if real shifts are structured rather than uniform, a weighted freedom measure may predict compatibility better than the raw count.
  • Beyond the paper: the transform $q \mapsto 1 - (1-q)^K$ applies to any best-of-$K$ training signal, so the freedom-based analysis could extend to minimum-of-$N$ objectives in other domains, not just generative modelling.
  • Beyond the paper: a mediation experiment that varies $K$ while intervening on measured freedom, or varies freedom while holding $K$ fixed, would separate the exploration mechanism from the generalisation benefit that the paper's selection result suggests.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper argues that the 'third pretraining axis' claimed by Explorative Modeling (XM) is not generative expressivity but freedom, understood as the weakness of the constraint implied by a trained policy. The first half develops a formal theory: a coverage identity for best-of-K training (Theorem 1), an essential-infimum limit (Theorem 2), exact finite risk under identification loss (Theorem 3), rank-theoretic results showing that under uniform targets and balanced mass the best-of-K hit probability increases strictly with the support size and hence with the finite-language freedom factor (Theorems 4-6), a nonuniform-target optimum that approaches uniform support as K grows (Theorem 7), and a theorem that future compatibility under independent uniform nonempty demands is proportional to freedom (Theorem 8). The second half reports two experiments: Experiment 1 shows that sample-based Forward XM training with larger K increases or saturates the measured freedom in a finite 12-cell output lattice; Experiment 2 compares a freedom-based selector with a child-validation selector over XM-trained candidate pools and reports a 29/30 win with a +0.072239 gain in balanced-parent hit. The paper includes code, verification scripts, and an unusually candid limitations section.

Significance. The formal core is a genuine strength. Theorems 1-8 are stated with explicit assumptions, the key algebra in Theorems 5 and 7 checks out, and the appendix's claim that scripts verify the finite identities and numerical claims is the kind of reproducibility that should be credited. If the empirical selection result could be made causally clean, the paper would provide a useful bridge between the XM literature and the author's Stack Theory framework, giving a function-level quantity that can be measured and selected for. The significance is tempered, however, by two scope restrictions: the operational definition of freedom relies on a finite permission language that is not defined for dense generators, and Experiment 2's transductive design does not isolate freedom from access to unlabelled parent contexts. The theoretical contribution is therefore more secure than the paper's title-level empirical claim.

major comments (3)
  1. [10.2, Table 3] The central empirical claim that freedom selection improves XM under distribution shift is not isolated by Experiment 2. The freedom selector read 256 unlabelled parent contexts, while the child-validation selector read none, and the endpoint was computed from the same 240 deployment draws used to form the selector's parent-context features. Selection noise and endpoint noise are therefore not independent, and a transductive selector that uses parent contexts without any freedom instrument could plausibly produce a similar gain. The paper discloses this in Section 13 and calls the result a transductive common-pool comparison, but the abstract and conclusion still assert that 'selecting for freedom improved XM under distribution shift.' Because this experiment is the only direct evidence for the selection claim, the manuscript needs either a held-out audit split with selector-probe draws separated from endpoint draws, or a transductive control that uses the same parent contexts but not the freedom instrument, before the claim can be accepted as stated.
  2. [7, Definition 10, Theorem 8] The result that future compatibility is 'exactly proportional to freedom' is a consequence of the specific demand distribution chosen in Definition 10: independent draws uniform over nonempty subsets of the output set. The paper itself notes in the interpretation of Theorem 8 that a pooled demand law can reverse the total-freedom ordering. Thus the generalisation significance of freedom is conditional on a modelling assumption that is plausible but not empirically justified. This is not a mathematical flaw, but the conclusion's phrasing 'freedom decides future compatibility' overstates the status of Theorem 8; it should be presented as a theorem about a chosen demand model, with the dependence on that model made prominent in the abstract and conclusion.
  3. [13, Definition 5] The finite permission language identifies freedom with the support cardinalities of a generator at each context. For dense softmax or continuous generators, the raw support profile is constant and cannot rank models, as the paper acknowledges. Since the XM results that motivate the paper are in image, video, and language domains where outputs are dense, the title-level claim that 'the third axis is actually freedom' is not operational for those domains without a predeclared admissibility rule or a measured completion proxy. The paper should either restrict the central claim to fixed finite output languages or provide at least one demonstration on a non-finite output space; currently the gap between the formal framework and the motivating application is substantial.
minor comments (4)
  1. [2, 5, 6, 10] Several informal asides and footnotes ('Yay', 'Not invaded', 'Well, it is now', the footnote about LLMs ruining the word 'exact', and the 'years of my life spent in relative obscurity' footnote) are unprofessional for a journal submission and should be removed or rewritten.
  2. [10.2, Table 3] The column label 'Best-of-8 hit' is ambiguous because the rows are labeled by training K values; clarify that this is an evaluation-time best-of-8 statistic, not the training exploration budget.
  3. [13] The limitations section is unusually candid and should be kept, but the abstract should be reconciled with it; phrases such as 'the third axis is actually freedom' should be qualified to reflect the conditional scope stated in Section 13.
  4. [7] The sentence 'The measurable Stack Theory result is broader' refers to prior work without enough detail for the reader to verify the exchangeable-demand claim; a self-contained statement or a more precise citation of Bennett 2026 would help.

