{"id":"3f8b9cfc-b774-4856-885f-94d5742d1cfa","arxiv_id":"2505.10559","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Under a river-valley model of the loss landscape, the paper shows that valley fluctuations behave like heat, with learning rate as temperature, and derives a 1/t optimal decay schedule.","lead":"This paper treats large language model training as a thermodynamic process, mapping the learning rate to temperature and loss valleys to heat. If correct, it explains why warmup-stable-decay schedules work and gives a simple 1/t decay rule.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GPT-2 slope confounded by slow-loss D-dependence: the paper's own Eq. (6) and Fig. 4 suggest ℓ_s varies with ηmin via the decay-area term, so the fitted 110 may not be thermal.","rationale":"I agree with the reader's CONDITIONAL verdict. The toy-model derivations are internally consistent, and the toy annealing phase diagram supports the optimal schedule in the simplified setting. The main unresolved threat is the identifiability of the thermal slope in the GPT-2 experiment. The paper's own Eq. (6) and Fig. 4 provide a quantitative mechanism by which the fitted slope could be spurious: the learning-rate sum D varies with ηmin through the decay-phase area, and the measured loss sensitivity to D is large enough to account for a substantial fraction of the observed 110ηmin trend. Because the paper's central empirical confirmation rests on this slope, the condition for accepting the LLM-level claim should be that the thermal interpretation survives a matched-D sweep. I do not see a basis for REJECT: the theoretical core is sound, and the concern is an identifiability condition that can be tested directly.","tokens_in":17523,"tokens_out":9806,"duration_ms":97574,"concrete_test":"Re-run the Fig. 2 GPT-2 sweep with D matched across ηmin, e.g., adjust the stable-phase length so that every schedule has the same total learning-rate sum (for ηmin = 0, add 500 stable steps at η = 6e-4; for intermediate ηmin, interpolate), and report validation loss vs ηmin with at least 3 seeds and error bars. If the slope drops from ≈110 toward ≈50 (the D-only prediction), the thermal attribution fails; if it remains ≈110, the ℓs-constancy assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative confirmation of NTL rests on identifying the GPT-2 slope dℓ/dηmin = 110 (Sec. 3.3) with the thermal coefficient NC in Lf = NCη. The paper's defense is that 'we can assume that ℓs does not vary much across decay schedules.' That assumption is not tested and is suspect on the paper's own evidence. In Eq. (6), the final loss is ℓfinal = ℓ(D, ηmin) + corrections, with D the total learning-rate sum, and Sec. 5 shows ℓ is sensitive to D (dℓ/dD ≈ 0.1 in Fig. 4). The WSD sweep in Fig. 2 does not hold D fixed: with 2k warmup, 7k stable at 6e-4, and 1k cosine decay, D ≈ 5.1 + 500·ηmin, so ΔD ≈ 0.3 across the sweep. This predicts Δℓs ≈ 0.1 × 0.3 = 0.03, roughly half the observed Δℓ ≈ 0.066 (= 110 × 6e-4). The slope 110 is therefore consistent with slow-loss variation through D alone; the inference N ≈ 5×10^6 is not identified unless D is controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Neural Thermodynamic Laws (NTL), a framework that maps stochastic optimization of large language models in a 'river-valley' loss landscape to thermodynamics. It analyzes a toy model with a fast quadratic valley direction x and a slow river direction y. With a fixed learning rate, the valley dynamics reach a Gaussian steady state whose average fast loss is linear in the learning rate and independent of curvature (equipartition), motivating an effective temperature T~η and a heat capacity C. During decay, the authors derive an optimal 1/t learning-rate schedule for the fast component. In the slow dynamics, the averaged fast loss produces an entropic force proportional to -∇ log a(y). The paper validates the optimal schedule in toy experiments and reports that GPT-2 pretraining validation loss is linear in the final learning rate η_min with slope 110, interpreted as N·C with N≈5×10^6 valley directions. It also draws analogies to the three laws of thermodynamics, Fourier's law, and entropic trapping, and gives practical guidance for WSD schedules.","tokens_in":17828,"tokens_out":16270,"duration_ms":150561,"significance":"The