{"id":"8a9be781-0618-446f-88d0-a39bba162c90","arxiv_id":"2412.01417","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Transformers trained on random elementary cellular automata can predict unseen rules fairly well one step ahead, but multi-step planning degrades unless the model is deeper or trained with future-state or rule prediction losses.","lead":"This paper trains Transformer models to predict the next state of random elementary cellular automata and shows they can generalize to rules never seen in training. It reports that multi-step planning accuracy drops sharply and that deeper models and auxiliary rule prediction improve longer rollouts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-scaling claim in Fig. 3 is confounded by parameter count and an undefined accuracy threshold; the layer requirements are not established.","rationale":"The reader's weakest assumption concerned the absence of error bars and single-run threshold readings in the depth experiment. My analysis agrees that the depth-scaling claim is the most load-bearing, but sharpens the concern: the independent variable (layer count) is confounded with parameter count and, likely, optimization budget because the paper does not report training details or controls. Even with error bars, the result would not establish a computational-depth requirement unless a matching-parameter shallow-wide model is tested. The reader's suggestion of a non-transformer baseline is useful but less directly targeted; a wide-transformer control is the minimal experiment that would resolve the confound. The generalization-across-rules finding is supported by the train/test rule split and high next-step accuracy (0.96 vs 0.5 chance), so that part of the central claim appears sound. Therefore the overall CONDITIONAL verdict remains appropriate: the paper should either add the controls and clearly defined thresholds or soften the depth claim. My read does not change the verdict.","tokens_in":6817,"tokens_out":10897,"duration_ms":100391,"concrete_test":"Retrain O-O models for layer counts {1,2,4,7,10} with d_model=512, 8 heads, using ≥5 random seeds each and identical, reported training budget and LR schedule. Pre-register an accuracy threshold (e.g., mean per-bit accuracy ≥0.90 averaged over the test set) for 'solving' each planning step. Additionally train matched-parameter baselines with different depth/width trade-offs under the same budget: (a) 1 layer, d_model=1024 (≈ parameters of 4-layer d=512); (b) 2 layers, d_model=724 (~same params); (c) 4 layers, d_model=512 plus double training steps. If the shallow-wide models solve step 4 at 1–2 layers, or if threshold crossings across seeds vary by >1 layer, the depth-scaling conclusion is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that each additional planning step requires more computational layers rests solely on Figure 3, which shows O-O accuracy as a function of layer count (1–11) at fixed d_model=512 and 8 heads. This manipulation changes parameter count and training dynamics along with depth, and no training budget, learning-rate schedule, seeds, or error bars are reported. The paper never states the accuracy threshold used to declare that a step is 'predicted' (e.g., 0.9?), so the threshold readings (step 2 at 4 layers, step 3 at 7, step 4 at 10) are unverifiable and may be artifacts of optimizer convergence or capacity. No shallow-wide control (e.g., 1–2 layers with larger d_model) was run, so the result cannot distinguish 'depth enables sequential computation' from 'more parameters help'. Since the abstract and conclusion use this finding to motivate recurrence and adaptive computation for LLMs, the depth conclusion is load-bearing; if it fails, a key practical implication of the paper is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies whether Transformer encoders can learn the local rules of Elementary Cellular Automata (ECA) from observed orbits and use them for multi-step planning. Four training objectives are compared: next-state prediction (O-S), multi-step prediction (O-O), joint state and rule prediction (O-SR), and state prediction given the rule (RO-S). The authors report that the model reaches about 0.96 per-bit accuracy for next-state prediction on held-out rules, that look-ahead accuracy drops for k≥2, that O-SR improves longer look-ahead over O-O, that autoregressive rollouts are more accurate than direct look-ahead, and that deeper models are needed to predict more steps ahead, with step 1 needing 2 layers, step 2 needing 4, step 3 needing 7, and step 4 needing 10 layers.","tokens_in":6995,"tokens_out":8508,"duration_ms":71688,"significance":"If the claims hold, the paper offers a clean testbed for studying how Transformers abstract discrete rules and a concrete demonstration that multi-step planning scales with depth, with implications for training objectives (longer horizons, rule prediction) and architecture (recurrence, adaptive computation) in LLMs. The setup is well chosen: test rules are disjoint from training rules, so next-state accuracy is evidence of generalization rather than memorization, and the authors provide code and data (GitHub link). The main limitation is that the experimental evidence is currently too thin—single runs, no error bars, and an undefined accuracy threshold for the depth scaling—so the central claims are plausible but not yet established.","major_comments":[{"comment":"The