{"id":"0aa010b7-e427-48dd-9739-2d98496e8b3f","arxiv_id":"2607.26059","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Frozen random CNN feature extractors in PPO yield sparse readouts whose active-neuron count is claimed to track task complexity, but the flagship deterministic-Pong numbers are contradicted by the paper's own appendices.","lead":"A single-author study reports that PPO agents with frozen random CNN backbones compress Atari tasks into extremely few active readout neurons—1–11 for Pong, 19–26 for Breakout, ~42 for Space Invaders—and that ablating those neurons collapses performance. The paper pitches this as a diagnostic of intrinsic task complexity, but several headline numbers conflict with the paper's own tables and depend on hand-set thresholds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1–3 active-neuron claim rests on a single uncalibrated activation threshold (0.01); without a threshold sweep, the headline sparsity numbers and scaling trend could be artifacts of the cutoff.","rationale":"The reader's verdict is CONDITIONAL and identifies the cross-game state-variable estimates as the weakest assumption. I agree that those estimates are fragile, but I find a more foundational operationalization issue in the active-neuron threshold. The abstract's central '1–3 neurons' and the scaling relationship depend on mean activation > 0.01, a single uncalibrated cutoff. Without a threshold sweep, the specific numbers and the 'sparsity scales with task complexity' trend lack demonstrated robustness. This is a concrete, addressable concern: it requires re-analysis of existing activation logs rather than new experiments. It does not overturn the reader's conditional verdict—the paper already needs revision and better data sharing—but it sharpens the condition under which the central claim should be accepted. I therefore recommend keeping the verdict at CONDITIONAL, unchanged from the reader's assessment, while adding the threshold-sweep requirement as a condition. The concern is not an accusation of error; it is a request for evidence that the headline magnitude is not an artifact of measurement choice. The absence of code/data makes it impossible to check independently, so the test is specified as a release-and-recompute step for the authors.","tokens_in":23606,"tokens_out":7947,"duration_ms":76230,"concrete_test":"Have the authors release the per-neuron mean-activation distributions for all seeds and recompute Table 1, Table 2, and Figure 2 under thresholds 0.001, 0.003, 0.01, 0.03, and 0.1. If deterministic Pong's active count changes from 1–3 to more than 10 at 0.001, or if the cross-game rank ordering changes at any threshold, then the headline sparsity and scaling claims require revision. The test should also re-run the FC1-Keep/FC1-Remove ablations at each threshold to see whether the identified 'active set' is stable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantities—FC1 active counts of 1–3, 5–11, 19–26, 42—are all defined by an arbitrary threshold: a neuron counts as active only if its mean post-ReLU activation over 1,000 sampled evaluation frames exceeds 0.01 (Section 3.4). The paper reports no distribution of per-neuron mean activations and no threshold sensitivity analysis. We cannot tell whether this cutoff selects a natural sparse subset or simply slices off a tail of a smooth distribution. If many neurons sit just below 0.01, lowering the threshold to 0.001 could materially change the reported counts and the cross-game rank ordering in Table 1 and Figure 2, making the 'emergence of sparsity' claim partly an artifact of the ruler. The paper's own FC1-Keep ablation shows that sub-threshold neurons are not necessary for performance, which partially mitigates the threat to the existence of a causal bottleneck, but it does not rescue the quantitative active-count values used to support the complexity-scaling claim. This concern is independent of the reader's flagged x-axis issue: even if the state-variable estimates are accepted, the y-axis (active count) must be shown to be robust to the threshold choice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports that deep RL agents trained with a frozen, randomly initialized CNN feature extractor spontaneously learn sparse fully-connected representations without any explicit sparsity objective. It claims that in the first fully-connected layer (FC1, 3,136->64), Pong agents use as few as 1-3 of 64 neurons, and that the number of active FC1 neurons scales with estimated task complexity across Pong, Freeway, Breakout, and Space Invaders. Ablations show that removing the active neurons collapses performance to random play, and a longitudinal sweep indicates that the active set locks early in training, well before reward becomes positive. The paper also contrasts frozen and trainable CNNs in Breakout using active count and participation ratio.","tokens_in":1677,"tokens_out":1645,"duration_ms":45240,"significance":"The phenomenon is interesting and, if confirmed, would connect reservoir computing, random feature methods, and implicit sparsity in RL in a way that is potentially useful for diagnosing task complexity and designing efficient readouts. The ablation protocol is a genuine strength: demonstrating necessity of the active set with a clean intervention across two PPO implementations and multiple games is the kind of evidence that separates a real effect from a threshold artifact. The paper is also transparent about small sample sizes and the lack of a proof. However, the central quantitative claims rest on a single activation threshold and hand-assigned complexity estimates, and there is an internal inconsistency between the headline numbers and the full results table. These issues materially affect the strength of the conclusions as currently stated.","major_comments":[{"comment":"The headline number for deterministic Pong is internally inconsistent. The abstract and Table 1 state 1-3 active FC1 neurons, but Appendix C, Table 7 lists five SB3 frozen deterministic seeds with active counts 3, 4, 6, 8, 8 (range 3-8); the 1-2 values come only from SF seeds. Please reconcile the reported ranges and either present the full distribution or clearly separate SB3 and SF results. Because the sparsity claim and the cross-game scaling in Figure 2 depend on these counts, this discrepancy needs a precise correction.","section":"Abstract/Table 1 vs. Appendix C, Table 7"},{"comment":"All quantitative sparsity results depend on the threshold that classifies a neuron as active if its mean post-ReLU activation over 1,000 frames exceeds 0.01. The paper provides no histogram of per-neuron mean activations and no sensitivity analysis. If many neurons have activations just below 0.01, lowering the threshold to, say, 0.001 could materially increase the reported counts and potentially alter the ordering in Table 1 and Figure 2. The FC1-Keep ablation shows that sub-threshold neurons are not needed for performance, which supports a causal bottleneck, but it does not establish the quantitative active-count values used for the complexity-scaling claim. Please add a threshold sweep (e.g., 0.001, 0.005, 0.01, 0.05, 0.1) and report whether the headline ranges and cross-game monotonicity are stable.","section":"Section 3.4 (active-neuron definition), Fig. 2/Table 1"},{"comment":"The x-axis of Figure 2 is based on hand-set estimates of 'policy-relevant state variables' (Pong ~5, Freeway ~10-15, Breakout ~15-25, Space Invaders ~40+). The paper notes that substituting AtariARI labeled counts would change the slope but preserve rank order, but no such alternative is plotted or quantified. Because the claim that active count 'tracks task complexity' is central, please provide a robustness check with at least one independent complexity measure (e.g., AtariARI counts, a learned state-dimension estimator, or a simple proxy like the number of moving objects). If the monotonic relation only holds under one particular ordinal assignment, the claim needs to be weakened.","section":"Section 3.3, Table 16 (task-complexity estimates)"},{"comment":"The phrase 'emergent sparsity without any sparsity-inducing objective' is strong. ReLU itself naturally produces exact zeros, and the 0.01 threshold treats near-zero mean activations as inactive. The paper's own Appendix N.6 shows that an L1-regularized linear probe needs ~1,000 features, while the ReLU network uses ~3, so the nonlinearity is doing important work. The ablation evidence supports the causal claim, but the 'spontaneous emergence' framing should be tempered unless the threshold-robustness analysis shows that sparsity is not merely a tail of the ReLU output distribution. I recommend explicitly stating this caveat in the main text.","section":"Section 6.1 / Appendix M (mechanism and \"emergent sparsity\" wording)"}],"minor_comments":[{"comment":"The 'Frozen trend' line is described as a qualitative visual guide, not a fitted regression. The word 'trend' may mislead readers into thinking there is a fitted quantitative relationship. Consider labeling it 'illustrative' or removing it from the figure.","section":"Figure 2"},{"comment":"The 'Final PR' column for c10 is listed as 13.31 in Table 4, while Table 2 reports 13.73 for the same seed. If these come from different sweeps or checkpoint timings, please clarify; otherwise unify the values.","section":"Table 4"},{"comment":"The within-game scaling claim is appropriately hedged as 'suggestive' with n=3 in the main text, but the abstract states 'three identical Pong seeds produce 5, 7, and 11 active neurons' without noting the wide error bars or the binary nature of the reward split (+14 vs. +18). A short caveat in the abstract would improve precision.","section":"Abstract / Section 4.3"},{"comment":"The text says activation statistics are computed from 1,000 frames sampled during evaluation, but elsewhere it says 50-100 evaluation episodes. Please clarify how many episodes and how many frames per episode are used for the activation sample, since this affects the reliability of the 0.01 threshold.","section":"Section 3.4 / Appendix N.4"}],"recommendation":"major_revision","confidential_remarks":"The ablation methodology and breadth of experiments are genuinely commendable, and the core phenomenon is likely real in some form. My main concern is that the paper's quantitative headline numbers are not yet robust: the deterministic-Pong discrepancy between abstract/Table 1 and Appendix C, and the absence of a