toy-model derivations in Sections 3–5 are internally consistent and exactly solvable, the toy experiments verify the predicted optimal decay schedule, and the framework yields falsifiable predictions (the linear ℓ_f–η relation, the 1/t schedule, and the dependence on the learning-rate sum D). These are genuine strengths. If the quantitative LLM result were validated, the paper would offer a significant mechanistic connection between optimizers and thermodynamics. However, the LLM evidence is currently too weak to support the claimed quantitative confirmation: a single small model, no error bars, a slope used to fit N rather than to test a pre-specified constant, an untested assumption about ℓ_s, and an unbridged application of SignGD theory to Adam. The significance is therefore conditional on the experimental identification being fixed.","major_comments":[{"comment":"The identification of the fitted slope dℓ/dη_min = 110 with the thermal coefficient NC is confounded by slow-loss variation with the learning-rate sum D. In the WSD sweep, D = 0.6 (warmup) + 4.2 (stable) + 0.3 (decay) + 500 η_min = 5.1 + 500 η_min, so ΔD = 0.3 across η_min ∈ [0, 6×10^-4]. The paper's own Fig. 4 shows validation loss depending on D; using a slope of roughly 0.1 as read from that figure, ΔD = 0.3 predicts Δℓ_s ≈ 0.03, about half of the observed Δℓ ≈ 0.066. The statement in Sec. 3.3 that 'ℓ_s does not vary much across decay schedules' is therefore unsupported and conflicts with Eq. (6), which lists ℓ(D, η_min) as the dominant term. The slope 110 and the inferred N ≈ 5×10^6 are not identified unless D is held fixed or the ℓ_s(D) contribution is subtracted.","section":"Sec. 3.3; Eq. (6); Figs. 2(b), 4"},{"comment":"The quantitative comparison uses the SignGD steady-state coefficient L_f = N√(π/32)σ_g η, but the GPT-2 runs use Adam with finite β1 and β2 plus adaptive scaling, not SignGD. The footnote that SignGD is a special case of Adam when (β1, β2) = (0, 0) does not supply a bridging argument: standard Adam's update and noise properties differ. Without a stated and tested assumption that Adam has the same linear thermal-loss law and coefficient, the reported N is an interpretation of a fitted slope rather than a quantitative prediction. The estimate σ_g ≈ 7×10^-5 is also not documented; N is inversely proportional to σ_g, so this free parameter drives the headline number.","section":"Sec. 3.3; Table 1; footnote 3"},{"comment":"The paper states that Δ_anneal ≈ 0 when the decay phase is at least 3k steps (Sec. 6), but the GPT-2 experiment in Fig. 2 uses a 1k-step cosine decay. By the paper's own criterion, the final losses in Fig. 2(b) may include a non-negligible annealing correction, so the identification ℓ ≈ ℓ_f(η_min) used to interpret the slope 110 is not supported. The experiment should either use a decay phase long enough for quasi-static equilibrium or explicitly model the annealing correction.","section":"Sec. 6 vs. Sec. 3.3"},{"comment":"The LLM evidence rests on a single GPT-2-small run per η_min value with no error bars or repeated seeds. The quantitative claim is a linear fit over a small number of points, with the linear regime selected post hoc; run-to-run variance could change the fitted slope substantially. Reporting means and standard errors over several seeds is necessary to support the claimed coefficient 110 and the resulting N.","section":"Sec. 3.3"}],"minor_comments":[{"comment":"The text uses 'SignSGD' in the first paragraph but 'SignGD' throughout the rest of the paper; the terminology should be unified.","section":"Sec. 3.1"},{"comment":"In the definition of entropy, the slow-variable dependence is written with x (e.g., 'S(x) ∝ −1/2 log a(x)') even though a is a function of the slow variable y; this is confusing and should be corrected to a(y).","section":"Sec. 5.2"},{"comment":"The half-time t_h for SignGD is typeset with an ambiguous square-root expression; it should be written unambiguously (e.g., √(2π) σ_g/(aη)) to match the Appendix C derivation.","section":"Table 1 and Sec. 4.1"},{"comment":"The experiment description omits the number of data points in the linear fit, the validation split details, and the Adam hyperparameters (β1, β2, ε, weight decay); these details are needed for reproducibility.","section":"Sec. 3.3"},{"comment":"The claim that the three learning-rate curves 'align reasonably well' is judged visually; a quantitative alignment metric or residual reporting would strengthen the conclusion about entropic forces.","section":"Sec. 