{"id":"a777258a-198e-4795-a8b9-19be30141dd6","arxiv_id":"2608.10260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single low-rank, sparsely trained lens family decodes residual, attention, and MLP states in models up to 70B parameters, revealing that visible and causally effective locations for a behavior can differ.","lead":"OmniLens makes 'lens' interpretability tools for large language models much cheaper, using low-rank math and sampling tricks so a dense set of internal states can be decoded at once, even for a 70-billion-parameter model. It reproduces several published interpretability results at lower cost and suggests that where a behavior is easiest to see inside a model is not where it is easiest to change.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The visibility-vs-intervention decoupling in §5 rests on component rankings from low-rank Subset-KL lenses; Appendix K's 4–5 hidden-state resolution shift and the absence of a full-rank componentwise reference leave that ordering unverified.","rationale":"The reader's weakest_assumption is exactly the one I find most load-bearing: the approximation must preserve the relative ordering of hookpoints, not merely task-level utility. The technical scaling contribution is well supported — the rank ablation, the unbiasedness theorem, the validated memory decomposition, and the fused-kernel benchmarks all check out — and I would not challenge that part. But the headline interpretability conclusion is an inference from small-n ranking comparisons made with the very approximations whose layerwise fidelity is weakest: early-layer agreement with the full-rank reference is poor (Table 7), prediction depth shifts by 4–5 hidden states (Appendix K), and at 8B the only full-rank reference reads through a half-block-mismatched translator. At 70B, no full-rank reference exists and the paper itself reports that targeted head ablation is no better than random, so the decoupling claim is supported only at GPT-2 and 8B. The paper is unusually transparent about these caveats in Appendices J, K, and L, which is why the appropriate disposition is the same CONDITIONAL rather than rejection. My concrete test is deliberately placed at 8B, where a full-rank componentwise reference is feasible; no such reference can exist at 70B under the paper's single-device lens placement, so verifying the ranking at 8B is the decisive check.","tokens_in":39566,"tokens_out":5354,"duration_ms":57556,"concrete_test":"Train a full-rank expanded-hookset lens stack (all 194 sites) on LLaMA-3-8B with the same Top-k+IS budget and schedule — this is within the paper's own memory accounting, since the full-rank optimizer bucket is ~39.1GB (Figure 14) and fits an A100-40GB with FSDP and a small microbatch — then rerun the DART visibility ranking and ToxIn intervention sweeps behind Figure 6d. Report the Spearman correlation between detected toxicity and intervention effectiveness across the six component types with a bootstrap confidence interval and a permutation test. If the correlation remains significantly negative and mlp_out remains the most effective target, the decoupling survives; if the ordering flips or the interval includes zero, the finding is an artifact of the rank/subset approximation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The new scientific finding in §5 — that hookpoints where a behavior is most visible need not be those where intervention is most effective — depends on preserving the relative ordering of hookpoints by visibility and by intervention efficacy across component types. The paper's own fidelity statistics show this preservation is not established. Table 7 gives Top-k+IS early-layer Kendall τ = 0.243 versus the full-rank reference (Top-k is −0.077); Appendix K reports Subset-KL lenses resolve predictions 4–5 hidden states later than the full-KL reference, and Figure 9 shows early-layer agreement degrades sharply. At 8B the only full-rank reference is residual-only and decodes heads through a half-block-mismatched block-input translator (Appendices J and L), and the selector comparison in Table 22 is single-run. At 70B there is no full-rank reference, and Appendix J states targeted head ablation is no more effective than random (Figure 30), so the 'outside attention' conclusion rests on GPT-2 and 8B. If the low-rank/subset approximation shifts visibility scores for MLP versus attention sites, the reported Spearman −0.43 over six component types, with no confidence interval or permutation test, is not evidence of decoupling.