{"id":"47525c3a-fcaa-4fac-a311-d622cbb9b554","arxiv_id":"2607.27805","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Greedy LLM inference shows bounded, jump-like sensitivity to sub-token perturbations that the authors interpret as chaotic, with attention expanding and normalization suppressing perturbations.","lead":"A physics group treated a 1.5-billion-parameter language model as a chaotic dynamical system, showing that tiny hidden-state nudges stay invisible for a while and then abruptly flip the model's output. They argue attention stretches these perturbations while normalization keeps the dynamics bounded — evidence, they say, that LLM 'thinking' has chaotic signatures.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Chaos claim lacks its defining signature: pre-flip hidden-state divergence is non-exponential; jump-like token flips may be threshold effects rather than chaotic stretching.","rationale":"The reader's CONDITIONAL verdict already identifies the missing Lyapunov exponent as a core weakness. My analysis sharpens it into a concrete internal tension: the only clean signal (latent phase, identical tokens) shows contraction, while the post-flip divergence is contaminated by context feedback. This makes the chaos classification rest on a threshold/branching phenomenon rather than on exponential stretching of the continuous map. The operator-level Jacobian results are exact and valuable, but isolated singular values >1 do not imply positive Lyapunov exponents for the composed map; the product of Jacobians along the orbit can still contract in all directions. The recurrence plots and correlation dimensions are qualitative and methodologically vulnerable (fixed recurrence rate, no Theiler window/surrogates), so they cannot independently support the chaos label. Therefore the single decisive experiment is the token-locked divergence test: it directly estimates the maximal Lyapunov exponent of the hidden-state dynamics conditioned on a fixed token sequence, eliminating the discrete-feedback confound. If the test fails, the abstract's 'demonstrate ... chaos' overclaims; if it passes, the paper's conclusion is substantially supported. Since the reader's verdict was already CONDITIONAL, my read leaves that verdict unchanged while making the condition explicit.","tokens_in":21015,"tokens_out":7945,"duration_ms":74981,"concrete_test":"Token-locked divergence analysis. Generate a greedily-decoded reference trajectory from an unperturbed prompt. For an ensemble of N=50 perturbed initial embeddings at the same r values used in the paper, run the model while forcing the token choice at every step to equal the reference token (e.g., zero out all non-reference logits). Compute the averaged hidden-state distance (cosine and Euclidean) between reference and perturbed runs over the first L=200 steps, avoiding saturation. Fit d(t)=d0 exp(λt) over the pre-saturation interval and report λ with confidence intervals. If λ is not significantly positive, or if distances decay/saturate, the hidden-state map is not chaotic and the paper's central classification should be weakened; if λ>0 robustly, the chaos claim survives. This directly supplies the missing maximal Lyapunov exponent while excluding token-feedback contamination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is deterministic chaos, defined in Sec. III.A by exponential divergence of nearby states (Eq. 3, λ>0). Yet Sec. V.B concedes that 'the discreteness of the token space prevents a robust global estimate of the maximal Lyapunov exponent' and that 'we do not observe a smooth increase of the distance but instead it is always characterized by large jumps.' This is not a mere technical gap: the defining quantitative signature is absent. Worse, the observed latent-then-jump pattern is confounded by the discrete token bottleneck. In the latent phase both trajectories generate identical tokens, and the paper reports the hidden-state distance may shrink with a negative finite-time exponent (Fig. 7). If the continuous hidden-state map were chaotic, same-token-conditioned trajectories should still diverge exponentially on average. After the first token flip the two runs have different context tokens, so all subsequent separation is caused by different exogenous inputs rather than by intrinsic divergence of nearby states in the same system. Thus the evidence is compatible with a stable (or weakly contracting) continuous map whose decisions occasionally cross a threshold, after which feedback produces large differences. Boundedness plus input-sensitivity, without a demonstrable positive Lyapunov exponent for the hidden-state dynamics, does not establish chaos.