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is at the boundary of ML theory and physics-inspired heuristics. The toy model is solid, but the empirical validation is too thin for the strength of the claims. The central quantitative confirmation depends on resolving the D-confound, the Adam bridge, and the annealing-length inconsistency. I would encourage the authors to strengthen the GPT-2 experiments (control D, report seeds, use a sufficiently long decay, and address the Adam gap) before publication. The 'laws' terminology in the title and abstract is stronger than what the text demonstrates; softening or caveating it would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the toy model is the real contribution. The derivations for SGD/SignGD on a river-valley landscape are clean, the steady-state Gaussian width and the optimal 1/t decay with explicit half-time th are correct as far as I checked, and the toy experiments confirm the phase diagram (b=0.5, th=2/aη). The thermodynamic packaging — temperature=η, heat capacity, entropic force — is expositionally nice and the authors are appropriately cautious in the limitations section.\n\nThe GPT-2 evidence is the soft spot. One 124M model, 10k steps, no error bars or seeds; the linear loss-ηmin relation is fit and then used to infer N≈5×10^6 valley directions, so the 'prediction' is not pre-specified. The theory is for SGD/SignGD, but the experiments use Adam, and the paper's only bridge is that SignGD is Adam with β1=β2=0, which standard Adam is not. There is also a citation gap: the Gaussian steady state of SGD and the 1/t annealing rate are classical results in stochastic approximation (Robbins-Monro and successors), and the paper presents them as new without engaging that literature.\n\nAbout the stress-test note: the specific D-confound arithmetic doesn't land. In Fig. 4, loss decreases as the learning-rate sum D increases (the curves are downward-sloping). Since the WSD sweep has D = 5.1 + 500 ηmin, increasing ηmin raises D and should lower ℓs, not raise it. That means the observed slope of 110 is a lower bound on the thermal coefficient, not a slope contaminated by ℓs variation in the direction claimed. So the quantitative concern should be retracted. The broader design point — the sweep does not hold D fixed, so the slope is a total derivative rather than ∂ℓ/∂ηmin — is fair and should be addressed with a controlled experiment.\n\nBottom line: worth a serious referee. The toy model is formally sound and may have practical value for LR schedule design. The LLM validation needs to be stronger before the thermodynamic claims for real LLMs can be accepted. I would send it to review with a request for reproducible experiments (seeds, error bars, fixed D) and better citations.","headline":"Clean toy-model thermodynamics with a well-tested 1/t decay rule; the GPT-2 confirmation is suggestive but not load-bearing, and the D-confound note reverses the sign of Fig. 4's slope.","tokens_in":18366,"tokens_out":7836,"would_cite":true,"duration_ms":72252,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM training is a thermodynamic process: the learning rate acts as temperature, and the thermal part of the final loss grows linearly with it.","keywords":["neural thermodynamic laws","river-valley loss landscape","equipartition theorem","learning rate schedules","large language model training","thermal loss","entropic force","1/t decay schedule"],"falsifier":"Scan the final learning rate $\\eta_{\\min}$ while varying the decay-phase length (for example, 1k, 3k, and 10k steps): the thermodynamic claim predicts that the fitted slope in $\\ell = 3.145 + 110\\eta_{\\min}$ stays the same whenever the decay is long enough for quasi-static annealing, whereas a slope that changes with decay length would show that the constant-$\\ell_s$ assumption fails and the equipartition reading is unsupported.","tokens_in":17297,"feed_emoji":"🌡️","tokens_out":14125,"duration_ms":114418,"temperature":0.7,"pith_summary":"The paper sets out to establish that training a large language model is