{"id":"fd99cd84-881f-449a-a622-5f0a12e01df1","arxiv_id":"2501.07359","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Linear probes of Llama activations show an abstraction gradient, but the 70B model deviates with double peaks and anti-persistent adjacent-layer attention.","lead":"A layer-by-layer probing study of Llama models finds that abstraction generally increases with depth, but the larger model shows two distinct peaks for relation and analogy coding and an alternating pattern between neighboring attention layers. The findings are new and relevant for interpretability, though they rest on small datasets and lack statistical validation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The adjacent-layer anti-persistence claim is likely a differencing artifact: Spearman lag-1 autocorrelation of first differences is negative even for independent per-layer noise, so the coordination finding needs a surrogate null test.","rationale":"The reader's verdict of CONDITIONAL is appropriate, but the specific weakness I identify is more pointed than the broad probe-comparability concern. The double-peak pattern is visually consistent and deserves conditional acceptance as an empirical observation. However, the adjacent-layer coordination claim is derived from a statistic whose null behavior is strongly negative: differencing independent noise yields a lag-1 autocorrelation near -0.5, so an observed -0.33 cannot be interpreted without a null model. This does not invalidate the double peaks or the overall mapping of layer-wise abstraction, but it removes one of the paper's most surprising mechanistic conclusions. The reader's weakest assumption about probe accuracy comparability is related, though not identical; my concern is that even if probe accuracy is a faithful measure, the anti-persistence statistic has an inherent mathematical artifact. The proposed surrogate test is cheap and decisive. Until it is run, the coordination claim should be treated as speculative, and the paper should remain a conditional, hypothesis-generating contribution rather than a confirmed finding.","tokens_in":8386,"tokens_out":5053,"duration_ms":55587,"concrete_test":"Construct a surrogate null for each 80-layer accuracy series in Figures 5c-i (and the corresponding series in the 3b model). Fit a smooth trend (e.g., cubic smoothing spline or moving average) to the accuracy sequence; compute residuals. Generate 10,000 surrogate series by adding bootstrap-resampled residuals to the fitted trend, preserving layer count, then recompute the Spearman lag-1 autocorrelation of first differences for each surrogate to form a null distribution. If the observed per-experiment rho values (and their cross-experiment mean) fall within the 95% interval of the null, the anti-persistence statistic is a differencing artifact and the adjacent-layer coordination claim fails; if they fall outside the null tail, the coordination claim survives this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is not the visual double peaks themselves but the quantitative claim about adjacent-layer anti-persistence (Experiment 5, Analyses 2). The authors compute the Spearman autocorrelation of first differences of layer-wise accuracies, obtain mean rho = -0.33 (SD = 0.12), and interpret this as coordination between adjacent attention layers. However, differencing any noisy accuracy curve induces negative lag-1 autocorrelation. If y_l = smooth(l) + noise_l with independent layer-wise noise, then d_l = y_{l+1} - y_l and d_{l+1} = y_{l+2} - y_{l+1} share the term y_{l+1}; the theoretical lag-1 autocorrelation is -0.5 for pure iid noise. The observed -0.33 is therefore exactly the range expected from a smooth trend plus uncorrelated probe noise, and it is not evidence of 'zigzag' specialization. The authors provide no confidence intervals, surrogate baselines, or permutation tests. The cross-experiment consistency (mean rho = 0.31) also does not rule out this artifact because all experiments share the same probe-estimation noise structure. Thus the emergent 'coordination between adjacent layers' half of the central claim is currently unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper probes layer-wise representations in two Llama models (Llama-3.2-3b and Llama-3.3-70b-Instruct) using linear classifiers and regressions on activations from small texts. It reports a broad abstraction hierarchy in the small model, with item-level semantics peaking early, two-item relations deeper, and four-item analogies deeper still. It also reports two deviations in the larger model: a double-peak pattern in the representation of relations and analogies, and an anti-persistent zigzag in attention-layer probe accuracies that the authors interpret as coordination between adjacent layers. The manuscript is written as an extended abstract, with no code or data release and no error bars on the main curves.","tokens_in":8646,"tokens_out":3758,"duration_ms":37675,"significance":"If the results are robust, the paper makes a useful empirical contribution by mapping layer-wise functional specialization in modern open-weight LLMs and by identifying scaling-related deviations from a simple hierarchy. The double-peak result and the proposed 'buried concept' analysis are potentially relevant to interpretability and activation steering. However, the paper's impact depends heavily on the stability of the probe-accuracy curves; the current lack of error bars, control