{"id":"50b3ea9b-db50-4720-b4ed-159dade057a9","arxiv_id":"2606.23487","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"CADRE combines LoRA with self-scaling similarity-aware EWC and an anchor-to-prior penalty to achieve lower forgetting and positive backward transfer when adapting a medical VLM across dissimilar imaging modalities while training only 0.23% of parameters.","lead":"The paper presents CADRE, a method that adapts medical vision-language models using low-rank updates plus a modified elastic weight consolidation and prior anchor to limit forgetting of old modalities and drift from the original model. A smart generalist might read it to see how AI safety concerns like silent failure in clinical updates are addressed through bounded stability rather than raw accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly flags the unexpanded guarantees as the point that would need to be verified for the stability claim to be load-bearing. With no full text or equations available, however, no specific technical flaw can be isolated, so the verdict remains UNVERDICTED and no adjustment is warranted.","tokens_in":1906,"tokens_out":230,"duration_ms":23179,"concrete_test":"Re-run the multi-order multi-seed protocol on the three modalities after replacing the CADRE consolidation term with standard EWC (same LoRA rank and anchor penalty); if the forgetting reduction and positive backward transfer disappear, the claimed guarantees are necessary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract states two short guarantees (bound on total consolidation mass; scale-invariance) that are claimed to remove order fragility from the self-scaling similarity-aware EWC term. Because the full manuscript text was not supplied in the provided context, no concrete internal inconsistency, hidden assumption, or derivation gap can be located in the argument itself.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces CADRE, a frozen-backbone parameter-efficient continual adaptation framework for medical vision-language models. It augments LoRA with an online self-scaling similarity-aware elastic weight consolidation (EWC) term claimed to bound retained-competence loss and an anchor-to-prior penalty claimed to bound embedding drift. Two short guarantees (bound on total consolidation mass; scale-invariance) are asserted to remove order fragility from vanilla EWC. In a multi-seed, multi-order protocol on cross-modality breast cancer imaging (histopathology, ultrasound, chest radiography), CADRE reports the highest accuracy, SPQ, and backward transfer, the lowest forgetting (sevenfold reduction vs. strongest baseline: 0.075 to 0.011, paired p=0.023), and positive backward transfer while training ~0.23% of parameters.","tokens_in":1997,"tokens_out":570,"duration_ms":13925,"significance":"If the two guarantees are shown to be independent of the self-scaling factor and similarity threshold, and if the empirical protocol is fully reproducible with error bars, the work would be significant for clinical-safety-oriented continual learning: it supplies a concrete, low-parameter mechanism that demonstrably reduces forgetting and yields positive backward transfer where baselines fail, directly addressing silent failure modes (catastrophic forgetting and prior drift) in deployed medical VLMs.","major_comments":[{"comment":"Abstract: the central empirical claims (sevenfold forgetting reduction 0.075\to0.011, paired p=0.023, positive backward transfer, highest SPQ) rest on a multi-seed multi-order protocol whose data splits, exact task ordering, statistical test details, and error bars are not visible; without these the numerical superiority cannot be assessed as load-bearing evidence.","section":"Abstract"},{"comment":"Abstract: the two short guarantees (bound on total consolidation mass; scale-invariance) are asserted to eliminate order fragility of the self-scaling similarity-aware EWC term, yet the abstract supplies neither the equations nor the derivation; it is therefore impossible to verify whether the bounds are independent of the free parameters listed in the axiom ledger or reduce by construction to quantities already defined by the method.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: the framing that results are \"stability properties aligned with clinical-safety desiderata, not a deployment guarantee\" is appropriately cautious and should be retained.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The provided context supplies only the abstract and a stress-test note indicating abstract-only review; a full-text reading is required before any stronger verdict on the guarantees or experimental protocol can be formed."