depth-scaling conclusion that \"each additional planning step requires more computational layers\" is not supported as presented. The thresholds (step 2 at 4 layers, step 3 at 7, step 4 at 10 layers) are read from single runs with no stated accuracy criterion for \"predicting a step,\" and no error bars or seeds are reported. Additionally, because d_model=512 and 8 heads are held fixed while the layer count changes from 1 to 11, the experiment conflates depth with parameter count and training dynamics; a shallow-wide control is not reported. This point is load-bearing because the abstract and conclusions use this result to motivate recurrence and adaptive computation time.","section":"3 Results and Discussion (Figure 3)"},{"comment":"All accuracy figures appear to come from a single training run per condition; no random seeds, error bars, or significance tests are reported. Consequently, comparative statements such as \"the O-SR model outperformed the O-O model\" (0.85 vs 0.75 for k=2,3) and the \"near-perfect accuracy\" of RO-S for k=0,1 are not quantitatively supported. Please report at least a small number of seeds with mean and standard deviation, or otherwise justify the stability of the observed differences.","section":"3 Results and Discussion (Figures 2 and 3)"},{"comment":"The training configuration is underspecified: the paper does not state the optimizer, learning rate and schedule, batch size, number of training steps, or compute budget, and for the O-SR objective it does not specify how the state and rule losses are weighted. Without these details the experiments are not reproducible, and the claim that rule prediction in the loss \"enhances the ability to form internal representations\" cannot be separated from arbitrary training choices. Please add full training details and, ideally, a sensitivity check on the loss weight.","section":"2 Methods"}],"minor_comments":[{"comment":"The text refers to \"Figure 3A,\" but the figure has no panel A; the caption reads \"Figure 3: Adding layers improves prediction of ECA orbit.\" Please correct the cross-reference.","section":"3 Results and Discussion"},{"comment":"The coupon-collector estimate Tmin treats each cell-time observation as an independent draw, but adjacent cells in the same time step have overlapping neighborhoods, so the draws are correlated. The agreement with Figure 2A is therefore heuristic; please state this caveat or provide a more careful estimate.","section":"2 Methods"},{"comment":"The definition of \"look-ahead steps\" could be clearer: Figure 2B includes k=0 while the text defines k∈{1,2,3}; please state explicitly what k=0 denotes (presumably next-state prediction) and how \"per-bit accuracy averaged over free runs\" is computed.","section":"2 Methods / Figure 1"},{"comment":"The O-SR model's rule prediction accuracy is never reported, so the claim that it forms an internal representation of the rule is only indirectly evidenced. Reporting the rule-prediction accuracy (and how rule outputs are decoded) would strengthen the mechanistic interpretation.","section":"3 Results and Discussion"},{"comment":"In Figure 2D, the comparison between autoregressive (AR) and look-ahead (LA) predictions needs a precise definition of both modes; in particular, for AR, specify whether the model consumes its own previous output tokens as context and whether teacher forcing is used at any stage.","section":"3 Results and Discussion"},{"comment":"The phrase \"we also confirmed that the model's depth plays a crucial role\" overstates the evidence given the single-run, threshold-free analysis in Figure 3; please soften the wording to match the strength of the results.","section":"4 Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a compact workshop preprint with interesting results but limited methodological detail. For a journal submission, the authors would need to significantly expand the experimental reporting; I would recommend major revision rather than rejection because the core questions are worth pursuing and the existing code/data make the required additions feasible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the ECA Transformer preprint carefully. The stress-test note is right: the depth-scaling claim in Fig. 3 is not established, and that matters because it drives the paper's practical message about recurrence and adaptive computation. But before that, let me give credit: the four-task comparison (O-S, O-O, O-SR, RO-S) on held-out rules is a clean, useful setup. The finding that adding future-state or rule-prediction losses improves multi-step look-ahead is plausible and consistent with existing intuitions about auxiliary objectives. The most interesting datapoint for me is that the RO-S model, given the rule explicitly, still degrades sharply beyond two steps - that is a real constraint story about transformers, not about rule inference. The held-out rule split and the coupon-collector estimate for Tmin are nice touches.\n\nThe soft spots are mostly about evidence thickness. Every figure appears to come from a single run: no error bars, no seeds, no significance tests. For the qualitative curves that is bad but tolerable. For the depth claim, it is disqualifying as stated. Fig. 3 changes layer count at fixed d_model and head count, so parameter count and training dynamics change with depth. There is no shallow-wide control, no stated accuracy threshold for 'begins to predict', and no training budget or LR schedule. The specific readings - step 2 at 4 layers, step 3 at 7, step 4 at 10 - are presented as facts but are read off single curves. That is overclaiming.