{"id":"6524de18-c2af-4ada-b249-72fbc48218ae","arxiv_id":"2606.29282","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ScaleErasure erases unsafe concepts in next-scale AR image generation via minimal logit guidance from unsafe and safe conditioned forward passes.","lead":"The paper introduces ScaleErasure, an inference-time method that erases unsafe concepts from next-scale autoregressive image generators by running two extra forward passes and guiding the most relevant logits toward safe outputs across scales, tokens, and channels. A smart generalist might read it to see how safety controls can be added to emerging image models without retraining or major capability loss.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Assumption that unsafe/safe conditioned pass differences isolate unsafe logits without collateral shifts under entanglement","rationale":"The reader's weakest assumption is exactly the load-bearing step; the abstract-only review correctly flags it, and the method description supplies no additional mechanism (e.g., orthogonality proof or ablation on entangled features) that would remove the risk.","tokens_in":1710,"tokens_out":292,"duration_ms":9565,"concrete_test":"On a held-out set of prompts containing both the target unsafe concept and an orthogonal safe concept (e.g., 'cat' when erasing 'violence'), run ScaleErasure and measure (a) success rate on the unsafe concept and (b) change in CLIP similarity or human preference for the orthogonal concept versus the unguided baseline; if (b) degrades by more than the variance of the baseline, isolation has failed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the two extra forward passes (unsafe concept vs. safe concept) yield logit differences that, after selection across scales/tokens/bit-channels, affect only the target unsafe semantics. In next-scale AR, early-scale tokens compress multiple semantics into the same representations; nothing in the described procedure guarantees the difference vector is orthogonal to unrelated content. If the selected dimensions still contain entangled signals, guidance will alter non-target semantics or introduce artifacts even if the headline metrics improve.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes ScaleErasure, an inference-time method for precise concept erasure in next-scale autoregressive image generation. It performs two additional forward passes (unsafe concept and safe concept) to compute logit differences, then selectively guides the target generation's logits across three dimensions—scales, tokens, and bit channels—to steer away from unsafe semantics while minimizing changes to unrelated content. The central claim is that this minimal intervention outperforms adapted baselines under severe semantic entanglement at early scales and largely preserves general generative capability, with code released.","tokens_in":1790,"tokens_out":461,"duration_ms":28972,"significance":"If the isolation of unsafe logits holds, the work addresses an unexplored gap in safety techniques for the emerging next-scale AR paradigm, where early-scale compression creates high entanglement. The inference-time, parameter-free nature and open code are practical strengths that could enable safer deployment without retraining.","major_comments":[{"comment":"§3 (Method): The procedure defines logit guidance directly from the difference of the two conditioned forward passes, but provides no derivation or bound showing that the selected dimensions (after scale/token/bit-channel masking) are orthogonal to non-target semantics. Under the entanglement described in the introduction, this risks collateral shifts; a concrete test (e.g., measuring change in unrelated concept scores before/after guidance) is needed to support the 'precise' and 'minimal' claims.","section":"§3"},{"comment":"§4 (Experiments): The abstract asserts outperformance and capability preservation, yet the soundness assessment notes absence of quantitative metrics, ablations on the three selection dimensions, or error analysis in the provided text. Without tables reporting, e.g., unsafe-concept success rate vs. FID or CLIP similarity on safe concepts, the central claim that the method 'enables effective erasure' cannot be verified as load-bearing.","section":"§4"}],"minor_comments":[{"comment":"Notation for the three-dimensional selection mask is introduced without an explicit equation; adding a compact definition (e.g., Eq. (X)) would improve clarity.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on our manuscript. The comments highlight important areas for strengthening the theoretical justification and empirical validation of ScaleErasure. We address each point below and will incorporate revisions to improve the paper.","responses":[{"response":"We acknowledge that a formal derivation or bound demonstrating orthogonality after masking would provide stronger theoretical support. Deriving such a bound is non-trivial due to the severe semantic entanglement at early scales in next-scale AR models, which is the core challenge highlighted in the introduction. The logit selection is motivated by empirical differences between unsafe and safe conditioned passes. To directly address the request for a concrete test, we will add an empirical evaluation in the revised §3 measuring changes in unrelated concept scores (via CLIP-based probes on safe concepts) before and after guidance. This will quantify collateral shifts and better support the 'precise' and 'minimal' claims.","revision_made":"yes","referee_comment":"[§3] §3 (Method): The procedure defines logit guidance directly from the difference of the two conditioned forward passes, but provides no derivation or bound showing that the selected dimensions (after scale/token/bit-channel masking) are orthogonal to non-target semantics. Under the entanglement described in the introduction, this risks collateral shifts; a concrete test (e.g., measuring change in unrelated concept scores before/after guidance) is needed to support the 'precise' and 'minimal' claims."},{"response":"We agree that the current presentation would benefit from more explicit quantitative reporting to make the claims load-bearing. While the manuscript includes comparative experiments and qualitative results, we accept the assessment that dedicated tables, ablations on the scale/token/bit-channel dimensions, and error analysis are missing from the provided text. In the revision, we will expand §4 with tables reporting unsafe-concept success rates alongside FID and CLIP similarity on safe concepts, plus ablations isolating each selection dimension and error analysis (e.g., failure cases). This will allow direct verification of outperformance and capability preservation.","revision_made":"yes","referee_comment":"[§4] §4 (Experiments): The abstract asserts outperformance and capability preservation, yet the soundness assessment notes absence of quantitative metrics, ablations on the three selection dimensions, or error analysis in the provided text. Without tables reporting, e.g., unsafe-concept success rate vs. FID or CLIP similarity on safe concepts, the central claim that the method 'enables effective erasure' cannot be verified as load-bearing."