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces OmniLens, a trained-lens framework that attaches one low-rank, identity-residual translator family to residual-stream, attention, and MLP hookpoints, and trains it with two Subset-KL objectives: a biased Top-k truncation and an unbiased Top-k+IS importance-sampled estimator. The authors report large reductions in trainable parameters and peak memory, train a 482-lens ensemble on LLaMA-3-70B, run eight optimization steps on LLaMA-3.1-405B as a feasibility demonstration, and reproduce three lens applications: prompt-injection detection, multi-hop memory injection, and toxicity localization. The paper's central scientific claim beyond scaling is that model-wide lens coverage reveals a decoupling between where a behavior is most visible in lens readouts and where intervention is most causally effective, with a reported Spearman correlation of -0.43 between detected toxicity and intervention reduction across six component types at 8B.","tokens_in":39775,"tokens_out":6845,"duration_ms":68491,"significance":"If the scaling results hold, OmniLens is a substantial systems contribution: it combines two independent cost reductions (low-rank translators and subset-KL training), ships reproducible code, gives a correct unbiasedness proof for the Top-k+IS estimator in Appendix E, and validates its memory model against measured GPT-2 and 70B runs. Enabling dense trained-lens coverage at 70B and a training path at 405B is genuinely valuable for the interpretability community. The decoupling claim, if confirmed with proper statistical support, would be an important empirical constraint on how lens readouts are used, because it would show that visibility rankings do not generally imply intervention rankings. However, as detailed below, the evidence for that claim is currently not sufficient: it rests on six component-type aggregates with no uncertainty quantification and on approximation choices whose effect on the relevant rankings is not verified against a full-rank componentwise reference.","major_comments":[{"comment":"The headline decoupling result, Spearman -0.43 between detected toxicity and intervention effectiveness at 8B, is computed from six component-type aggregates with no confidence interval, no permutation test, and no cross-seed spread on the intervention axis. The paper itself notes in Appendix J that the selector comparison in Table 22 is a single-run evaluation and that differences of a few points are within run-to-run noise. As presented, the correlation cannot be distinguished from sampling noise. Please report bootstrap confidence intervals over lens seeds and prompts, and a permutation test over component labels, or explicitly demote the claim to a qualitative observation.","section":"§5, Fig. 6d"},{"comment":"The decoupling claim requires that the low-rank/Subset-KL approximation preserve the relative ordering of hookpoints by visibility and by intervention efficacy. The paper's own fidelity statistics do not establish this. Table 7 reports early-layer Kendall tau@100 of only 0.243 for the recommended Top-k+IS lens against the full-rank reference (and -0.077 for Top-k), and Appendix K states that Subset-KL lenses resolve predictions 4-5 hidden states later than the full-KL reference. Since toxicity detection is a fine-grained, partly early-layer signal, this is a real risk that the component rankings in Fig. 6d are artifacts of the approximation. A concrete remedy is to compute the same detection-vs-intervention Spearman at GPT-2 with the full-rank full-KL reference, which is trainable at that scale, and show that the low-rank and full-rank component orderings agree; at 8B no full-rank componentwise reference exists, so that check is essential rather than optional.","section":"Appendix K, Table 7"},{"comment":"The abstract states that 'the most effective interventions lie outside the attention heads examined by prior lens studies' without qualification, but the supporting evidence is limited to GPT-2 and 8B. At 70B, Appendix J reports that ablating the flagged heads is no more effective than ablating random heads (Fig. 30), so the outside-attention conclusion has no 70B support; the 8B selector comparison in Table 22 is single-run. Please qualify the abstract and Section 6 claims to 'at GPT-2 and 8B', or provide an intervention result at 70B that identifies an effective non-attention site.","section":"Abstract, §5, Appendix J"}],"minor_comments":[{"comment":"The six component-type points in the main-text figure are not directly labeled in the rendered figure; please add explicit labels or a legend so the reader can map each point to attn_in, attn_out, resid_mid, mlp_in, mlp_out, and resid_post without cross-referencing the text.","section":"Fig. 6d"},{"comment":"Appendix L states that the