{"id":"5b852ca6-442a-49a6-a777-e5fa78bd9382","arxiv_id":"2402.17762","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.","lead":"Large language models contain a tiny number of neuron activations that are vastly larger than the rest, sometimes 100,000 times bigger. These massive activations stay nearly constant across inputs and act as fixed bias terms that shape attention patterns.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Generalization of input-independent bias role and attention effect from limited models to all LLMs","rationale":"The reader's weakest_assumption directly identifies the same empirical-scope limitation. Because the paper is observational and the strongest_claim is framed universally, broadening the test set is the single check that would either confirm or falsify the load-bearing step; no internal inconsistency or formal gap is visible from the given material.","tokens_in":1622,"tokens_out":317,"duration_ms":20264,"concrete_test":"Re-run the constancy and attention-concentration measurements on Llama-3-70B and Mistral-7B using the paper's exact detection threshold, plus 500 out-of-distribution prompts (code-switched, adversarial, long-context); report per-layer max/min ratio and attention entropy change; if any model shows >5% relative variance in the 'massive' values or loss of attention concentration, the generalization claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that massive activations remain nearly constant across inputs and act as indispensable biases that concentrate attention, but this is asserted after characterization on an unspecified small set of models. The abstract states the phenomenon occurs 'across various LLMs' and values 'largely stay constant regardless of the input,' yet provides no quantitative bound on variance, no enumeration of tested architectures or prompt distributions, and no ablation showing indispensability (e.g., zeroing the activations and measuring downstream degradation). Without those controls, the universality and functional necessity remain unsecured.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript reports an empirical observation of 'massive activations' in large language models: a small number of activations with values orders of magnitude larger than the rest (e.g., 100,000x). These activations are characterized across various LLMs, shown to remain largely constant across inputs, to function as indispensable bias terms, and to induce concentration of attention probabilities onto their corresponding tokens (with resulting implicit biases in self-attention outputs). The same phenomenon is examined in Vision Transformers, and code is released.","tokens_in":1725,"tokens_out":570,"duration_ms":22368,"significance":"If the core empirical claims hold after tighter controls, the work supplies a concrete, reproducible handle on an internal LLM regularity that directly shapes attention behavior. The release of code is a clear strength for follow-up work on model analysis and potential interventions.","major_comments":[{"comment":"Abstract and characterization sections: the claim that massive activations 'function as indispensable bias terms' and 'lead to the concentration of attention probabilities' rests on observational correlations but provides no ablation (e.g., zeroing the identified activations and measuring downstream perplexity or task degradation) or quantitative bound on input variance; without these the indispensability and causal attention effect remain unsecured.","section":"Abstract"},{"comment":"Results on LLMs: the statement that the phenomenon occurs 'across various LLMs' and values 'largely stay constant regardless of the input' lacks an enumerated list of architectures, prompt distributions, or statistical summary (mean/variance of activation magnitude across inputs); the absence of these controls makes the universality claim difficult to evaluate.","section":"Characterization of massive activations"},{"comment":"Attention analysis: the mechanism linking massive activations to attention concentration and implicit bias terms is described qualitatively but lacks explicit equations or controlled before/after measurements showing how the large constant values alter the softmax distribution relative to a baseline without them.","section":"Attention concentration"}],"minor_comments":[{"comment":"Notation for activation magnitude thresholds and 'massive' criteria should be defined explicitly (e.g., a precise multiple or percentile) rather than relying on the example '100,000 times larger'.","section":"Introduction"},{"comment":"Figure legends and captions would benefit from stating the exact models, layers, and input types shown so readers can assess representativeness without cross-referencing text.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is primarily an empirical observation note rather than a theoretical or algorithmic contribution; this may affect fit for venues that prioritize novel methods over descriptive findings, though the reproducibility via code mitigates some concerns."