}],"tokens_in":1387,"tokens_out":534,"duration_ms":42571,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main contribution is a new inference-time procedure for next-scale autoregressive generators. It runs two extra conditioned passes, subtracts the logits, and then picks and steers only the dimensions across scales, tokens, and bit channels that appear most tied to the target concept. This is presented as the first erasure work in this specific paradigm, where early scales pack multiple semantics together.\n\nWhat stands out is the attempt to keep the change minimal by restricting guidance to selected coordinates rather than shifting the whole output. The authors correctly flag the entanglement issue that standard diffusion-style erasure would not directly transfer.\n\nThe obvious gap is the complete absence of numbers. The abstract says the method outperforms adapted baselines and preserves capability, yet gives no FID, CLIP scores, erasure success rates, or even qualitative examples. Without those, it is impossible to judge whether the three-dimensional selection actually avoids shifting unrelated content. The stress-test point lands here: nothing in the described procedure shows that the difference vector from the two extra passes is free of entangled signals, so the guidance could still alter non-target semantics even if headline metrics later look good.\n\nThe work is aimed at researchers building safety tools for the emerging next-scale AR family. It deserves a serious referee if the full paper contains reproducible experiments and ablations that address the isolation question; otherwise the central claim remains an unverified assertion. I would send it to review on the strength of the problem framing and the explicit code release, but would expect the reviewers to demand quantitative evidence before acceptance.","headline":"ScaleErasure claims a first method for concept erasure in next-scale AR models via three-way logit selection from safe/unsafe pass differences, but the abstract supplies zero metrics or examples so the isolation claim stays untested.","tokens_in":2285,"tokens_out":395,"would_cite":false,"duration_ms":22040,"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":"ScaleErasure erases unsafe concepts in next-scale autoregressive image generators by guiding selected logits at inference time.","keywords":["concept erasure","next-scale autoregressive","image generation","inference-time intervention","logit guidance","unsafe content","semantic entanglement"],"falsifier":"Apply ScaleErasure to a next-scale model and observe that unsafe concepts still appear in a measurable fraction of outputs or that non-target regions show consistent quality degradation relative to the unmodified model.","tokens_in":2612,"feed_emoji":"🖼️","tokens_out":650,"duration_ms":29143,"temperature":0.7,"pith_summary":"Next-scale autoregressive models compress semantic information at early scales, which creates tight entanglement between unsafe concepts and unrelated image content. ScaleErasure counters this by running the model twice more during inference, once conditioned on the unsafe concept and once on a safe counterpart, then using the output difference to steer only the most relevant logits. The steering happens across scales, tokens, and bit channels so that intervention stays minimal. A reader would care because the technique adds safety controls to these generators without any retraining or loss of broad generative ability. Experiments indicate the approach removes target concepts more cleanly than adapted baselines while keeping image quality intact.","feed_headline":"Logit guidance erases unsafe concepts in next-scale generators","feed_subtitle":"Two extra passes at inference steer relevant predictions away from unsafe ideas while preserving output quality.","key_machinery":"Three-dimensional logit selection and guidance (scales, tokens, bit channels) driven by output differences from unsafe-conditioned and safe-conditioned forward passes.","core_discovery":"ScaleErasure is an inference-time concept erasure method for next-scale autoregressive image generation that performs minimal intervention by precisely selecting and guiding predicted logits most relevant to the unsafe concept. It achieves this through two additional forward passes conditioned on the unsafe concept and a corresponding safe concept, then leverages their outputs to guide target logits away from unsafe semantics toward safe ones. Logits selection and guidance are conducted across three dimensions—scales, tokens, and bit channels—to enable effective erasure under severe semantic entanglement while largely preserving general generative capability.","pith_inferences":["The same difference-based guidance pattern could be tested on other autoregressive generation tasks such as video or audio.","Careful selection of the safe concept may be needed to avoid introducing new unintended biases in the outputs.","Combining this inference-time step with light fine-tuning could further strengthen erasure on particularly entangled concepts."],"forward_implications":["Next-scale autoregressive models can receive concept-level safety controls directly at inference without retraining.","Multi-dimensional guidance limits side effects on image content unrelated to the erased concept.","The method outperforms simple adaptations of earlier erasure techniques when semantic entanglement is severe.","Precise logit steering preserves the model's overall ability to generate diverse, high-quality images."],"fun_headline_variants":["ScaleErasure uses minimal logit intervention for concept erasure in AR images","Inference-time logit guidance erases unsafe concepts precisely in next-scale gen","Extra forward passes guide relevant logits away from unsafe concepts","Precise erasure of unsafe concepts via scale token and channel logit selection","ScaleErasure steers logits across scales tokens and channels at inference"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The differences between the two extra forward passes isolate exactly the logits tied to the unsafe concept without shifting unrelated content or introducing new artifacts.","fun_headline_variants_meta":{"raw":{"variants":["ScaleErasure uses minimal logit intervention for concept erasure in AR images","Inference-time logit guidance erases unsafe concepts precisely in next-scale gen","Extra forward passes guide relevant logits away from unsafe concepts","Precise erasure of unsafe concepts via scale token and channel logit selection","ScaleErasure steers logits across scales tokens and channels at inference"]},"model":"grok-4.3","cost_usd":0.003761,"raw_usage":{"total_tokens":1955,"prompt_tokens":686,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":37612000,"prompt_tokens_details":{"text_tokens":686,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1184,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":686,"tokens_out":85,"duration_ms":15363,"temperature":1.0,"reasoning_tokens":1184,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T07:59:57.938179+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Apply ScaleErasure to a next-scale model and observe that unsafe concepts still appear in a measurable fraction of outputs or that non-target regions show consistent quality degradation relative to the unmodified model.","supporting_citations":[],"review_version":1}