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper treats autoregressive LLM inference as a discrete-time dynamical system in hidden-state space. It injects small norm-controlled perturbations into initial token embeddings, generates trajectories under greedy decoding, and measures divergence in both hidden-state and sentence-embedding spaces. The central empirical pattern is 'latent-then-jump': two trajectories produce identical tokens for a while (with possibly shrinking hidden-state distance), then a single token flip is followed by rapid separation and saturation. Supporting analyses include recurrence plots, Grassberger–Procaccia correlation dimensions, an attention-window ablation, and an operator-level Jacobian decomposition of SwiGLU, attention, RMSNorm, and residual connections, plus a mean-field formula for the feed-forward expansion. The paper concludes that LLMs exhibit 'hallmark signatures of chaos' and operate near the edge of chaos, with attention/MLP layers stretching perturbations and normalization/residual connections folding or preserving them.","tokens_in":21225,"tokens_out":5668,"duration_ms":54394,"significance":"If the chaos interpretation could be made rigorous, this would be a significant contribution: it would connect the nonlinear coupling of self-attention to sensitivity, boundedness, and possible edge-of-chaos operation in trained LLMs, with implications for interpretability, robustness, and reproducibility. The paper has real strengths: the Jacobian derivations in Sec. V.A (Eqs. 29–32) are exact and clearly presented; Fig. 3 shows the mean-field formula tracks the measured expansion across layers; the attention-window experiment (Fig. 11) is a constructive causal test identifying attention as the carrier of divergence; and the two-model replication plus multi-metric consistency are valuable. However, the decisive quantitative evidence for deterministic chaos — a positive Lyapunov exponent on the continuous hidden-state map, or an equivalent discriminator against non-chaotic threshold models — is missing, and the recurrence and dimension analyses rely on methods known to overcall chaos on correlated stochastic signals. The contribution is therefore promising but currently under-supported at the central claim.","major_comments":[{"comment":"The defining criterion for chaos stated in Eq. (3) is a positive maximal Lyapunov exponent. The paper explicitly concedes that 'the discreteness of the token space prevents a robust global estimate of the maximal Lyapunov exponent' and that divergence is 'always characterized by large jumps' rather than smooth exponential growth. The observed latent-then-jump pattern, including negative finite-time exponents before the first token flip (Fig. 7), is exactly what a stable or weakly contracting continuous update plus a threshold readout would produce; after the first token flip the two trajectories are driven by different context tokens and their separation no longer measures divergence of nearby states in the same system. To support the chaos claim, the authors need a positive Lyapunov exponent (or finite-time Lyapunov exponent) estimated on same-token-conditioned segments, and/or a quanti","section":"V.B, Eq. (3)"},{"comment":"The recurrence-plot evidence is qualitative: thresholds are fixed by recurrence rate (3% or 5%), and the comparison with Lorenz and stochastic panels is visual. Diagonal-line structures can occur in non-chaotic, quasi-periodic, or even noise-driven processes after coarse-graining. Provide recurrence quantification statistics (determinism, laminarity, diagonal-line length distribution) computed from the RPs, and compare against surrogate data (e.g., Fourier or phase-randomized surrogates preserving autocorrelation) and against a null model. Without this, 'structural similarities' do not discriminate chaos from other dynamical classes.","section":"V.D, Fig. 12"},{"comment":"The Grassberger–Procaccia correlation dimension is applied to raw, serially correlated trajectory points without Theiler-type temporal decorrelation, explicit scaling-region selection, or surrogate analysis. For short, high-dimensional, autocorrelated trajectories, the slope of C(ε) can mimic a low fractal dimension even for stochastic or quasi-periodic signals. The random-vector comparison in Fig. 14 is not a surrogate control because it destroys temporal correlations. Report D2 with a Theiler window, bootstrap confidence intervals, and surrogate comparisons; otherwise the fractal-attractor conclusion is unsupported.","section":"V.E, Eq. (21)"},{"comment":"The text and conclusion describe Eq. (30) as a 'predictive, weight-only' formula, but it depends on measured activation densities s_x,d_x from the same forward pass used for validation (Fig. 3). As a closed-form reorganization