{"id":"bc5c2904-2159-4ecd-b387-cc026d5543a1","arxiv_id":"2606.00132","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"FoLoRA applies generalized Rayleigh-quotient optimization to LoRA updates so that directions are gated by downstream utility divided by a pretraining-proxy forgetting penalty.","lead":"The paper proposes FoLoRA, which uses a generalized Rayleigh quotient to score fine-tuning update directions by task utility per unit of estimated forgetting penalty. A smart generalist might read it to understand a new optimization framing for keeping foundation models capable after specialization.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"First-order preservation condition may fail to accurately proxy nontarget capability degradation","rationale":"The reader's weakest assumption is precisely the load-bearing step: the first-order proxy must be sufficient. The abstract supplies no derivation, ablation, or correlation check that would confirm this, so the experimental superiority remains conditional on that untested link. Full text would be needed to see whether any such validation appears; absent it, the concern stands.","tokens_in":1660,"tokens_out":332,"duration_ms":15278,"concrete_test":"After FoLoRA adaptation on one of the reported tasks (e.g., math), compute the per-direction forgetting penalties from the first-order condition and compare their ranking against measured performance drops on a held-out suite of nontarget benchmarks (MMLU, HumanEval subsets, etc.) never seen during training or proxy sampling; if rank correlation (Kendall tau) falls below 0.6 across 5 random seeds, the penalty does not accurately reflect degradation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The method scores update directions via a generalized Rayleigh quotient whose denominator is a forgetting penalty derived from a first-order preservation condition on sampled pretraining-proxy activations. For the headline experimental claim (strongest preservation-adaptation balance) to follow, this linear penalty must reliably rank directions by their true effect on nontarget capabilities. Nothing in the provided description establishes that the first-order term dominates higher-order interactions, distribution shifts, or capability-specific nonlinearities that arise during adaptation; the proxy sampling procedure is also left unspecified at the level needed to guarantee coverage of the capabilities being preserved.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes Foundation-Preserving LoRA (FoLoRA), a forgetting-aware optimization method for adapting foundation models. It constructs a forgetting penalty from a first-order preservation condition on activations sampled from the pretrained model (pretraining-proxy data) and a task utility on downstream activations. Update directions are scored via a generalized Rayleigh quotient of utility per unit penalty; the resulting spectral basis is used for direction-wise gated Adam updates that attenuate low-utility directions. Experiments on math, code, and instruction-following adaptation tasks are reported to show that FoLoRA attains the best aggregate preservation of non-target capabilities while improving target-task performance relative to baselines.","tokens_in":1804,"tokens_out":614,"duration_ms":24042,"significance":"If the first-order preservation condition reliably ranks directions by their effect on non-target capabilities, the work supplies a training-time mechanism for regulating the adaptation-preservation trade-off that is more flexible than fixed initialization or constraint methods. The use of a generalized Rayleigh quotient to induce a spectral coordinate system for gated updates is a technically interesting construction that could apply to other constrained fine-tuning settings. The proxy-data construction by direct sampling from the pretrained model avoids dependence on a single external calibration set and is a concrete strength. The significance is limited by the absence of an explicit derivation or quantitative validation that the linear penalty dominates higher-order effects.","major_comments":[{"comment":"Abstract and method description: the first-order preservation condition is invoked to define the forgetting penalty, yet no explicit statement of the condition, the precise form of the generalized Rayleigh quotient, or any supporting derivation is supplied. Without these, it is impossible to verify whether the quotient reduces to an implicit utility-penalty hyperparameter or whether the linear term suffices to proxy non-target capability degradation.","section":"Abstract / method"},{"comment":"§4 (experiments): the headline claim that FoLoRA achieves the 'strongest preservation-adaptation balance' is stated without quantitative results, error bars, ablation tables, or statistical tests. This absence makes the empirical superiority load-bearing for the central contribution impossible to evaluate.","section":"§4"},{"comment":"Proxy-data construction paragraph: sampling from the pretrained model is presented as avoiding a single proxy dataset, but the procedure for choosing the sampling distribution is unspecified. If the distribution itself depends on the model under adaptation, the forgetting penalty estimate risks circularity that would undermine the preservation guarantee.","section":"Method / proxy data"}],"minor_comments":[{"comment":"Notation for the generalized Rayleigh quotient and the gated Adam update rule should be introduced with explicit equations rather than descriptive prose.","section":"Method"},{"comment":"The abstract would benefit from a single-sentence statement of the first-order condition and the exact Rayleigh-quotient objective.