threshold sensitivity analysis, make the exact sparsity magnitudes and the scaling law fragile. The authors should be asked to reconcile the numbers and provide the threshold sweep and alternative-complexity robustness checks before the paper can be recommended for acceptance. If those checks overturn the scaling trend, the paper would still be a valid empirical observation, but its central claim would need to be substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the core phenomenon is likely real, and the paper is worth taking seriously, but the abstract and Table 1 overclaim in ways the paper's own appendix contradicts. The best evidence is the ablation logic: FC1-Remove crashes Pong and Breakout to random play, FC1-Keep preserves performance, and the effect replicates across SB3 and Sample Factory for Pong. The 47-checkpoint longitudinal sweep showing the active set locks before reward emerges is also nice and reported honestly with n=3 caveats.\n\nWhat's new: no one, as far as I know, has measured readout activation sparsity this way in frozen random CNN RL, and the width-scaling result (absolute active count stable across widths) does push back against a trivial capacity-fraction explanation. The Freeway boundary case, where ablation shows vestigial neurons and traps trainable seeds too, is a good-faith complication and they handle it correctly.\n\nSoft spots, in rough order:\n\n1. The 1–3 deterministic Pong headline doesn't match their own Table 7 (SB3 deterministic seeds are 3–8; the 1–2s come from a different SF setting). The abstract and Table 1 need to be fixed before this goes anywhere.\n\n2. The active-count y-axis rests on a single 0.01 mean-activation threshold, with no threshold sweep. FC1-Keep tells us those neurons are sufficient, which partially rescues the bottleneck claim, but the scaling trend across games could shift if the threshold moves. This needs a sensitivity analysis.\n\n3. The x-axis state-variable estimates are hand-assigned. They're honest about it being ordinal and say AtariARI substitution preserves rank order, so this is a minor-to-moderate issue, not fatal.\n\n4. \"Competitive rewards\" in Breakout is only true for deterministic-trained frozen agents evaluated in-distribution. The progressive-sticky frozen agents get 50–65 vs the trainable 298 at the same sticky=0.25 evaluation. That sentence in the abstract overreaches.\n\n5. The \"four games\" ablation claim lacks visible support for Space Invaders in the manuscript I read. Freeway ablations are mixed by design. So \"crashes across four games\" is not supported as stated.\n\nNo code/data either, which for a measurement-heavy empirical paper is a real gap.\n\nWho benefits: people working on RL interpretability, reservoir computing, or efficient readouts. With corrected numbers and a threshold sweep, this could be a solid empirical contribution. I'd send it to review but expect revision.","headline":"The frozen-random-CNN sparsity phenomenon is probably real, but the headline numbers overstate it; worth refereeing after the authors fix the abstracts and add a threshold sensitivity analysis.","tokens_in":24377,"tokens_out":3423,"would_cite":false,"duration_ms":31182,"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":"A reinforcement learning agent with a CNN frozen at random initialization solves Pong through 1–3 of 64 readout neurons, with sparsity scaling to task complexity.","keywords":["emergent sparsity","frozen random features","deep reinforcement learning","information bottleneck","active neuron count","participation ratio","task complexity","random projection"],"falsifier":"Re-run the active-neuron analysis with thresholds ranging from 0.001 to 0.1; if Pong's 1–3 'active' neurons balloon to 50–64 under a slightly looser threshold, the sparsity is a measurement artifact. Alternatively, find a same-recipe frozen Pong seed that reaches expert reward with zero active FC1 neurons, which would mean the ablation signature is not necessary for competence.","tokens_in":23405,"feed_emoji":"🕹️","tokens_out":6820,"duration_ms":54896,"temperature":0.7,"pith_summary":"This paper claims that when a deep reinforcement learning agent's convolutional feature extractor is frozen at random initialization, gradient descent on the trainable readout spontaneously discovers extremely sparse representations: deterministic Pong is solved through only 1–3 of 64 neurons in the first fully-connected layer, with 5–11 under sticky-action stochasticity. The sparsity is not engineered—no pruning, regularization, or sparse-inducing objective—and the number of active neurons grows with the task's estimated complexity across four Atari games. Ablations show the active set is both necessary and sufficient: removing it crashes reward to random play across two independent policy-gradient implementations, while keeping only it preserves performance. Longitudinal checkpoints show the active set locks in by 15–30M steps while reward turns positive 35–105M steps later, ruling out the idea that neurons died because the policy converged. A sympathetic reader would care because, if right, a frozen random projection plus ordinary gradient descent exposes the effective dimensionality of a task that end-to-end training