{"id":"d11ae82f-2c7c-42ed-9621-fefe4a8cd304","arxiv_id":"2605.30524","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Sequential post-training of LLMs induces representation collapse that correlates with reduced plasticity, weaker generalization, and poorer calibration, with lightweight interventions tested to mitigate it.","lead":"This paper studies whether chaining multiple post-training stages on large language models causes their internal representations to collapse into low-rank and homogeneous spaces. A smart generalist might read it to learn about potential limits on how adaptable LLMs remain after repeated fine-tuning and safety tuning.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Measurement suite may fail to isolate causal role of collapse vs. confounding stage effects","rationale":"The reader's weakest assumption directly identifies the measurement suite and causal demonstration as the load-bearing point; the concern above is a precise technical elaboration of that same gap rather than a new objection.","tokens_in":1636,"tokens_out":276,"duration_ms":14039,"concrete_test":"Add a yoked control condition that applies the same total data volume and stage count but randomizes ordering while measuring the same collapse and plasticity metrics; if the correlation between collapse scores and plasticity loss disappears or reverses while behavioral gains remain, the causal isolation claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the defined suite (hidden-state anisotropy, logit concentration, token trajectories, LoRA update norms) both quantifies collapse and establishes its causal effect on later plasticity, OOD generalization, and calibration under controlled orderings. Because sequential stages also alter cumulative data exposure, optimization trajectory, and effective capacity, any observed correlation between collapse metrics and downstream metrics could be driven by those confounders rather than the geometric concentration itself. The interventions (replay, feature refresh, diversity regularization) are evaluated, but without an explicit test that decouples the geometric signature from the other changes, the predictive link remains unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper examines whether sequential post-training stages in LLMs induce representation collapse, characterized by low-rank, anisotropic, and homogeneous feature spaces. It introduces a measurement suite spanning hidden-state anisotropy, logit concentration, token trajectories, and LoRA update norms, applies it across controlled orderings of SFT, preference optimization, safety tuning, math/code specialization, and long CoT, and hypothesizes that such collapse causally reduces later-stage plasticity, OOD generalization, and calibration. Lightweight interventions (mixed-domain replay, feature refresh, diversity regularization, LoRA decorrelation) are evaluated as mitigations that preserve behavioral gains.","tokens_in":1761,"tokens_out":501,"duration_ms":15038,"significance":"If the measurement suite successfully isolates a causal geometric mechanism rather than stage-order confounders, the work would supply a concrete, testable account of why multi-stage post-training often degrades adaptability and would directly inform practical recipe design for preserving future learnability.","major_comments":[{"comment":"Measurement suite section: the claim that the defined metrics (hidden-state anisotropy, logit concentration, token trajectories, LoRA norms) establish a causal link between representation collapse and reduced plasticity/OOD performance requires an explicit decoupling experiment. Because stage ordering simultaneously changes cumulative data exposure, optimization trajectory, and effective capacity, any observed correlation could be driven by those factors; the manuscript must show that the geometric signature predicts the downstream metrics even after controlling for the confounders.","section":"Measurement suite section"},{"comment":"Intervention evaluation section: the reported gains from replay, feature refresh, and diversity regularization must be accompanied by controls that verify the interventions act through the collapse metrics rather than through other mechanisms (e.g., simply increasing effective data diversity). Without such mediation analysis or matched ablations, it remains unclear whether the interventions succeed by mitigating the hypothesized geometric cause.","section":"Intervention evaluation section"}],"minor_comments":[{"comment":"Notation for the anisotropy and concentration metrics should be defined with explicit formulas and normalization details in the main text rather than deferred to an appendix.","section":null},{"comment":"The abstract states the central hypothesis but supplies no quantitative outcomes; the introduction or results section should include a concise summary table of key effect sizes for the collapse–plasticity relationship.