\n\nThere is also a smaller overstatement: the abstract says models 'abstract the underlying rules', but the evidence is generalization over a sample of arity-5 functions. That is within-class inductive generalization, not demonstration of a mechanism. The authors hedge later, so it's an abstract-level exaggeration rather than a fatal flaw.\n\nCitation pattern is fine; prior CA-and-transformer work is cited and the contribution is positioned as an extension. Code and data are promised at a GitHub link, but with no commit or config, reproduction requires mailing the author.\n\nWho is this for? Researchers working on multi-step planning in transformers or on toy dynamical systems as reasoning testbeds. They will find the task definitions useful and the depth caution instructive. A serious referee could fix the paper by asking for seeds, error bars, a shallow-wide control, and a softening of the depth conclusion. I would send it to review rather than desk-reject.","headline":"Clean toy-domain task comparison, but the depth-scaling result is not supported by the reported single-run experiments.","tokens_in":7519,"tokens_out":3575,"would_cite":true,"duration_ms":30092,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37B15","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a Transformer trained on random Elementary Cellular Automaton orbits abstracts the underlying Boolean rule and that the ability to plan multiple steps ahead is controlled by the number of layers.","keywords":["elementary cellular automata","transformers","rule abstraction","multi-step planning","depth scaling","autoregressive generation","chain-of-thought","Boolean functions"],"falsifier":"Repeat the O-O depth experiment with multiple random seeds and learning-rate schedules at layer counts 2, 4, 7, and 10, and record accuracy distributions. If the apparent thresholds shift by more than a layer or disappear when seeds vary, the claim that each planning step requires additional layers would not be supported; alternatively, a wider 4-layer model that predicts step 3 as well as the 7-layer model would falsify the depth-specific conclusion.","tokens_in":6600,"feed_emoji":"🧩","tokens_out":4979,"duration_ms":41357,"temperature":0.7,"pith_summary":"This paper asks whether a Transformer can infer the local rule of an Elementary Cellular Automaton from raw orbit data and then use that rule to plan ahead. It claims that a Transformer trained on random orbits with held-out rules generalizes across Boolean functions of fixed arity, so the network is abstracting a rule class rather than memorizing examples. The paper also claims that adding future-state or rule prediction to the training loss helps the model form better internal rule representations, which improves longer-horizon prediction and autoregressive generation. Finally, it reports a depth-scaling effect: predicting each additional step into the future requires more Transformer layers, suggesting that fixed-depth models are the bottleneck for multi-step planning.","feed_headline":"More planning steps demand more Transformer layers","feed_subtitle":"Training on random orbits lets Transformers abstract Boolean rules; deeper models plan further ahead.","key_machinery":"The central object is the Elementary Cellular Automaton (ECA), a one-dimensional lattice of binary cells updated synchronously by a local rule $\\rho: \\{0,1\\}^{2r+1}\\to\\{0,1\\}$; with $W=20$ and $r=2$ there are $2^{32}\\approx4.3\\times10^9$ possible rules. The machinery is a Transformer encoder with masked tokens standing for unknown future bits or the unknown rule, trained under four task variants: O-S (orbit to next state), O-O (orbit to several future states), O-SR (orbit to next state plus rule), and RO-S (given orbit and rule, predict state). The rule-recovery estimate $T_{\\min}=2^{2r+1}(\\ln 2^{2r+1}+\\gamma)/W\\approx6.47$ from the coupon collector's problem tells how many orbit steps are needed, and the depth sweep on the O-O task is what carries the claim that each planning step needs more layers.","core_discovery":"On the paper's own terms, the central discovery is that a Transformer encoder with full self-attention can learn to predict the next state of an Elementary Cellular Automaton from a ten-step orbit and, more strongly, can do so for local rules never seen in training, meaning it has abstracted the underlying Boolean function rather than memorized a specific rule. When asked to jump several steps ahead without intermediate context, accuracy falls sharply, from about 0.96 for next-state prediction to below 0.75 for two or three steps ahead. Training with an orbit-to-orbit objective or with simultaneous rule prediction both improve planning, with the rule-prediction variant (O-SR) generalizing better at longer horizons, while explicitly supplying the rule (RO-S) does not rescue multi-step jumps. A depth experiment shows that four-step planning requires about ten layers, leading the author to conclude that sequential computation in Transformers is bounded by depth and that recurrence or adaptive computation time are promising remedies.","pith_inferences":["If the depth-to-horizon scaling holds beyond this toy setup, fixed-depth models should systematically fail on