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and detailed comments. We address each major point below and will incorporate clarifications and additional material in the revised manuscript.","responses":[{"response":"We agree that an explicit statement and derivation are needed. In the revision we will insert a dedicated subsection (new §3.2) that states the first-order preservation condition as the linear term in the Taylor expansion of activation change, defines the forgetting penalty matrix B from the expected squared norm of this change on pretraining-proxy activations, and presents the generalized Rayleigh quotient as the solution to max_u (u^T A u) / (u^T B u) where A encodes downstream utility. The resulting eigenvectors supply the spectral basis for the gated updates; this is not equivalent to a scalar hyperparameter. We will also add a short paragraph discussing the linear approximation as a computationally tractable proxy.","revision_made":"yes","referee_comment":"[Abstract / method] Abstract and method description: the first-order preservation condition is invoked to define the forgetting penalty, yet no explicit statement of the condition, the precise form of the generalized Rayleigh quotient, or any supporting derivation is supplied. Without these, it is impossible to verify whether the quotient reduces to an implicit utility-penalty hyperparameter or whether the linear term suffices to proxy non-target capability degradation."},{"response":"We accept that the experimental claims require fuller quantitative support. The revision will expand §4 with complete tables reporting mean and standard deviation over five random seeds for all tasks, ablation tables isolating the contribution of the Rayleigh-quotient gating, and paired statistical tests (t-tests with Bonferroni correction) comparing FoLoRA against each baseline on both target and non-target metrics.","revision_made":"yes","referee_comment":"[§4] §4 (experiments): the headline claim that FoLoRA achieves the 'strongest preservation-adaptation balance' is stated without quantitative results, error bars, ablation tables, or statistical tests. This absence makes the empirical superiority load-bearing for the central contribution impossible to evaluate."},{"response":"The sampling distribution is generated once from the frozen pretrained model using a fixed collection of generic prompts drawn from public corpora that contain no downstream-task examples. Because the model parameters remain unchanged during sampling, the procedure is independent of the subsequent adaptation. The revision will add an explicit paragraph (new §3.4) describing the prompt sources, temperature, and number of samples, together with a short argument confirming the absence of circularity.","revision_made":"yes","referee_comment":"[Method / proxy data] Proxy-data construction paragraph: sampling from the pretrained model is presented as avoiding a single proxy dataset, but the procedure for choosing the sampling distribution is unspecified. If the distribution itself depends on the model under adaptation, the forgetting penalty estimate risks circularity that would undermine the preservation guarantee."}],"tokens_in":1469,"tokens_out":616,"duration_ms":20742,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core move is to score LoRA directions with a generalized Rayleigh quotient that divides task utility by a forgetting penalty built from a first-order preservation condition on activations sampled from the base model itself. This produces a spectral system for gating Adam steps so low-utility directions get attenuated.\n\nWhat is new is the specific combination of the quotient with direction-wise gating and the proxy sampling step that avoids a fixed external dataset. The experiments on math, code, and instruction-following adaptation are presented as showing better target performance with stronger aggregate preservation than baselines.\n\nThe abstract supplies none of the actual equations, the precise form of the quotient, or any quantitative results, error bars, or ablation tables. Without those it is impossible to tell whether the Rayleigh coordinate system is doing independent work or simply encoding a tunable trade-off hyperparameter.\n\nThe main soft spot is the untested claim that the linear first-order penalty on proxy activations will reliably rank directions by their true effect on nontarget capabilities. Higher-order interactions, distribution shifts, and capability-specific nonlinearities are not addressed, so the stress-test concern stands. The proxy construction also carries a circularity risk that is not resolved in the given description.