Circularity Check

2 steps flagged · score 4.0 of 10

Best-of-K identities are self-contained, but Theorem 8 makes 'future compatibility ∝ freedom' true by construction, and the broader generalisation claim leans on the author's own prior work.

  1. self definitional [Section 7, Definition 10 and Theorem 8]
    "Definition 10 (Independent nonempty demands). For every unseen context c∈C_u, draw a future demand S_c ∼ Unif(2^Y \ {∅}) independently across contexts. A profile F is compatible with the demands when S_c ⊆ F(c) for every c∈C_u. ... Theorem 8 (Future compatibility is proportional to freedom). Let a_c := |F(c)|. Under independent nonempty demands, P(F is compatible) = ∏_{c∈C_u} (2^{a_c}−1)/(2^{|Y|}−1)."

    Under the uniform nonempty demand law, the number of demands compatible with F(c) is exactly 2^{a_c}−1, which Proposition 1 had already defined as the local freedom factor. Compatibility probability in Theorem 8 is therefore the same product as w(π_F) by construction; choose any other demand law and the proportionality fails. The theorem restates Definition 10 plus Proposition 1 rather than independently deriving that freedom controls future compatibility. This is the load-bearing step that gives freedom its 'generalisation significance'.

  2. self citation load bearing [Section 7, paragraph after Theorem 8]
    "The measurable Stack Theory result is broader. Let U be the unseen region of a task and let Bπ := Ext(π)∩U be the unseen buffer of a correct policy. Under a µ-exchangeable demand law, generalisation probability is a nondecreasing function of µ(Bπ), and strict under nondegeneracy [Bennett, 2026, 2025]."

    The paper's central premise that 'freedom' (formerly 'weakness') is the generalisation-relevant axis is supported by the author's own prior work (Bennett 2023, 2025, 2026), not by an independent source or by a fully self-contained derivation in this paper. The only in-paper derivation, Theorem 8, is definitional as noted above. Thus the interpretive claim that the third axis is freedom, and that freedom rather than mode count or exploration is the operative quantity, is load-bearing on self-citations. The best-of-K mathematics is independent, so this step is partial rather than total.

full rationale

The core derivation chain for the best-of-K identities (Theorems 1–7) is self-contained: Theorem 1 follows from the tail-integral identity, Theorem 3 from independence, Theorem 4 from Jensen, Theorem 5 from a derivative calculation, and Theorem 7 from KKT. No fitted parameters are involved, and the accompanying script regenerates the figures. I therefore do not treat the core XM/freedom relation as circular. The score is raised by two constructional steps. First, Theorem 8's demand model (Definition 10) is chosen so that the number of demands compatible with a profile F is exactly ∏(2^{a_c}−1), which Proposition 1 already defined as freedom; the 'future compatibility ∝ freedom' theorem is therefore a restatement of the definitions, not an independent derivation of freedom's generalisation significance. Second, the broader claim that weakest policies are likeliest to generalise is imported from Bennett's own prior work (2023, 2025, 2026), as the paper acknowledges in Section 7. Section 13 also candidly states the Experiment 2 confound: 'Its freedom selector read unlabelled parent contexts, while the child-validation selector did not. The same 240 deployment draws were used to form the selector's parent-context features and the reported parent-hit estimate, so selection and endpoint noise were not independent.' This is a disclosed external-validity limitation rather than a circular derivation, and I weigh it as such. Net: the formal identities are independent, but the interpretive axis 'freedom' and its generalisation significance are partly definitional and self-citational.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central theory depends on several assumptions the reader does not pay for upstream: deployment consistency, a finite exclusion vocabulary, and a specific future-demand law. The mathematical identities themselves are self-contained, but the generalisation claim and the measured freedom values inherit these assumptions. Empirical hyperparameters (admissibility radius, formula-bank size, gate) are fitted or hand-chosen for the experiments.