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive report. The two major comments correctly identify gaps in establishing causality for the collapse-plasticity link and in validating the mechanism of the proposed interventions. We address each point below and commit to revisions that strengthen these aspects without altering the core claims or experimental scope.","responses":[{"response":"We acknowledge that controlled stage orderings alone do not fully isolate the geometric signature from confounders such as cumulative data exposure and optimization trajectory. In the revision we will add a decoupling experiment that holds total tokens and optimization steps fixed while varying only the presence of collapse-inducing stages (via replay buffers that restore diversity without changing data volume). We will then report partial correlations and regression coefficients showing that collapse metrics remain predictive of plasticity and OOD metrics after these controls.","revision_made":"yes","referee_comment":"[Measurement suite section] Measurement suite section: the claim that the defined metrics (hidden-state anisotropy, logit concentration, token trajectories, LoRA norms) establish a causal link between representation collapse and reduced plasticity/OOD performance requires an explicit decoupling experiment. Because stage ordering simultaneously changes cumulative data exposure, optimization trajectory, and effective capacity, any observed correlation could be driven by those factors; the manuscript must show that the geometric signature predicts the downstream metrics even after controlling for the confounders."},{"response":"The referee is right that the current intervention results lack explicit mediation or matched ablations. In the revision we will include (i) a mediation analysis regressing downstream gains on both intervention type and measured change in collapse metrics, and (ii) matched ablations that increase data diversity through non-geometric means (e.g., random token shuffling) and show they do not produce the same plasticity or calibration benefits. These additions will be reported alongside the existing tables.","revision_made":"yes","referee_comment":"[Intervention evaluation section] Intervention evaluation section: the reported gains from replay, feature refresh, and diversity regularization must be accompanied by controls that verify the interventions act through the collapse metrics rather than through other mechanisms (e.g., simply increasing effective data diversity). Without such mediation analysis or matched ablations, it remains unclear whether the interventions succeed by mitigating the hypothesized geometric cause."}],"tokens_in":1323,"tokens_out":475,"duration_ms":21211,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this work tracks how sequential post-training compresses representations and checks a handful of fixes, yet the link from collapse to weaker later adaptation rests on correlations that other training changes could explain.\n\nWhat is new is the controlled comparison of stage orderings (SFT then preference tuning, safety, math specialization, long CoT) using metrics on hidden-state anisotropy, logit concentration, token trajectories, and LoRA update norms. The interventions—mixed replay, feature refresh, diversity regularization, and decorrelation—are straightforward and directly tied to the measurements.\n\nThe paper does a solid job making the measurement suite explicit enough that others could apply it. Framing collapse as something that matters for future learnability rather than just a geometric side effect is a useful shift.\n\nThe soft spot is exactly the one in the stress-test note. Sequential stages also change cumulative data, optimization path, and effective capacity, so any observed drop in plasticity or calibration could trace to those factors instead of the concentration itself. The interventions are run, but without an ablation that varies the geometric signature while holding the other variables fixed, the predictive claim stays unproven. The abstract gives no effect sizes or controls, which makes it hard to gauge how much of the downstream signal is real.\n\nThis is for people who design or debug multi-stage post-training pipelines. A reader who cares about representation dynamics during adaptation would get concrete ideas from the metrics. It deserves a serious referee because the question is current and the measurements are reproducible, even if the causality needs tightening.\n\nRecommendation: send to review, but ask for explicit tests that separate collapse from the other stage effects.","headline":"The paper measures representation collapse across post-training stages with a new suite and tests interventions, but the claimed causal effects on plasticity look vulnerable to stage confounders.","tokens_in":2272,"tokens_out":409,"would_cite":false,"duration_ms":18817,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Sequential post-training compresses LLM internal representations into low-rank spaces that limit later plasticity, generalization, and calibration.","keywords":["representation collapse","sequential post-training","large language models","plasticity","out-of-domain generalization","calibration","LoRA updates","measurement