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below and have revised the manuscript to incorporate additional experiments, documentation, and quantitative analyses as requested.","responses":[{"response":"We agree that explicit causal evidence strengthens the claims. In the revised manuscript we add ablation experiments that zero the identified massive activations and report the resulting perplexity increase on held-out validation sets together with performance drops on downstream tasks. We also supply quantitative bounds on input variance, showing that the standard deviation of massive-activation magnitudes across 10,000 diverse prompts is orders of magnitude smaller than the mean value.","revision_made":"yes","referee_comment":"[Abstract] Abstract and characterization sections: the claim that massive activations 'function as indispensable bias terms' and 'lead to the concentration of attention probabilities' rests on observational correlations but provides no ablation (e.g., zeroing the identified activations and measuring downstream perplexity or task degradation) or quantitative bound on input variance; without these the indispensability and causal attention effect remain unsecured."},{"response":"We accept that greater specificity is needed. The revision includes a dedicated table that enumerates every architecture examined (Llama-2 7B/13B, Mistral-7B, Gemma-7B, and additional models), the exact prompt distributions (C4, The Pile, and synthetic random sequences), and statistical summaries (mean, variance, and range) of activation magnitudes computed over 10,000 inputs.","revision_made":"yes","referee_comment":"[Characterization of massive activations] Results on LLMs: the statement that the phenomenon occurs 'across various LLMs' and values 'largely stay constant regardless of the input' lacks an enumerated list of architectures, prompt distributions, or statistical summary (mean/variance of activation magnitude across inputs); the absence of these controls makes the universality claim difficult to evaluate."},{"response":"We have expanded the attention section with explicit equations that show how a large constant added to the pre-softmax logits produces the observed probability concentration. We further include controlled before/after measurements that subtract the mean massive-activation value from the attention scores and quantify the resulting change in attention entropy and output bias.","revision_made":"yes","referee_comment":"[Attention concentration] Attention analysis: the mechanism linking massive activations to attention concentration and implicit bias terms is described qualitatively but lacks explicit equations or controlled before/after measurements showing how the large constant values alter the softmax distribution relative to a baseline without them."}],"tokens_in":1290,"tokens_out":546,"duration_ms":36244,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core observation is straightforward: in several LLMs a handful of activation values sit 100,000 times above the rest, stay nearly constant across inputs, and appear to function as bias terms that concentrate attention on their tokens. The authors map where these occur, show the pattern in both language and vision transformers, and supply code, which is the right way to present an empirical claim like this. That combination of measurement plus reproducibility is what makes the work usable right away for people doing interpretability or pruning experiments. The attention-concentration link follows directly from the numbers they report and does not rely on extra assumptions. The main limitation is scope. The abstract says the effect appears across various models but gives no table of exact architectures, prompt distributions, or variance bounds, so the claim that the values are largely constant and indispensable rests on whatever set they actually ran. An ablation that zeros the massive activations and measures downstream change would tighten the functional story, but it is not required for the basic finding to stand. This is the sort of targeted empirical note that fits a methods or analysis track rather than a flagship result. I would bring it to a reading group to check the numbers on a couple of models we have locally, and I would cite the observation if the constancy holds under our own prompts. It is worth sending to referees because the measurement is clean, the code is public, and the phenomenon is new enough that others should be able to build on or refute it quickly.","headline":"The paper spots a small set of input-independent 'massive' activations that act as fixed biases and skew attention in LLMs, backed by measurements and released code.","tokens_in":2171,"tokens_out":372,"would_cite":true,"duration_ms":13653,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"echoes","rs_module":"Cost.JcostCore","rs_theorem":"Jcost_unit0","paper_passage":"their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs"},{"relation":"unclear","rs_module":"Foundation.LedgerForcing","rs_theorem":null,"paper_passage":"these massive activations lead to the concentration of attention probabilities to their corresponding tokens"}],"headline":"Massive activations as input-independent biases in LLMs show no direct connection to RS J-cost fixed points or ledger structure","alignment":"orthogonal","rationale":"The paper empirically identifies constant large activations acting as biases that concentrate attention, which superficially echoes the fixed point of J at 1 (defect=0) and bias-like behavior in RS cost minima. However, it provides no engagement with the RS forcing chain (distinction to J uniqueness to φ to D=3 to ledger), treating the phenomenon as an LLM-specific observation rather than a derived consequence of zero-parameter recognition dynamics. No RS theorems are invoked or paralleled.","tokens_in":281045,"confidence":"low","tokens_out":270,"duration_ms":30348,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"lean_confirmation":{"model":"grok-4.3","status":"out_of_scope","citations":[],"rationale":"This is an empirical ML paper whose load-bearing premise is observational (activation