tasks, and a null model for the autocorrelation analysis leaves the central emergent claims under-supported. The manuscript is honest about its speculative interpretations, and several analyses are explicitly labeled inconclusive, which is a positive feature.","major_comments":[{"comment":"The adjacent-layer anti-persistence claim is not supported as stated. For a sequence y_l = s_l + epsilon_l with independent layer-wise noise, the lag-1 autocorrelation of first differences d_l = y_{l+1} - y_l tends to -0.5 in the pure-noise limit; with a smooth trend plus moderate noise, values near the reported mean rho = -0.33 (SD = 0.12) are expected. The paper provides no surrogate null, permutation test, or confidence interval against this differencing baseline. Because coordination between adjacent layers is one of the three headline emergent effects, the analysis needs a null model (e.g., permuting layer labels or comparing to smoothed baselines) before this claim can be accepted.","section":"Experiment 5, Analyses 2"},{"comment":"The double-peak claim is based on visual inspection of single cross-validation runs without error bars, confidence intervals, or significance tests. The paper states that 'all four experiments show two robustly distinct peaks' and locates them at layers 12-16 and 25-33, but it does not quantify the uncertainty in peak locations or amplitudes, does not test against a unimodal alternative, and does not report variability across folds. Bootstrap or permutation-based intervals, plus a statistical comparison of the double-peak model against a single-peak model, are needed to support the emergence claim.","section":"Experiment 5, Analyses 1"},{"comment":"The core assumption that cross-validated probe accuracy measures the degree to which a layer encodes a property is not validated by control tasks. Accuracy can vary across layers because of feature geometry, layer dimensionality, or probe trainability, so the layer-wise accuracy curves may reflect artifacts of the probing setup rather than representational content. The paper should include control probing experiments—for example, probing the same activations for random labels or for a property known to be absent—to demonstrate that the observed peaks and dips are specific to the semantic property of interest.","section":"Methods, all experiments"},{"comment":"The attribution of the observed differences to scaling is weakened by confounds between the two models. Llama-3.2-3b and Llama-3.3-70b-Instruct differ not only in parameter count but also in architecture, training data, and instruction tuning. The paper should acknowledge this limitation explicitly and, if feasible, include a comparison with a similarly-sized base model or a third model size to strengthen the scaling interpretation.","section":"Experiment 5, model comparison"}],"minor_comments":[{"comment":"The text refers to 'Experiments 5 and 6' when discussing buried concepts, but the manuscript contains only Experiments 1 through 5; this should be corrected to 'Experiments 4 and 5'.","section":"Conclusion, first paragraph"},{"comment":"There is a typo in 'thy e attention layers' that should read 'the attention layers'.","section":"Experiment 5, Analyses 2"},{"comment":"Subfigure labels are inconsistently capitalized in the captions and the text (e.g., 'Figure 2C' versus '2c'); please standardize.","section":"Figures 2 and 3"},{"comment":"No code or data release is mentioned; for reproducibility, sharing the probing datasets, the exact list of word pairs and analogies, and the analysis scripts is strongly recommended.","section":"General"},{"comment":"The reference list omits prior work on probing artifacts (e.g., control tasks for linear probes) and on statistical issues in layer-wise autocorrelation; adding such references would help position the claims and address the potential artifact concerns.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the differencing artifact is well-founded and should be addressed with a surrogate null analysis before publication. The paper also needs error bars or confidence intervals on the main accuracy curves. The manuscript is an extended abstract with potentially interesting findings, but the current evidence for the emergent coordination and double-peak claims is not yet sufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a layer-wise probing study of Llama models, and the one thing to know is: the double-peak story for the 70B model is worth a look, but the adjacent-layer 'anti-persistence' claim is almost certainly a differencing artifact, not a finding. If you read it, read Figures 5 and 6 with that in mind.\n\nWhat's actually new: the layer-wise maps for Llama-3.2-3b and 3.3-70b across item semantics, two-item relations, four-item analogies, and the buried-text manipulation. The hierarchical progression in the small model (items early, relations middle, analogies deeper) replicates the BERT-era results but on modern Llama. The double peak in the 70B attention/FFN outputs around layers 12-16 and 25-33, if real, is a genuinely surprising pattern. The buried-text experiments, showing that early-sentence information persists deep into the network after ~100 filler tokens, are a nice addition and directly relevant to steering.\n\nThe paper is transparent: cross-validated linear probes, clear description of datasets and fold structures, and the author explicitly labels the double-peak interpretations as speculative.