suite"],"falsifier":"Running the same later adaptation stage on two models that differ only in measured representation concentration but show identical plasticity, out-of-domain accuracy, and calibration scores would falsify the claimed predictive link.","tokens_in":2544,"feed_emoji":"📉","tokens_out":713,"duration_ms":18129,"temperature":0.7,"pith_summary":"The paper examines whether chains of post-training stages on large language models gradually compress hidden representations into low-rank, anisotropic, and homogeneous spaces. It introduces a measurement suite covering hidden states, logits, token trajectories, and LoRA updates, applied across supervised fine-tuning, preference optimization, safety tuning, and specialization under controlled stage orders. If the central hypothesis holds, this concentration directly reduces how flexibly the model can adapt in subsequent stages, how well it handles out-of-domain inputs, and how accurately its output probabilities reflect true uncertainty. The authors also test simple countermeasures such as mixed-domain replay and diversity regularization to retain behavioral gains while keeping representations more open for future learning.","feed_headline":"Sequential post-training collapses LLM representations and cuts later adaptability","feed_subtitle":"Measurements show concentration predicts weaker plasticity, out-of-domain performance, and calibration across stage orders.","key_machinery":"The measurement suite tracking rank, anisotropy, and homogeneity of hidden states, logits, token trajectories, and LoRA updates, which quantifies representation collapse and its link to reduced future learnability.","core_discovery":"Sequential post-training causes representation collapse, measured as progressive reduction in rank, increase in anisotropy, and rise in homogeneity across hidden states, logits, token paths, and parameter updates. This collapse is not merely geometric but predicts measurable drops in plasticity during later adaptation stages, weaker performance on out-of-distribution tasks, and degraded probability calibration. Controlled experiments varying stage order show that certain sequences accelerate the collapse while others slow it, and lightweight interventions including replay buffers and regularization terms can reduce collapse without erasing the gains from each post-training step.","pith_inferences":["Training pipelines could insert diversity checks after each major stage to decide whether to continue or reset.","The same concentration pattern may appear in non-LLM sequential learning settings such as chained reinforcement learning agents.","Routine monitoring of the measurement suite could guide when to apply corrective interventions during production post-training runs."],"forward_implications":["Models that reach higher representation concentration after early post-training stages exhibit measurably lower plasticity when a new task is introduced.","Out-of-domain generalization declines as hidden-state homogeneity increases across the sequence of training stages.","Probability calibration worsens in proportion to the degree of representation collapse induced by prior stages.","Certain orderings of fine-tuning, preference optimization, and specialization accelerate collapse more than others.","Mixed-domain replay and diversity regularization preserve measurable future learnability while retaining stage-specific behavioral improvements."],"fun_headline_variants":["Sequential post-training collapses LLM representations","Post-training chains compress LLM features and reduce plasticity","Representation collapse from sequential tuning limits LLM adaptability","Sequential tuning collapses hidden states and cuts LLM plasticity"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The defined measurements of hidden states, logits, token trajectories, and LoRA updates isolate representation collapse and establish its causal connection to reduced plasticity and generalization under the tested stage sequences.","fun_headline_variants_meta":{"raw":{"variants":["Sequential post-training collapses LLM representations","Post-training chains compress LLM features and reduce plasticity","Representation collapse from sequential tuning limits LLM adaptability","Sequential tuning collapses hidden states and cuts LLM plasticity"]},"model":"grok-4.3","cost_usd":0.004356,"raw_usage":{"total_tokens":2160,"prompt_tokens":620,"num_sources_used":0,"completion_tokens":53,"cost_in_usd_ticks":43562000,"prompt_tokens_details":{"text_tokens":620,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1487,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":620,"tokens_out":53,"duration_ms":16222,"temperature":1.0,"reasoning_tokens":1487,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T08:08:34.088418+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the same later adaptation stage on two models that differ only in measured representation concentration but show identical plasticity, out-of-domain accuracy, and calibration scores would falsify the claimed predictive link.","supporting_citations":[],"review_version":1}