planning problems whose solution requires more sequential rule applications than the network has layers; that is a testable prediction for arithmetic or multi-hop reasoning benchmarks.","The O-SR result suggests rule-identification losses could act as a regularizer that compresses the hidden state into a causal rule; one could test this by probing the model's hidden vectors for the rule and correlating probe accuracy with planning accuracy.","The coupon-collector estimate implies that the number of orbit steps needed to identify a rule grows exponentially in neighborhood radius; scaling ECAs to larger $r$ should make rule abstraction and planning degrade in a predictable way.","Because RO-S fails at multi-step jumps despite having the rule, the bottleneck is not rule knowledge but the propagation of intermediate states; architectures with explicit memory or recurrent state updates should be tested on the same task."],"forward_implications":["Transformers trained on random ECA orbits can predict the next state of unseen local rules with high per-bit accuracy, so rule abstraction from data alone is possible for Boolean functions of fixed arity.","Adding future-state or rule prediction to the loss improves both next-state accuracy and autoregressive rollout quality; the rule-prediction variant generalizes best at look-ahead steps.","Explicitly giving the rule (RO-S) does not fix multi-step planning beyond one or two steps, indicating the difficulty is in propagating intermediate state information, not in knowing the rule.","Autoregressive generation outperforms direct multi-step look-ahead, so models are better at short state-by-state prediction than at jumping ahead without intermediate context.","The number of layers required for accurate O-O prediction grows with planning horizon (2 layers for step 1, 4 for step 2, 7 for step 3, 10 for step 4), suggesting depth is the key architectural resource for sequential computation."],"supporting_citations":[{"why":"Defines the Transformer encoder architecture used as the base model throughout the experiments.","marker":"[8]"},{"why":"CellPyLib generates the ECA orbits for the fixed lattice size and radius used in training and testing.","marker":"[20]"},{"why":"Chain-of-thought prompting supplies the analogy that predicting intermediate steps should improve multi-step planning.","marker":"[36]"},{"why":"Universal Transformers motivate the use of recurrence and adaptive computation time to overcome the depth bottleneck claimed in the conclusions.","marker":"[13]"},{"why":"The Turing-completeness result underpins the premise that Transformers could in principle simulate the sequential rule application required for planning.","marker":"[15]"}],"fun_headline_variants":["Transformers abstract cellular automata rules they've never seen","Depth limits Transformers' multi-step planning in cellular automata","Rule-prediction loss boosts Transformer planning horizons","Orbit training helps Transformers generalize Boolean rules","Deeper Transformers, longer paths in cellular automata"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth-scaling conclusion assumes that the exact layer counts at which accuracy first rises are real signals rather than random variation between training runs; no repeated runs or error bars are reported.","fun_headline_variants_meta":{"raw":{"variants":["Transformers abstract cellular automata rules they've never seen","Depth limits Transformers' multi-step planning in cellular automata","Rule-prediction loss boosts Transformer planning horizons","Orbit training helps Transformers generalize Boolean rules","Deeper Transformers, longer paths in cellular automata"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3089,"prompt_tokens":897,"completion_tokens":2192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2128}},"tokens_in":513,"tokens_out":2192,"duration_ms":16648,"temperature":1.0,"reasoning_tokens":2128,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:22:56.915564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the O-O depth experiment with multiple random seeds and learning-rate schedules at layer counts 2, 4, 7, and 10, and record accuracy distributions. If the apparent thresholds shift by more than a layer or disappear when seeds vary, the claim that each planning step requires additional layers would not be supported; alternatively, a wider 4-layer model that predicts step 3 as well as the 7-layer model would falsify the depth-specific conclusion.","supporting_citations":[{"cited_title":"Attention is All you Need","cited_arxiv_id":null,"evidence_quote":"Defines the Transformer encoder architecture used as the base model throughout the experiments."},{"cited_title":"Chain-of-thought prompting elicits reasoning in large language models","cited_arxiv_id":null,"evidence_quote":"Chain-of-thought prompting supplies the analogy that predicting intermediate steps should improve multi-step planning."},{"cited_title":"Uni- versal transformers","cited_arxiv_id":null,"evidence_quote":"Universal Transformers motivate the use of recurrence and adaptive computation time to overcome the depth bottleneck claimed in the conclusions."},{"cited_title":"Attention is turing-complete","cited_arxiv_id":null,"evidence_quote":"The Turing-completeness result underpins the premise that Transformers could in principle simulate the sequential rule application required for planning."}],"review_version":1}