magnitudes, constancy, bias role). Shape-of-logic is a formalization of physical reality from one distinction and contains no relevant theorems. Status is out_of_scope per guidelines.","tokens_in":280822,"confidence":"moderate","tokens_out":182,"duration_ms":31452,"inferential_bridge":"The paper's central claims rest on empirical observations and measurements across a limited set of LLMs (e.g., LLaMA2, Mixtral) and inputs; no mathematical/structural identity is asserted that could be machine-checked. Shape-of-logic contains no theorems about neural activations, attention, or LLM internals.","load_bearing_premise":"Massive activations (few activations with values ~100,000x larger than others) have constant values independent of input and function as indispensable bias terms, causing attention concentration to their tokens.","cache_read_input_tokens":64,"cache_creation_input_tokens":0},"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Large language models contain a small number of massive activations that remain constant across inputs and act as indispensable bias terms.","keywords":["massive activations","large language models","transformer","attention mechanism","bias terms","self-attention","vision transformers"],"falsifier":"Measuring the largest activations on two very different inputs inside the same layer of a new LLM and finding that their relative magnitudes or absolute values change by more than a small constant factor.","tokens_in":2535,"feed_emoji":"","tokens_out":652,"duration_ms":19462,"temperature":0.7,"pith_summary":"The paper reports that LLMs consistently produce a handful of activations whose values are orders of magnitude larger than all others. These massive activations change very little when the input changes and therefore function as fixed additive biases inside the network. Because they are so large they dominate the attention scores, causing probability mass to concentrate on the tokens that produce them. The same pattern appears in both language and vision transformers. Characterizing this mechanism clarifies why certain tokens receive outsized influence in every forward pass.","feed_headline":"Massive constant activations act as fixed biases in LLMs","feed_subtitle":"A few input-invariant high-magnitude values concentrate attention on their tokens and shape every output.","key_machinery":"Massive activations: the small set of high-magnitude, nearly input-invariant activation values that serve as fixed bias terms and drive attention concentration.","core_discovery":"We observe an empirical phenomenon in Large Language Models (LLMs) -- very few activations exhibit significantly larger values than others (e.g., 100,000 times larger). We call them massive activations. First, we demonstrate the widespread existence of massive activations across various LLMs and characterize their locations. Second, we find their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs. Third, these massive activations lead to the concentration of attention probabilities to their corresponding tokens, and further, implicit bias terms in the self-attention output.","pith_inferences":["Interpreting LLMs may become simpler by isolating these few constant terms rather than analyzing every activation.","Model compression or editing techniques could treat the massive activations as a separate, editable bias vector.","The same mechanism may appear in other sequence models, offering a route to test architectural universality.","Training procedures that explicitly regularize or initialize these large constant values could change convergence behavior."],"forward_implications":["Attention probability mass concentrates on the tokens that produce the massive activations.","Self-attention outputs contain implicit bias terms traceable to these constant activations.","The pattern extends to Vision Transformers, suggesting a general transformer property.","Because the activations act as indispensable biases, altering or removing them would change model output distributions.","Model scaling laws and internal dynamics must account for these persistent high-magnitude terms."],"fun_headline_variants":["Massive activations serve as constant biases in LLMs","LLMs depend on input-invariant massive activations","Constant high-value activations bias transformer attention","Massive activations concentrate attention on key tokens"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The observed constancy of the largest activation values and their bias-like behavior holds for every LLM architecture and every input distribution.","fun_headline_variants_meta":{"raw":{"variants":["Massive activations serve as constant biases in LLMs","LLMs depend on input-invariant massive activations","Constant high-value activations bias transformer attention","Massive activations concentrate attention on key tokens"]},"model":"grok-4.3","cost_usd":0.004914,"raw_usage":{"total_tokens":2371,"prompt_tokens":596,"num_sources_used":0,"completion_tokens":54,"cost_in_usd_ticks":49137000,"prompt_tokens_details":{"text_tokens":596,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1721,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":596,"tokens_out":54,"duration_ms":17352,"temperature":1.0,"reasoning_tokens":1721,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-16T06:59:19.393219+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Measuring the largest activations on two very different inputs inside the same layer of a new LLM and finding that their relative magnitudes or absolute values change by more than a small constant factor.","supporting_citations":[],"review_version":1}