of the Jacobian it is useful, but as a predictor it is circular. Test it on held-out inputs or reframe it as a diagnostic decomposition; also state explicitly which quantities are static weights and which are data-dependent.","section":"V.A.1.a, Eq. (30)"},{"comment":"The perturbation magnitude r∈[2.5×10^-4, 5×10^-4] overlaps or lies below the reported FP16 activation spacing (≈4.88×10^-4 in [0.5,1.0]-scale activations). The paper does not show that the injected embedding-space perturbation remains above the rounding noise of the actual pipeline. Please quantify the effective perturbation after FP16 quantization, compare with a perturbation-free control (e.g., two runs with identical inputs under FP32/FP64), and verify the divergence phenomenology in higher precision. This is necessary to rule out numerical noise as the source of the 'nearby' trajectory differences.","section":"Appendix D / V.A.5"}],"minor_comments":[{"comment":"The caption uses 'Lyapunov timeseries' and 'lyapunov exponent' for what is actually a finite-time local slope of log-distance. Please use 'finite-time exponent' or 'local divergence rate' to avoid conflating this with the true Lyapunov exponent.","section":"Fig. 7 caption"},{"comment":"The equation uses a continuous derivative d/dt but the data are discrete. Specify the finite-difference estimator used to compute the slope.","section":"Eq. (33)"},{"comment":"The non-monotonic attention-window curve is a key causal claim, but the figure appears to show a single ensemble/prompt. Add error bars, number of prompts, and statistical significance, and state the perturbation magnitude used (the text says 10^-4, while Appendix D lists 3.5–4×10^-4).","section":"V.C, Fig. 11"},{"comment":"The phrase 'hallmark signatures of chaos' is stronger than what Sec. V.B delivers. Consider phrasing such as 'chaos-like signatures' or 'sensitivity consistent with chaos' unless a Lyapunov-based discriminator is added.","section":"Abstract and Conclusions"},{"comment":"The statement 'available from the corresponding author upon reasonable request' limits reproducibility. I recommend releasing code, trajectories, and recurrence plots, especially since several thresholds and scaling regions are chosen manually.","section":"Data Availability"},{"comment":"Reference [48] appears as a numbered footnote-like entry rather than a proper reference; several preprints are undated or have future arXiv numbers (e.g., [15], [17], [18]) — please add accessible identifiers and dates in the final version.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for a nonlinear-dynamics journal, and the Jacobian/operator-level analysis is solid and publishable as a standalone contribution. The obstacle is the overreach from 'sensitivity plus boundedness' to 'chaos' without a Lyapunov-based discriminator; the recurrence and dimension evidence, as analyzed, will not persuade practitioners. I would encourage the authors to either substantially strengthen the chaos diagnostics (continuous-state Lyapunov estimates, surrogates, Theiler decorrelation, precision controls) or reframe the paper's claim to 'sensitive, bounded, non-periodic dynamics in LLM inference' with chaos as an open hypothesis. No concerns about novelty or citation integrity beyond the minor reference issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: read this for the exact Jacobian decomposition of the transformer sub-blocks and the attention-window ablation; be wary of the word \"chaos\" until the authors actually measure it. The paper does good, honest work on where perturbations grow and shrink per layer, and then stretches it into a classification the evidence doesn't support.\n\nWhat's genuinely new: Sec. V.A derives the Jacobian of each architectural piece — attention (split into nonlinear routing vs. static mixing), SwiGLU MLP, RMSNorm as a scaled orthogonal projection, residual as identity — and measures spectral norms per layer. That part is exact, clean, and consistent with the mean-field tradition. The attention-window ablation (Sec. V.C), showing non-monotonic first-divergence time vs. window size, is a nice causal probe pointing to attention as the carrier of perturbation growth. The mean-field formula for MLP expansion is a useful diagnostic, though it borrows activation densities from the same forward pass, so \"predictive\" oversells it.