detection captures of Table 16 and the whole-model audit predate the 250-step cosine-annealed schedule and use earlier-schedule checkpoints. Please make this schedule mismatch visible in the main text, since the three case studies are presented as a single coherent evaluation and the inconsistency is currently easy to miss.","section":"Appendix L"},{"comment":"The caption's statement 'No full-rank reference is trainable at 70B (Section 4)' is ambiguous: Appendix G shows that LoRA+Full-KL fits at 35.5GB, so the missing reference is a full-rank translator, not a full-KL objective. Please rephrase to 'no full-rank translator reference is trainable at 70B'.","section":"Fig. 6d caption"},{"comment":"In the proof of Theorem 1, the display for the tail expectation omits the explicit average over the k_tail draws; the surrounding text explains it, but the equation would be clearer if the 1/k_tail sum appeared explicitly before taking the expectation.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The systems contribution is strong and the scaling results appear credible, but the paper's headline interpretability finding is currently under-supported. I would recommend major revision focused on validating the decoupling claim against a full-rank reference at GPT-2 and adding proper uncertainty quantification at 8B. If those additions are not feasible, the authors should clearly reposition the decoupling as a tentative observation rather than a central result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The scaling machinery is real: OmniLens' low-rank identity-residual translators plus the Subset-KL objectives make dense trained-lens coverage practical at 70B, and the 405B eight-step run shows the training path exists on commodity hardware. The paper ships code, and the memory model is validated against measured runs. That part is solid. The headline scientific finding — that hookpoints where a behavior is most visible need not be those where intervention is most effective — is plausible but not established. That claim rests on a Spearman −0.43 over six component types at 8B with no confidence interval, a single-run head selector comparison, and no full-rank componentwise reference at 70B. The paper's own Appendix K shows Subset-KL lenses resolve predictions 4–5 hidden states later than the full-KL reference and early-layer agreement is poor; if the approximation shifts component rankings, the decoupling could be an artifact. The stress-test note is fair. It is not a takedown: the authors are unusually transparent about these limits, and the decoupling may survive better measurement, but as written it is a suggestive observation, not a result.\n\nWhat is genuinely new: the hookpoint-agnostic lens family, the unbiased Top-k+IS estimator with the correct caveat that unbiasedness applies to raw stochastic gradients, the fused kernel that avoids gathering the subset, and the dense 482-lens ensemble at 70B. The reproduction of prior case studies at lower cost is useful validation, and the parameter/Frobenius-energy analysis showing low rank is a constraint rather than compression is well done. Citation pattern looks fair to me; self-citations are to specific prior implementations that are actually extended.\n\nSoft spots in proportion: besides the decoupling, the 70B full-rank reference is absent by design, so the fidelity of the density claim at 70B is not directly verified; the reduced evaluations (200 prompts, 40 prompts) are acknowledged. The detection study reproduces the original's framework, but the logit lens also detects well on the easy tasks, and the paper handles that by noting it. None of this undermines the scaling contribution.\n\nThis paper is for interpretability researchers who want dense readouts on large models and for systems people who care about memory-efficient distillation; the rigorous-estimator crowd will appreciate Appendix E. It deserves a serious referee. My recommendation: send to review, with the expectation of revision. The scaling and systems results are worth publishing; the visibility-vs-actionability claim should be either hardened with error bars, a full-rank reference at 8B for all component types, and a permutation test, or demoted to a hypothesis.","headline":"A solid scaling contribution whose headline interpretability claim needs better statistics before it is treated as established.","tokens_in":40371,"tokens_out":2377,"would_cite":true,"duration_ms":25189,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"OmniLens claims that one low-rank, identity-residual lens family can attach to any model-width hookpoint, making dense trained-lens coverage practical on models as large