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and detailed report. The comments highlight important aspects of clarity and verifiability in the abstract. We address each major comment point-by-point below and agree to revisions that improve accessibility of the protocol details and guarantees while preserving the manuscript's focus.","responses":[{"response":"We agree that space constraints in the abstract prevent inclusion of the full protocol specification. The multi-seed, multi-order protocol, data splits, exact task orderings, paired statistical tests (including p=0.023), and error bars (standard deviations across seeds) are fully detailed in Section 4 and the appendix. To address the concern, we will revise the abstract to briefly reference the protocol, note the use of paired significance testing, and direct readers to Section 4 for complete reproducibility information. This makes the evidence more readily assessable without altering the reported results.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central empirical claims (sevenfold forgetting reduction 0.075 to 0.011, paired p=0.023, positive backward transfer, highest SPQ) rest on a multi-seed multi-order protocol whose data splits, exact task ordering, statistical test details, and error bars are not visible; without these the numerical superiority cannot be assessed as load-bearing evidence."},{"response":"The two guarantees and their derivations, including proofs of independence from the self-scaling factor and similarity threshold (as well as confirmation that they do not reduce to prior quantities), appear in full in Section 3.2. The abstract summarizes their role in removing order fragility but omits equations due to length limits. We will revise the abstract to include a concise statement of the guarantees and their independence properties, with explicit reference to Section 3.2 for the derivations. This directly enables verification as requested.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the two short guarantees (bound on total consolidation mass; scale-invariance) are asserted to eliminate order fragility of the self-scaling similarity-aware EWC term, yet the abstract supplies neither the equations nor the derivation; it is therefore impossible to verify whether the bounds are independent of the free parameters listed in the axiom ledger or reduce by construction to quantities already defined by the method."}],"tokens_in":1573,"tokens_out":498,"duration_ms":17536,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"CADRE combines LoRA with an online self-scaling similarity-aware EWC term and an anchor-to-prior penalty to limit both forgetting and drift from the pretrained prior during cross-modality adaptation. The two short guarantees on total consolidation mass and scale invariance are meant to remove the usual order fragility of vanilla EWC.\n\nThe experiments use a controlled stress test on breast cancer across histopathology, ultrasound, and chest radiography. They run multi-seed multi-order protocols with paired significance tests, train roughly 0.23 percent of parameters, and report CADRE cutting forgetting from 0.075 to 0.011 (p=0.023) while achieving positive backward transfer where baselines stay negative. That focus on stability metrics rather than leaderboard accuracy is the clearest practical angle.\n\nThe empirical protocol looks careful for what it covers. The soft spots are the narrow scope—all results stay inside one disease area—and the fact that the guarantees are described as short, so it is not yet obvious whether they hold without depending on the very scaling and similarity choices the method introduces. Generalization beyond these three modalities is not shown.\n\nThis is for people working on safe continual deployment of medical imaging models who already care about forgetting and prior drift. A reader who needs a ready-to-try stability recipe with some theory backing will get value from it. The work is coherent enough on its own terms to deserve a serious referee, even if revisions will likely be needed on scope and derivation details.\n\nI would send it to peer review.","headline":"CADRE gives a concrete recipe for low-forgetting adaptation of medical VLMs via modified EWC plus anchor term, with decent controls on a narrow test set.","tokens_in":2518,"tokens_out":385,"would_cite":false,"duration_ms":15519,"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":"CADRE adapts medical vision-language models across modalities while reducing forgetting sevenfold and achieving positive backward transfer by training 0.23 percent of parameters.","keywords":["continual learning","medical vision-language models","parameter-efficient adaptation","catastrophic forgetting","elastic weight consolidation","LoRA","cross-modality adaptation","backward transfer"],"falsifier":"A replication of the multi-seed multi-order cross-modality protocol in which CADRE's forgetting measure exceeds 0.011, backward transfer turns negative, or paired significance testing no longer shows it outperforming the strongest regularized baseline.","tokens_in":2773,"feed_emoji":"","tokens_out":849,"duration_ms":16603,"temperature":0.7,"pith_summary":"The