obscures.","feed_headline":"Frozen random CNN solves Pong with 1–3 active neurons","feed_subtitle":"With CNN features fixed at random, Pong needs just 1–3 of 64 readout neurons; harder games recruit more.","key_machinery":"The load-bearing object is a frozen, randomly initialized convolutional stack—three conv layers producing 3,136 features—followed by trainable fully-connected layers (FC1: 3,136→64; FC2: 64→64) and policy/value heads; formally, this is a reservoir/random-feature architecture. The central observable is 'active-neuron count': FC1 neurons whose mean post-ReLU activation exceeds 0.01, complemented by PCA dimensionality and the participation ratio of FC2-output activations. Active-count measurement carries the task-complexity claim; the ablation protocol (zeroing active versus inactive neurons) carries the causal-necessity claim; and the longitudinal checkpoint sweep carries the temporal-commitme","core_discovery":"The central claim is that a fixed random projection followed by gradient descent acts as a compressive measurement device: the number of FC1 neurons a proximal-policy-optimization agent actually uses is governed by the task's intrinsic dimensionality, not by network width or a fixed fraction of capacity. Concretely, frozen-CNN agents activate 1–3 neurons for deterministic Pong, 19–26 for Breakout, and roughly 42 for Space Invaders, while trainable-CNN controls activate 55–64 under matched conditions. Within a single game at matched hyperparameters, three same-recipe Pong seeds produce 5, 7, and 11 active neurons, and the 5-neuron seed plateaus below expert reward while the others reach it—co","pith_inferences":["Beyond the paper: the same frozen-projection-plus-emergent-sparse-readout recipe could serve as a diagnostic for intrinsic dimensionality in any high-dimensional prediction setting with few underlying factors—genomics, imaging, or sensor arrays—where the paper offers this as conjecture but does not test it.","Beyond the paper: if early active-set lock generalizes, a direct testable extension is to rewire only the trainable head after lock (e.g., replace FC2 with a wider MLP) and ask whether reward ceiling lifts; the paper's capacity-lag pattern predicts it would not for a fixed projection.","Beyond the paper: the 0.01 activation threshold is a measurement choice; a threshold sweep (0.001 to 0.1) would reveal whether the 1–3-neuron count is a stable plateau or a knife-edge, which the current manuscript does not report."],"forward_implications":["Frozen-random-feature agents can be used as a no-penalty probe: the number of active readout neurons estimates how many independent state variables a task's policy actually uses.","Width-scaling results imply that increasing FC width does not change the absolute number of neurons recruited, so sparse bottlenecks reflect task structure rather than a fixed fraction of capacity.","Representational compression precedes competence: because the active set locks long before reward improves, downstream learning works within an immutable feature subspace, suggesting training interventions should focus on the readout rather than the features.","Frozen and trainable CNNs can reach comparable reward with structurally different bottlenecks—narrow-but-diverse versus broad-but-redundant—so active count alone is not a health metric; participation ratio is a complementary indicator.","Deployment is cheap: only the trained readout (about 9,400 parameters for a 3-neuron Pong solution) needs to be stored, since the random seed regenerates the frozen projection."],"fun_headline_variants":["Pong solved with 1-3 neurons via frozen random CNN","Frozen random CNNs compress Pong to 1-3 active neurons","Emergent sparsity: random CNN extractors solve Pong with 1-3 neurons","Task complexity dictates sparse neuron count in frozen CNNs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The cross-game scaling claim rests on hand-assigned estimates of how many state variables each game uses (Pong ~5, Breakout ~15–25, Space Invaders ~40+); if those ordinal complexity numbers are wrong, the claim that active-neuron count tracks task complexity loses its quantitative grounding.","fun_headline_variants_meta":{"raw":{"variants":["Pong solved with 1-3 neurons via frozen random CNN","Frozen random CNNs compress Pong to 1-3 active neurons","Emergent sparsity: random CNN extractors solve Pong with 1-3 neurons","Task complexity dictates sparse neuron count in frozen CNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3007,"prompt_tokens":905,"completion_tokens":2102,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":2035}},"tokens_in":649,"tokens_out":2102,"duration_ms":14594,"temperature":1.0,"reasoning_tokens":2035,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:54:46.968582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the active-neuron analysis with thresholds ranging from 0.001 to 0.1; if Pong's 1–3 'active' neurons balloon to 50–64 under a slightly looser threshold, the sparsity is a measurement artifact. Alternatively, find a same-recipe frozen Pong seed that reaches expert reward with zero active FC1 neurons, which would mean the ablation signature is not necessary for competence.","supporting_citations":[],"review_version":1}