\n\nThe soft spots are significant. There are no error bars, no significance tests, and no repetitions of the cross-validation; many datasets are tiny (60 word-pairs, 50 analogies), and there are no control probing tasks to show the accuracy curves aren't reflecting layer-wise trainability or dimensionality rather than representation. The author does not release code or data.\n\nThe anti-persistence claim in Analyses 2 is the clearest problem. The Spearman autocorrelation of first differences of noisy accuracy curves is expected to be negative even if layer-wise representations are independent. Pure iid noise gives -0.5; a smooth trend plus independent probe noise gives roughly the observed -0.33. So the 'coordination between adjacent attention layers' is unsupported without a surrogate null. This also undermines the cross-experiment consistency (rho=0.31), which could simply reflect shared probe-estimation noise.\n\nFor the double peaks themselves, the evidence is suggestive but not confirmatory; they could be probe artifacts given the missing controls.\n\nWho gets value: interpretability researchers and people working on activation steering will want to know the double-peak observation, and the buried-text result is a candidate for follow-up. The paper deserves a serious referee, but it needs substantial revision: add error bars/permutation tests, surrogate baselines for differencing, control probes, and ideally release data/code.\n\nMy recommendation: send it to peer review, and make sure the reviewers force the anti-persistence analysis to be redone with a proper null.","headline":"Layer-wise probing study with a probable artifact in the anti-persistence claim; double-peak finding is suggestive but lacks error bars and controls.","tokens_in":9132,"tokens_out":2274,"would_cite":false,"duration_ms":21261,"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":"The paper claims that scaling to a 70-billion-parameter model produces two distinct mid-network peaks for relational and analogical information, plus systematic layer-to-layer alternation in attention outputs, rather than a single smooth…","keywords":["large language models","layer-wise probing","functional hierarchy","semantic abstraction","scaling effects","attention layers","anti-persistence","activation steering"],"falsifier":"Run the same SVM and ridge probes with shuffled labels and repeated random seeds across Llama-3.3-70b layers: if the double peaks and negative first-derivative autocorrelations appear for random labels, or vanish with different seeds, the emergent structure is a probing artifact rather than a property of the model.","tokens_in":8177,"feed_emoji":"📊","tokens_out":8359,"duration_ms":72941,"temperature":0.7,"pith_summary":"The paper tries to establish that the functional organization of large language models is only partly a clean abstraction hierarchy. Using layer-by-layer classifiers on Llama models, it shows that in a smaller 28-layer model semantic properties appear in a sensible order—items, then two-item relations, then four-item analogies—but in a larger 80-layer model the same probes reveal two distinct mid-network peaks for relational and analogical information rather than one smooth gradient. It also reports that adjacent attention layers in the large model alternate systematically in what they encode, a zigzag pattern quantified as anti-persistence. If correct, these findings matter because they suggest qualitative changes in internal organization as models scale, with direct consequences for where embeddings are extracted from and how activations are steered.","feed_headline":"Probing a 70B Llama reveals double semantic peaks and zigzag attention","feed_subtitle":"Probes show a 70B Llama encodes relations twice and alternates attention, complicating extraction and steering.","key_machinery":"The central machinery is linear probing: for each layer, and separately for the residual stream, attention output, and feed-forward output, a support vector machine or ridge regression is trained with cross-validation to predict a semantic label from activations, and the resulting accuracy is treated as a measure of how linearly accessible that information is at that layer. The paper also uses the Spearman autocorrelation of the first differences of these accuracy series across neighboring layers to quantify layer-to-layer alternation, calling a consistently negative value anti-persistence. The probes convert the abstract question “what does each layer represent?” into a measurable curve, and the derivative autocorrelation turns the observed zigzags into a quantitative claim about coordination between adjacent attention layers.","core_discovery":"On its own terms, the paper's central discovery is that scaling a transformer from 28 to 80 layers changes the layer-wise map of semantic abstraction. In Llama-3.2-3b, probe accuracy peaks for item-level semantics in layers 2–7, for two-item relations in layers 8–12, and for four-item analogies in layers 10–15, consistent with a gradual hierarchy. In Llama-3.3-70b-Instruct, the same relational and analogical probes instead show two robustly separate accuracy peaks, around layers 12–16 and layers 25–33, with a valley between them; an overlay shows that the valley coincides with the rise of buried-context representations, whose peak lands near layer 35. The paper further reports that attention-layer outputs in the large model fluctuate from layer to