a thermodynamic process in a quantitative sense. Under the river-valley structure of the loss landscape, fast valley directions reach a Gaussian steady state whose average thermal loss is linear in the learning rate and independent of curvature, an equipartition result. On a GPT-style transformer the paper measures the final validation loss as roughly $3.145 + 110\\eta_{\\min}$ and interprets the slope as $NC$, giving about 5 million valley directions, only 4% of the parameters. If this interpretation is correct, learning-rate scheduling becomes an annealing problem, and the paper derives an optimal decay schedule $\\eta_t = (\\eta/2)/(1 + t/t_h)$.\n\n","feed_headline":"LLM loss obeys thermodynamics: thermal loss grows with learning rate","feed_subtitle":"A measured linear fit maps validation loss to thermal loss and yields an optimal 1/t learning-rate decay.","key_machinery":"The river-valley toy model $\\ell(x,y) = c(y) + \\frac12 a(y)x^2$, with fast variable $x$ and slow variable $y$, is the load-bearing object. For fixed $y$ the fast coordinate equilibrates to a Gaussian of width $\\sigma = d(\\eta,\\sigma_g)/\\sqrt{a}$, and the key identity is that the thermal loss $\\ell_f = \\frac12 a\\sigma^2 = C\\eta$ is independent of sharpness $a$. This calculation carries the thermodynamic mapping: $\\eta$ is temperature, $C$ is heat capacity, the exponential relaxation of $\\ell_f$ toward $C\\eta_B$ is thermal conduction, and the step-by-step minimization of $\\sigma_{t+1}^2$ with respect to $\\eta_t$ yields the optimal $1/t$ decay schedule.","core_discovery":"The central discovery is that the fast valley dynamics of stochastic gradient descent and sign descent converge to a Gaussian steady state whose width scales as $\\sigma \\approx \\sqrt{\\eta/(2a)}\\,\\sigma_g$ in the flat limit, so the averaged fast loss $\\ell_f = \\frac12 a\\sigma^2 = C\\eta$ is independent of the sharpness $a$. This is the neural analog of the equipartition theorem: every valley direction contributes the same thermal loss, so the total is $L_f = N C \\eta$, with the learning rate acting as temperature and the slope $C$ as heat capacity. The paper then uses this identity to explain why the final validation loss of a GPT-style model is linear in the final learning rate, and derives the optimal decay schedule by choosing $\\eta_t$ at each step to minimize the next Gaussian width, obtaining $\\eta_t = (\\eta/2)/(1 + t/t_h)$. The slow river dynamics is shown to feel an entropic force proportional to the gradient of $-\\frac12 \\log a$, so valleys that narrow along the river can trap the optimizer.","pith_inferences":["Extension: the inferred count $N \\approx 5\\times 10^6$ valley directions is a concrete microstate count; if the thermodynamic picture is right, the Hessian spectrum of a similarly trained GPT-style model should show roughly five million high-curvature directions, which an independent Hessian measurement could confirm.","Extension: the equipartition law predicts a precise batch-size trade-off, because gradient noise scales as $\\sigma_g \\propto 1/\\sqrt{B}$ for SGD, making the thermal slope $C$ scale as $1/B$ (or $1/\\sqrt{B}$ for sign updates); scanning batch size at fixed $\\eta_{\\min}$ would test this prediction directly.","Extension: the framework predicts that entropic trapping grows stronger as training proceeds and valleys sharpen, so the near-perfect loss-curve alignment seen in the early GPT experiment may fail at much larger compute budgets, which would be visible as a growing misalignment in learning-rate-sum plots."],"forward_implications":["Final loss separates into a river part controlled by the learning-rate sum $D$, a thermal part controlled by the final learning rate $\\eta_{\\min}$, and two correction terms from entropic forces and incomplete annealing: $\\ell_{\\mathrm{final}} = \\ell(D,\\eta_{\\min}) + \\Delta_{\\mathrm{entropic}} + \\Delta_{\\mathrm{anneal}}$.","Because $1/\\sigma_t^2$ grows as an arithmetic sequence under the optimal schedule, the thermal loss decays as $1/t$ and the learning rate as $\\eta_t = (\\eta/2)/(1+t/t_h)$, so $1/t$ decay is optimal for isotropic valleys rather than linear or cosine decay.","The first decay step should jump from $\\eta$ to roughly $\\eta/2$; continuous learning-rate schedules that start decaying