free parameters (4)
  • operational admissibility radius = 0.64
    A generated continuous vector is counted as occupying output cell j when its distance to the nearest one-hot prototype is within 0.64 (Experiments 1 and 2). This hand-chosen rule defines the permission profile on which freedom is measured; a different radius changes the freedom values.
  • freedom instrument residual coefficients = not reported in text
    Experiment 2's freedom selector combines a fixed primary freedom coefficient of 1 with nonnegative bounded residual terms fitted on 24 calibration worlds. The full formula bank (240 formulas) is shipped in code rather than described in the paper.
  • formula-bank size = 240
    Selected from {48, 120, 240} during calibration; the 240-formula bank was frozen before the 30 final worlds were evaluated.
  • child-fit gate = 0.97
    Selected from {0.97, 0.98, 0.99} during calibration; candidates whose child hit lies below 0.97 of the best child hit are excluded from freedom selection.
assumptions (4)
  • domain assumption Training candidate law equals deployed law (deployment consistency).
    Stated after Definition 6 and used throughout Sections 4 to 8 to turn best-of-K risk into a property of the deployed policy. Noise-searching hybrid XM variants violate this, as the paper notes in Section 13.
  • ad hoc to paper The finite exclusion vocabulary is the correct embodied language; output cells are symmetric and independent permissions.
    Definition 5 and Proposition 1 define freedom as the product of (2^{a_c} - 1) over contexts. This treats every output cell as an independent permission and makes support the only relevant property. Section 13 says richer languages can separate equal-cardinality supports.
  • ad hoc to paper Future demands at unseen contexts are independent and uniform over nonempty output subsets.
    Definition 10 and Theorem 8. This demand law is what makes future compatibility exactly proportional to freedom; a pooled demand law could reverse the ordering, as the paper itself explains.
  • domain assumption Candidates are conditionally i.i.d. given context and target.
    Definition 6 and Theorem 1 use independence to obtain the power K; Remark 1 gives the coupled-candidate version.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why the Third Axis Is Freedom." pith.science (2026). https://pith.science/paper/XYLEFJPB

@misc{pith2026260805423,
  author       = {Pith},
  title        = {Pith review of: Why the Third Axis Is Freedom},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYLEFJPB}},
  note         = {Machine review of arXiv:2608.05423}
}
abstract

In generative training, a model produces an output and is penalised for its difference from an example. With one output per comparison, a model that produces one common answer can outperform a model retaining a broader repertoire. Explorative Modeling (XM) produces $K$ outputs per comparison and updates on the closest, claiming exploration as a "third pretraining axis" associated with generative expressivity. Here I show the third axis is actually freedom, meaning the weakness of the constraint implied by a model's behaviour. Previous work showed freedom is a property of function rather than form. Parameters, architecture, minimum-description-length (MDL), and data can vary while the behavioural constraint remains unchanged. It was formally proved that weakest models are likeliest to generalise, and freedom selection beat MDL by 110-500\% in induction experiments. I prove average XM loss depends on the chance a candidate misses an acceptable region, with exploration raising miss probability to power $K$. For $K>1$, match probability rises with freedom. I then demonstrate empirically that XM optimises for freedom. In a Forward XM experiment, larger $K$ increased or saturated measured freedom, and increased freedom at every tested value under context-dependent targets. I trained XM candidate pools and compared validation selection with a freedom selector that read unlabelled parent contexts. Freedom won in 29 of 30 cases. Generative expressivity is a mode-count proxy for freedom, that discards the extension structure that gives freedom its generalisation significance. XM is a means, freedom an end, and selecting for freedom improved XM under distribution shift.

Figures

Figures reproduced from arXiv: 2608.05423 by the authors.