\n\nWhere the paper overreaches: the central claim is deterministic chaos. The defining signature — a positive maximal Lyapunov exponent — is explicitly unavailable; Sec. V.B concedes that discreteness of the token space prevents a robust global estimate, and the paper also admits the distance growth \"is always characterized by large jumps\" rather than smooth exponential divergence. In the latent phase the finite-time exponent is often negative. After the first token flip, the two runs have different contexts, so subsequent separation comes from different inputs rather than intrinsic stretching along the same map. That is a threshold-and-feedback story, not chaos. The Lorenz-like recurrence plots and GP dimension slopes without Theiler or surrogate corrections are suggestive but not conclusive. The perturbation magnitude sits at the stated FP16 floor, with some fuzziness about which space the floor applies to, and the second-model replication rests on an unpublished thesis. These are not minor complaints — they target the paper's thesis — but they don't destroy the descriptive phenomenology.\n\nWho gets value: people working on LLM reproducibility, interpretability, and the dynamics of transformers. The operator-level Jacobian results and the window-size dependence are likely to survive scrutiny. The \"edge of chaos\" interpretation is a hypothesis in need of stronger evidence. I'd send it to a serious referee: the good parts deserve review, and the overclaim needs to be confronted. The path forward is clear — surrogate testing, Theiler-corrected dimensions, quantitative recurrence statistics, a precise statement about numerical precision, and code/data release — or a reframed conclusion as bounded sensitivity with threshold dynamics.","headline":"Solid operator-level Jacobian and attention-window results, but the 'chaos' verdict is an interpretation the evidence doesn't force — the defining Lyapunov exponent is missing by the authors' own admission.","tokens_in":21820,"tokens_out":2506,"would_cite":true,"duration_ms":25706,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37D45","37M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that chain-of-thought LLM inference behaves like a bounded chaotic dynamical system: tiny prompt perturbations stay latent, then abruptly flip token choices and cause jump-like divergence, with self-attention and feed-forw","keywords":["chaos theory","large language models","Lyapunov exponents","recurrence plots","fractal dimension","self-attention","sensitivity to initial conditions","edge of chaos"],"falsifier":"Measure the first-divergence time as a function of perturbation radius r across the full linear-response range (r from the FP16 floor upward). In a genuinely chaotic system, halving r should delay divergence by a roughly constant number of tokens (logarithmic scaling); if instead the divergence time is essentially independent of r (or scales as a power law with exponent far from the chaotic prediction), the latent-then-jump pattern would be better explained by a deterministic threshold effect or rounding artifact than by chaotic amplification.","tokens_in":20750,"feed_emoji":"🌀","tokens_out":5739,"duration_ms":52258,"temperature":0.7,"pith_summary":"This paper treats chain-of-thought generation by a large language model as a discrete-time dynamical system and argues that its internal dynamics are chaotic. A tiny perturbation to the initial prompt embedding stays invisible in the output for many tokens, then abruptly flips a single token choice, after which the two trajectories separate rapidly while remaining bounded. The authors back this with an exact Jacobian analysis of the Transformer's building blocks: self-attention and feed-forward layers amplify perturbations, while normalization and residual connections confine them, producing the stretching-and-folding mechanism of classical chaos. They also find recurrence plots reminiscent of canonical chaotic attractors and lower-dimensional fractal structure in the final layers, leading them to place LLM inference at the edge of chaos — a regime that could explain why models are simultaneously stable and sensitive to small input changes.","feed_headline":"Chain-of-thought LLMs behave as chaotic systems, study says","feed_subtitle":"Small prompt changes stay hidden, then abruptly flip outputs, mimicking chaotic attractors.","key_machinery":"The central object is the hidden-state trajectory of the most recent token across generation steps, treated as a discrete-time map, together with the exact Jacobian of each Transformer sub-block. Self-attention and the feed-forward network are the stretching components (the maximum stretching factor of their Jacobians exceeds one at every layer); normalization acts as a scaled orthogonal projection that removes the radial, most-expansive perturbation component; and the residual connection preserves the surviving perturbation so it can accumulate across tokens. The first-divergence time — how many tokens pass before two perturbed trajectories pick different tokens — is the paper's characteris","core_discovery":"The