\n\nThis is for researchers working on parameter-efficient fine-tuning who want an automatic preservation mechanism inside the optimizer. It is worth sending to a serious referee so the derivations and experimental details can be checked; the framework is concrete enough to evaluate even if the current write-up is thin.","headline":"FoLoRA adds Rayleigh-quotient gating to LoRA updates but the first-order forgetting penalty lacks supporting math or evidence.","tokens_in":2295,"tokens_out":368,"would_cite":false,"duration_ms":14925,"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":"FoLoRA scores LoRA update directions via a generalized Rayleigh quotient of task utility over forgetting penalty to preserve non-target capabilities during adaptation.","keywords":["FoLoRA","LoRA adaptation","forgetting penalty","Rayleigh quotient","foundation model preservation","parameter-efficient fine-tuning","gated Adam updates","preservation-adaptation trade-off"],"falsifier":"An experiment in which FoLoRA updates produce equal or lower non-target task scores than standard LoRA while matching or exceeding target-task gains would refute the claimed advantage of the Rayleigh-quotient gating.","tokens_in":2577,"feed_emoji":"⚖️","tokens_out":651,"duration_ms":18439,"temperature":0.7,"pith_summary":"The paper proposes Foundation Preserving LoRA (FoLoRA), a method that regulates the adaptation-preservation trade-off dynamically during low-rank fine-tuning. It defines a forgetting penalty from first-order changes in pretraining-proxy activations and a task utility from downstream activations. These quantities form a generalized Rayleigh quotient whose eigenvectors supply a coordinate system for direction-wise gated Adam updates that down-weight low-utility directions. Calibration data for the penalty is obtained by sampling the pretrained model itself rather than a fixed proxy set. On math, code, and instruction-following tasks the resulting updates improve target performance while yielding the strongest aggregate retention of non-target skills among tested baselines.","feed_headline":"Rayleigh quotient ranks LoRA directions by utility per forgetting","feed_subtitle":"The resulting gated updates raise target performance while recording the strongest retention of non-target capabilities across math, code, a","key_machinery":"The generalized Rayleigh quotient that ranks candidate update directions by the ratio of downstream task utility to pretraining-proxy forgetting penalty, supplying the basis for gated Adam steps.","core_discovery":"FoLoRA constructs a spectral coordinate system from the generalized Rayleigh quotient of task utility divided by forgetting penalty, then performs direction-wise gated Adam updates that attenuate directions with low utility-to-penalty ratios, thereby improving target-task performance while achieving the best measured preservation of non-target capabilities.","pith_inferences":["If the first-order penalty correlates reliably with downstream forgetting, the same quotient construction could be tested on full fine-tuning or other adapter families.","The approach implicitly treats preservation as a per-direction resource constraint rather than a global regularization term.","Extending the calibration sampling to include synthetic or out-of-distribution prompts might further tighten the penalty estimate."],"forward_implications":["Directions that deliver high task utility per unit of estimated forgetting receive larger effective learning rates.","Sampling calibration activations directly from the pretrained model removes dependence on any single external proxy dataset.","The same spectral gating procedure applies across math, code, and instruction-following adaptation settings.","Aggregate preservation of non-target capabilities improves even as target performance rises."],"fun_headline_variants":["FoLoRA gates Adam with Rayleigh quotient on utility per forgetting","Direction wise gated updates from generalized Rayleigh quotient scoring","Rayleigh quotient defines spectral coordinates for LoRA preservation","FoLoRA uses task utility over forgetting penalty in LoRA updates","Pretraining proxies calibrate forgetting penalty in Rayleigh optimization"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"A first-order preservation condition computed on activations sampled from the pretrained model is sufficient to define a forgetting penalty that tracks actual degradation of non-target capabilities.","fun_headline_variants_meta":{"raw":{"variants":["FoLoRA gates Adam with Rayleigh quotient on utility per forgetting","Direction wise gated updates from generalized Rayleigh quotient scoring","Rayleigh quotient defines spectral coordinates for LoRA preservation","FoLoRA uses task utility over forgetting penalty in LoRA updates","Pretraining proxies calibrate forgetting penalty in Rayleigh optimization"]},"model":"grok-4.3","cost_usd":0.006109,"raw_usage":{"total_tokens":2855,"prompt_tokens":608,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":61087000,"prompt_tokens_details":{"text_tokens":608,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2170,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":608,"tokens_out":77,"duration_ms":16262,"temperature":1.0,"reasoning_tokens":2170,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T08:36:55.802225+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment in which FoLoRA updates produce equal or lower non-target task scores than standard LoRA while matching or exceeding target-task gains would refute the claimed advantage of the Rayleigh-quotient gating.","supporting_citations":[],"review_version":1}