as LLaMA-3-70B.","keywords":["trained lenses","low-rank translators","Subset-KL","LLM interpretability","hookpoint-agnostic","dense lens coverage","visibility-action decoupling","LLaMA-3-70B"],"falsifier":"At 8B, train full-rank full-KL lenses on all six hookpoint types and rerun the toxicity audit; if MLP outputs no longer give the largest toxicity reduction or the negative Spearman correlation between detection and intervention disappears, the visibility-action decoupling is an artifact of the low-rank or subset-KL approximation.","tokens_in":39319,"feed_emoji":"🔍","tokens_out":5941,"duration_ms":50027,"temperature":0.7,"pith_summary":"OmniLens claims that one lens family, an identity-based low-rank translator combined with vocabulary-subset KL training, can attach to any model-width hookpoint, making dense trained-lens coverage practical on models as large as LLaMA-3-70B. The paper demonstrates 482 lenses at 70B with 90.5% fewer trainable parameters than a full-rank residual-only stack, and up to 70% lower peak memory on GPT-2. It reproduces three published interpretability studies, prompt-injection detection, multi-hop memory injection, and toxicity localization, at substantially lower cost. Model-wide coverage yields a new empirical constraint: the hookpoint where a behavior is most visible is often not the hookpoint where intervening changes the behavior most, and the best intervention sites lie outside attention heads. If right, trained-lens interpretability stops being a sparse, component-specific tool and becomes a model-wide instrument for near-frontier models.","feed_headline":"482 lenses decode a 70B model's hidden states","feed_subtitle":"Low-rank translators and subset-vocabulary training make dense interpretability cheap, and reveal a visibility-action gap.","key_machinery":"The load-bearing mechanism is the identity-residual low-rank translator $L_{\\ell,u}=I+\\frac{\\alpha}{r}B_{\\ell,u}A_{\\ell,u}$, which turns each hookpoint's activation into residual-stream space before the frozen final normalization and unembedding; because it starts as the identity and only needs to agree on activations the model actually produces, rank $r=64$ captures most of a full-rank lens's fidelity at a fraction of the parameters. The second mechanism is Subset-KL training: Top-$k$ truncation renormalizes both distributions over the teacher's $k$ most probable tokens, while Top-$k$+IS computes the head exactly and importance-samples the tail with a lens-independent proposal, so the estimated objective and its gradients are unbiased for the full KL (Theorem 1). A fused CUDA kernel computes only the selected logits without materializing the gathered unembedding rows, removing the memory blow-up that subset objectives would otherwise reintroduce. Together these two mechanisms make per-hookpoint parameter growth linear in width and keep vocabulary-side memory under control, which is what allows 482 lenses on a 70B model.","core_discovery":"OmniLens establishes that a single low-rank lens family can decode any model-width activation—residual, attention, or MLP—and that with this universal family dense coverage is tractable at 70B scale. The translator is an identity-residual low-rank map $L_{\\ell,u}=I+\\frac{\\alpha}{r}BA$, trained against the model's frozen final readout, with parameters growing linearly rather than quadratically in width. Two Subset-KL objectives supply the memory lever: Top-$k$ truncation renormalizes over the teacher's most probable tokens, and Top-$k$+IS keeps those head tokens exact while importance-sampling the tail, giving unbiased stochastic gradients for the full KL. On a 482-lens LLaMA-3-70B ensemble the paper reproduces published prompt-injection detection, multi-hop memory injection, and toxicity localization results at lower cost. The new finding that model-wide coverage exposes is that visibility and causal actionability decouple: the components where toxic language is most detectable are negatively correlated (Spearman $-0.43$ at 8B) with the components where intervention reduces toxicity, and the most effective intervention targets lie outside the attention heads studied previously.","pith_inferences":["Beyond the paper, a user of the released framework can test the decoupling directly on any 8B model: select intervention sites by measured causal effect rather than by lens visibility and compare; if visibility-selected sites continue to underperform, the paper's constraint becomes a general design rule for interpretability tools.","Beyond the paper, the rank-64 default makes early-layer analyses the fragile case; the paper's own ablations show layer-0 top-1 agreement of only 56.7%, so component