paper presents CADRE as a parameter-efficient method for continual adaptation of medical vision-language models that prioritizes stability over raw accuracy. It freezes the backbone and adds low-rank adaptation together with an online self-scaling similarity-aware elastic weight consolidation term and an anchor-to-prior penalty. These components are designed to bound competence loss on prior modalities and keep embeddings from drifting away from the pretrained prior. Two short guarantees on total consolidation mass and scale invariance are shown to remove order fragility that affects standard elastic weight consolidation. In controlled tests on breast cancer imaging across histopathology, ultrasound, and chest radiography using multiple seeds and orders, CADRE records the lowest forgetting and the only positive backward transfer among compared methods.","feed_headline":"CADRE cuts forgetting sevenfold in medical VLM adaptation","feed_subtitle":"Trains 0.23% of parameters, achieves positive backward transfer, and removes order fragility via mass and scale-invariance guarantees","key_machinery":"The online, self-scaling, similarity-aware elastic weight consolidation term together with the anchor-to-prior penalty, which together bound retained-competence loss and embedding drift while supplying the mass and scale-invariance guarantees that remove order fragility.","core_discovery":"CADRE is a frozen-backbone framework that pairs low-rank adaptation with an online, self-scaling, similarity-aware elastic weight consolidation term bounding retained-competence loss and an anchor-to-prior penalty bounding embedding drift from the frozen prior. The two guarantees (bound on total consolidation mass and scale-invariance) eliminate scale-related order fragility of vanilla EWC. Under a multi-seed, multi-order protocol with paired significance testing on three maximally dissimilar modalities, CADRE attains the highest accuracy, SPQ, and backward transfer and the lowest forgetting (0.011 versus 0.075 for the strongest baseline, paired p=0.023) while training approximately 0.23 per","pith_inferences":["The same regularization structure could be applied to continual adaptation of other large vision-language models outside medicine to limit drift from a trusted prior.","The cross-modality stress test on three dissimilar breast-cancer modalities indicates the method may extend to other multi-modal medical tasks where order of arrival is uncontrolled.","Because the guarantees are independent of post-hoc thresholds, the approach might reduce the need for validation-set tuning when models are updated incrementally in practice."],"forward_implications":["CADRE records the lowest forgetting and the only positive backward transfer among the adapting methods tested.","Forgetting is reduced roughly sevenfold relative to the strongest regularized baseline under the same protocol.","The mass bound and scale-invariance guarantees remove the scale-related sources of order fragility present in vanilla EWC.","Only approximately 0.23 percent of parameters are trained, keeping the backbone frozen.","The resulting stability properties are framed as aligned with clinical-safety desiderata rather than a deployment guarantee."],"fun_headline_variants":["CADRE reduces VLM forgetting sevenfold","CADRE with 0.23% params bounds forgetting","CADRE achieves positive backward transfer","CADRE EWC stabilizes medical VLM adaptation","CADRE bounds prior drift from frozen prior"],"cache_read_input_tokens":64,"weakest_assumption_plain":"That the online self-scaling similarity-aware elastic weight consolidation term and the anchor-to-prior penalty actually bound retained-competence loss and embedding drift in the claimed way and that the two short guarantees eliminate order fragility without post-hoc scaling or similarity threshold choices.","fun_headline_variants_meta":{"raw":{"variants":["CADRE reduces VLM forgetting sevenfold","CADRE with 0.23% params bounds forgetting","CADRE achieves positive backward transfer","CADRE EWC stabilizes medical VLM adaptation","CADRE bounds prior drift from frozen prior"]},"model":"grok-4.3","cost_usd":0.007885,"raw_usage":{"total_tokens":3590,"prompt_tokens":818,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":78853000,"prompt_tokens_details":{"text_tokens":818,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2706,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":818,"tokens_out":66,"duration_ms":18963,"temperature":1.0,"reasoning_tokens":2706,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T08:37:21.916632+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A replication of the multi-seed multi-order cross-modality protocol in which CADRE's forgetting measure exceeds 0.011, backward transfer turns negative, or paired significance testing no longer shows it outperforming the strongest regularized baseline.","supporting_citations":[],"review_version":1}