Figure 1
Figure 1. Best-of-K hit probability for a uniform target over m = 20 valid output cells when the generator is uniform on a permitted cells. For K = 1 the curve is flat. For every K > 1 sampled coverage rises with the local freedom factor 2a − 1. Example 1 (Freedom alone does not rank finite-K access). Let m = 3 and K = 2. A generator with q (1) = (1/2, 1/2, 0) has support size two and hit probability H2(q (1)) = 1 2 . A gener… view at source ↗
Figure 2
Figure 2. The exact optimum from Theorem 7 for target probabilities (0 [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Experiment 1. Mean log freedom after sample-based Forward XM training. Points show [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Experiment 2. Paired balanced-parent improvement from prospectively calibrated freedom [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 5 canonical work pages

  1. [2]

    Michael Timothy Bennett

    URLhttps://doi.org/10.1007/978-3-031-65572-2_2. Michael Timothy Bennett. Are biological systems more intelligent than artificial intelligence?, 2024b. URL https://arxiv.org/abs/2405.02325. In press, Philosophical Transactions of the Royal Society B: Biological Sciences. Special issue on Hybrid agencies: crossing borders between biological and artificial w...

    arXiv

  2. [12]

    Ray Solomonoff

    URL https://arxiv.org/ abs/2601.12837. Ray Solomonoff. Complexity-based induction systems: Comparisons and convergence theorems.IEEE Transactions on Information Theory, 24(4):422–432,

    arXiv

  3. [1978]

    URLhttps://doi.org/10.1016/0005-1098(78)90005-5

    doi: 10.1016/0005-1098(78)90005-5. URLhttps://doi.org/10.1016/0005-1098(78)90005-5. Ricard Sol´ e, Luis F. Seoane, Jordi Pla-Mauri, Michael Timothy Bennett, Michael E. Hochberg, and Michael Levin. Cognition spaces: natural, artificial, and hybrid,

    doi:10.1016/0005-1098(78)90005-5

  4. [1982]

    URL https://doi.org/10.1109/TIT.1982

    doi: 10.1109/TIT.1982.1056489. URL https://doi.org/10.1109/TIT.1982. 1056489. Jorma Rissanen. Modeling by shortest data description.Automatica, 14(5):465–471,

    work page arXiv 1982

  5. [1997]

    URLhttps://doi.org/10.1162/neco.1997.9.1.1

    doi: 10.1162/neco.1997.9.1.1. URLhttps://doi.org/10.1162/neco.1997.9.1.1. Marcus Hutter, David Quarel, and Elliot Catt.An Introduction to Universal Artificial Intelligence. Chapman and Hall/CRC, 1st edition,

    doi:10.1162/neco.1997.9.1.1 1997

  6. [2012]

    Sepp Hochreiter and J¨ urgen Schmidhuber

    URL https://proceedings.neurips.cc/paper_ files/paper/2012/hash/cfbce4c1d7c425baf21d6b6f2babe6be-Abstract.html. Sepp Hochreiter and J¨ urgen Schmidhuber. Flat minima.Neural Computation, 9(1):1–42,

    work page 2012

  7. [2016]

    Ke Li and Jitendra Malik

    URL https://proceedings.neurips.cc/paper_files/paper/2016/hash/ 20d135f0f28185b84a4cf7aa51f29500-Abstract.html. Ke Li and Jitendra Malik. Implicit maximum likelihood estimation,

    work page 2016

  8. [2017]

    URL https://doi.org/ 10.1109/CVPR.2017.264

    doi: 10.1109/CVPR.2017.264. URL https://doi.org/ 10.1109/CVPR.2017.264. Alexi Gladstone, Heng Ji, and Yilun Du. Explorative modeling: Unlocking a third pretraining axis and end-to-end generation,

    doi:10.1109/cvpr.2017.264 2017

Show all 12 references
  1. [2018]

    Stuart P

    URL https://arxiv.org/ abs/1809.09087. Stuart P. Lloyd. Least squares quantization in PCM.IEEE Transactions on Information Theory, 28(2): 129–137, March

    arXiv

  2. [2023]

    doi: 10.1007/978-3-031-33469-6

    doi

  3. [2024]

    URL https://doi.org/ 10.1201/9781003460299

    doi: 10.1201/9781003460299. URL https://doi.org/ 10.1201/9781003460299. Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Cran- dall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. InAdvances in Ne...

    doi

  4. [2026]

    6Codenames for experiments 1 and 2, now meaningless letter salad

    URLhttps://arxiv.org/abs/2607.27372. 6Codenames for experiments 1 and 2, now meaningless letter salad. 25 Abner Guzm´ an-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. InAdvances in Neural Information Process...

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.