smoothly from $\\eta$ are suboptimal because neither keeping $\\eta$ nor dropping to $0$ decreases the variance.","Decaying all the way to zero is suboptimal because reaching $\\eta_{\\min}=0$ would take infinite time; the decay time needed to reach a small $\\eta_{\\min}$ is bounded by $O(1/\\eta_{\\min})$, independent of the stable learning rate.","Increasing the stable learning rate lowers the river loss by raising the learning-rate sum $D$ without requiring a longer decay phase, which is why a larger stable learning rate improves validation loss up to numerical stability limits."],"supporting_citations":[{"why":"It supplies the river-valley loss-landscape premise that motivates the fast/slow decomposition and the WSD analysis.","marker":"[5]"},{"why":"It introduces the warmup-stable-decay training recipe whose stable and decay phases are mapped to equilibrium and annealing.","marker":"[8]"},{"why":"It defines the edge-of-stability boundary $a=2/\\eta$ that limits the validity of the Gaussian steady-state formula.","marker":"[9]"},{"why":"It reports that 1-sqrt decay outperforms linear or cosine decay, an observation the derived $1/t$ schedule explains.","marker":"[10]"},{"why":"It provides the small transformer implementation used for the GPT validation-loss measurements.","marker":"[11]"},{"why":"It establishes the learning-rate-sum control of loss curves that the paper uses to probe entropic forces and the final-loss decomposition.","marker":"[12]"},{"why":"It documents that decaying the learning rate to zero is suboptimal, which the optimal-schedule analysis accounts for.","marker":"[13]"}],"fun_headline_variants":["Training LLMs like thermodynamics: loss = heat capacity × LR","Neural thermodynamics: why LLM loss grows with learning rate","Equipartition theorem for LLMs: loss = LR × heat capacity","LLM training thermodynamics: optimal 1/t LR schedule","Loss follows thermodynamics: linear in LR, optimal 1/t decay"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The empirical identification of the fitted slope $110$ as purely thermal loss requires that the slow river loss $\\ell_s$ does not change across the decay schedules, so that the measured final validation loss is a proxy for the thermal loss $\\ell_f$; if $\\ell_s$ shifts with $\\eta_{\\min}$, the inferred valley-direction count and the thermodynamic confirmation lose their quantitative meaning.","fun_headline_variants_meta":{"raw":{"variants":["Training LLMs like thermodynamics: loss = heat capacity × LR","Neural thermodynamics: why LLM loss grows with learning rate","Equipartition theorem for LLMs: loss = LR × heat capacity","LLM training thermodynamics: optimal 1/t LR schedule","Loss follows thermodynamics: linear in LR, optimal 1/t decay"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3388,"prompt_tokens":851,"completion_tokens":2537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2449}},"tokens_in":467,"tokens_out":2537,"duration_ms":16620,"temperature":1.0,"reasoning_tokens":2449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:07:23.841641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan the final learning rate $\\eta_{\\min}$ while varying the decay-phase length (for example, 1k, 3k, and 10k steps): the thermodynamic claim predicts that the fitted slope in $\\ell = 3.145 + 110\\eta_{\\min}$ stays the same whenever the decay is long enough for quasi-static annealing, whereas a slope that changes with decay length would show that the constant-$\\ell_s$ assumption fails and the equipartition reading is unsupported.","supporting_citations":[{"cited_title":"Scaling laws and compute-optimal training beyond fixed training durations","cited_arxiv_id":null,"evidence_quote":"It reports that 1-sqrt decay outperforms linear or cosine decay, an observation the derived $1/t$ schedule explains."},{"cited_title":"A multi-power law for loss curve prediction across learning rate schedules","cited_arxiv_id":null,"evidence_quote":"It establishes the learning-rate-sum control of loss curves that the paper uses to probe entropic forces and the final-loss decomposition."},{"cited_title":"modded-nanogpt","cited_arxiv_id":null,"evidence_quote":"It documents that decaying the learning rate to zero is suboptimal, which the optimal-schedule analysis accounts for."}],"review_version":1}