layer in how strongly they represent a probed property, quantified by a negative Spearman autocorrelation of first derivatives (mean $\\rho = -0.33$, standard deviation $0.12$) that is specific to adjacent layers and consistent across experiments. These patterns are presented as emergent effects of scale rather than as contradictions of hierarchy altogether.","pith_inferences":["A testable extension is to perturb or ablate the valley layers (roughly 17–24) in the large model and measure whether the second peak, long-context performance, and relational probes change coherently; this would test the paper's speculation that the valley is where local representations are integrated into global ones.","The anti-persistence claim lends itself to a null-model check: if layer outputs were independent, first-difference autocorrelations would scatter near zero, so a permutation test over experiment order would sharpen the consistency claim.","A practical implication the author leaves implicit: layer choice for feature extraction in downstream tasks should be validated per task and per model scale, because the optimal layer for a semantic property is not monotonic in depth once models get large.","The buried-concept results suggest that deep layers implement content-addressable compression rather than pure abstraction; a direct test would be to vary the filler length and see whether the late-layer probe signature moves smoothly with distance."],"forward_implications":["If the double-peak structure is real, embeddings for relational or analogical content should be taken from at least two depth bands (roughly layers 12–16 and 25–33 in an 80-layer Llama) rather than from a single “semantic” layer.","Activation steering that targets a concept's representation should account for buried contexts: deep layers retain a compressed form of early content, so late-layer steering may nudge behavior without forcing fixation on the concept.","The layer-wise map gives practical guidance for extracting embeddings and for interpretability: different transformer components (residual, attention, feed-forward) carry different amounts of the probed information at different depths.","Because anti-persistence appears in the larger model and is consistent across experiments, scaling may introduce an organized division of labor between adjacent attention layers rather than simply more of the same processing.","The valley between the two peaks may indicate a transition where local relational representations are integrated into more global ones, making mid-network layers a natural target for studying abstraction breaks."],"supporting_citations":[{"why":"Supplies the classic layer-wise hierarchy claim about BERT that the paper revisits for modern LLMs.","marker":"1"},{"why":"Advances the view that BERT rediscovers the classical NLP pipeline, the hierarchy hypothesis being tested.","marker":"2"},{"why":"Provides prior layer-by-layer evidence on which layers of generative LLMs reflect lexical semantics.","marker":"3"},{"why":"Supplies the linear structural probing methodology that the experiments adapt.","marker":"4"},{"why":"Reports an abstraction phase in transformers, a phenomenon the double peaks are compared against.","marker":"5"},{"why":"Documents the earlier rise-fall-rise layer accuracy pattern that the small-model results reproduce.","marker":"6"},{"why":"Provides the activation-steering technique that the layer-wise map is meant to inform.","marker":"7"}],"fun_headline_variants":["Scaling LLMs to 70B yields double abstraction peaks","70B Llama shows twin semantic peaks and zigzag attention","Emergent double peaks in LLM layer-wise semantics","Large models weave hierarchy with double peaks and shifts","Probing reveals 70B LLM's dual abstraction peaks and attention wiggles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a classifier's cross-validated accuracy on a layer's activations measures how strongly that layer encodes the tested property, so the observed peaks and zigzags reflect the model rather than artifacts of how easily different layers can be probed.","fun_headline_variants_meta":{"raw":{"variants":["Scaling LLMs to 70B yields double abstraction peaks","70B Llama shows twin semantic peaks and zigzag attention","Emergent double peaks in LLM layer-wise semantics","Large models weave hierarchy with double peaks and shifts","Probing reveals 70B LLM's dual abstraction peaks and attention wiggles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1582,"prompt_tokens":1111,"completion_tokens":471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":727,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":727,"tokens_out":471,"duration_ms":5030,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:43:15.538287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same SVM and ridge probes with shuffled labels and repeated random seeds across Llama-3.3-70b layers: if the double peaks and negative first-derivative autocorrelations appear for random labels, or vanish with different seeds, the emergent structure is a probing artifact rather than a property of the model.","supporting_citations":[{"cited_title":"& Seddah, D","cited_arxiv_id":null,"evidence_quote":"Supplies the classic layer-wise hierarchy claim about BERT that the paper revisits for modern LLMs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the activation-steering technique that the layer-wise map is meant to inform."}],"review_version":1}