central claim is that LLM inference is a bounded, sensitive, aperiodic dynamical system rather than a stable or purely stochastic process. Concretely, the paper demonstrates a latent-then-jump divergence pattern: after a sub-token perturbation is injected into the initial embedding, both trajectories agree on tokens for a while while the perturbation accumulates in hidden states; once it crosses the token-sampling threshold, a different token is chosen, and that single discrete difference feeds back into the context, causing fast, jump-like separation until distances saturate. The authors argue this matches the discrete-space picture of sensitivity to initial conditions and support it wi","pith_inferences":["If the edge-of-chaos picture is right, test-time choices like temperature, context length, or quantization would tune a model's position on the order-chaos boundary, which could predict when reasoning becomes brittle and when it becomes repetitive — a directly testable consequence.","The latent-then-jump divergence gives a cheap probe for 'latent reasoning': by measuring hidden-state divergence before token divergence, one could detect when a model is forming a decision without committing to output, possibly aiding interpretability.","The layer-wise transition from random-like early embeddings to structured, lower-dimensional late dynamics suggests that the fractal-dimension diagnostic could serve as a measure of how deeply a model is 'thinking' on a given prompt; this could be compared across prompts of different complexity.","The non-monotonic attention-window effect predicts that extending context length in modern LLMs may systematically alter the chaoticity of generation, which could be checked by comparing first-divergence time distributions across models with different context lengths."],"forward_implications":["Tiny numerical differences in prompts or floating-point rounding can cascade into very different generations, so deterministic-looking LLM inference is intrinsically sensitive to microscopic perturbations.","The effective dynamical state of an LLM is not a single hidden vector but a window of past states, so future analyses should use delay-coordinate or sliding-window representations rather than instantaneous states.","The stretch-then-fold layout — expansion through the middle layers followed by contraction before the output — may be a general property of trained Transformer language models, not an artifact of one architecture.","Because stochastic sampling (temperature > 0) disrupts the last-layer chaotic structure while leaving early layers unchanged, decoding strategies control not just output diversity but the underlying dynamical regime.","Lyapunov exponents and fractal dimensions in discrete token spaces cannot be measured as precise invariants; they function as comparative diagnostics across layers, prompts, and models."],"fun_headline_variants":["LLM hidden states are chaotic: tiny tweaks, big flips","LLM chaos: latent drift, then sudden token flips","Tiny prompt tweaks flip LLM outputs in chaotic bursts","Chaos in LLM reasoning: small changes, abrupt output jumps"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's chaos classification rests on treating the observed latent-then-jump divergence of discrete token trajectories as the discrete-space signature of sensitivity to initial conditions, even though the discreteness of the token space prevents a direct global Lyapunov-exponent estimate.","fun_headline_variants_meta":{"raw":{"variants":["LLM hidden states are chaotic: tiny tweaks, big flips","LLM chaos: latent drift, then sudden token flips","Tiny prompt tweaks flip LLM outputs in chaotic bursts","Chaos in LLM reasoning: small changes, abrupt output jumps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001538,"raw_usage":{"total_tokens":5958,"prompt_tokens":678,"completion_tokens":5280,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":5215}},"tokens_in":422,"tokens_out":5280,"duration_ms":29637,"temperature":1.0,"reasoning_tokens":5215,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:06:00.512659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the first-divergence time as a function of perturbation radius r across the full linear-response range (r from the FP16 floor upward). In a genuinely chaotic system, halving r should delay divergence by a roughly constant number of tokens (logarithmic scaling); if instead the divergence time is essentially independent of r (or scales as a power law with exponent far from the chaotic prediction), the latent-then-jump pattern would be better explained by a deterministic threshold effect or rounding artifact than by chaotic amplification.","supporting_citations":[],"review_version":1}