or layer rankings in the first few layers should be read with wider error bars than late-layer rankings.","Beyond the paper, the identity-residual translator plus a learned input map could extend the same family to individual attention heads, which the paper lists as future work; if that holds, per-head interpretability at 70B becomes a matter of engineering rather than method design."],"forward_implications":["Trained-lens interpretability scales to near-frontier models: a 482-hookpoint ensemble trains on LLaMA-3-70B within A100-40GB budgets, and eight optimizer steps execute on LLaMA-3-405B.","Component-agnostic coverage makes cross-component comparison routine: the same lens family reads residual, attention, and MLP activations, so rankings of visibility and intervention efficacy can be compared without retraining separate lens families.","Visibility and actionability can diverge: at 8B, the components with the strongest detected toxicity signal and the components whose modification most reduces toxicity are negatively correlated (Spearman $-0.43$), so localization studies that rely on readout visibility alone may pick the wrong intervention site.","Established lens applications survive the approximations: prompt-injection detection reaches mean AUROC 0.997 on the five classification tasks at every scale, memory injection captures 61–82% of achievable gain at $\\tau=4$, and toxicity localization reproduces the published head concentration.","Full-rank references are not needed for task-level utility: at 70B, where no full-rank reference is trainable, the Subset-KL lenses detect knowledge-task injections the logit lens misses."],"supporting_citations":[{"why":"Supplies the tuned-lens baseline, evaluation protocol, and the prompt-injection detection task OmniLens must reproduce.","marker":"Belrose et al. 2023"},{"why":"Defines the attention-head lens family and the per-head decoder cost that OmniLens generalizes.","marker":"Sakarvadia et al. 2023"},{"why":"Contributes the low-rank adaptation idea that the identity-residual translator parameterization builds on.","marker":"Hu et al. 2021"},{"why":"Provides the LoRA lens and toxicity-localization protocol (DART) that OmniLens extends to all hookpoint types.","marker":"Pettyjohn 2025"},{"why":"Supplies the RS-KD baseline that Top-k+IS improves on with its exact head and importance-sampled tail.","marker":"Anshumann et al. 2025"},{"why":"Supplies The Pile corpus used to train all lenses.","marker":"Gao et al. 2020"},{"why":"Supplies the 2WikiMultiHop dataset and multi-hop reasoning setup used in the memory-injection case study.","marker":"Ho et al. 2020"},{"why":"Introduces the logit lens that OmniLens uses as a frozen-readout baseline.","marker":"Nostalgebraist 2020"}],"fun_headline_variants":["Universal low-rank lens decodes 70B models at scale","OmniLens: 482 lenses for 70B reveal visibility-action gap","Low-rank translators make 70B interpretability affordable","One lens family fits any width: dense 70B coverage"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central new finding—that visible hookpoints differ from effective intervention sites—rests on the assumption that its rank-64 and subset-vocabulary approximations preserve the relative ordering of hookpoints; the paper's own appendix shows the approximations resolve predictions several hidden states later than the full-KL reference, so if the approximations shift rankings, the decoupling could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Universal low-rank lens decodes 70B models at scale","OmniLens: 482 lenses for 70B reveal visibility-action gap","Low-rank translators make 70B interpretability affordable","One lens family fits any width: dense 70B coverage"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":3040,"prompt_tokens":1078,"completion_tokens":1962,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":1889}},"tokens_in":694,"tokens_out":1962,"duration_ms":13392,"temperature":1.0,"reasoning_tokens":1889,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:11:04.974090+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At 8B, train full-rank full-KL lenses on all six hookpoint types and rerun the toxicity audit; if MLP outputs no longer give the largest toxicity reduction or the negative Spearman correlation between detection and intervention disappears, the visibility-action decoupling is an artifact of the low-rank or subset-KL approximation.","supporting_citations":[{"cited_title":"2025 , howpublished =","cited_arxiv_id":null,"evidence_quote":"Provides the LoRA lens and toxicity-localization protocol (DART) that OmniLens extends to all hookpoint types."}],"review_version":1}