{"id":"0c4760ce-d0a2-4537-824d-027696210e49","arxiv_id":"2608.09521","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen shared activation bus with one encoder-decoder adapter per model lets probes, a single SAE, and a carrier NLA be reused across compatible language models.","lead":"The paper builds a shared activation space that lets several language models connect through one small adapter pair per model. This could let interpretability tools such as probes, autoencoders, and explanation models be built once and reused across many models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bus's added value is untested: the compatibility screen already requires near-perfect linear maps to a reference, so tool transfer may be inherited from the screen rather than from the frozen bus.","rationale":"The reader's weakest_assumption identifies the same load-bearing spot: the label-free screen pre-selects linearly alignable models, so the bus's near-perfect alignment and tool transfer may be inherited from the screen rather than produced by the bus. I agree, and I sharpen it into a concrete missing control: a ridge-map baseline. The paper is honest about restricting attention to compatible models, and it reports genuine supporting evidence—sharply bimodal screening scores, leave-one-model-out onboarding, ten additional checkpoints with all 99 behavioral-axis assignments correct, and calibration saturation at 10K positions. None of that, however, isolates the bus architecture from the screen's selection criterion. The screen itself already provides one linear map per model to a common reference, which is functionally an interface; the bus adds shared MLPs, decoders, and frozen coordinates. Without comparing against that baseline, the central claim is underdetermined: the reported universality could be a restatement of the compatibility filter rather than a property of the bus. The condition for acceptance should include this baseline comparison. I therefore keep the reader's CONDITIONAL verdict unchanged; the concern is substantial but addressable, not a refutation.","tokens_in":19560,"tokens_out":8380,"duration_ms":81053,"concrete_test":"Run a baseline in which the interface is replaced by the compatibility-screen ridge maps: fit the closed-form ridge maps from each model's z-scored activations to Llama-3.2-3B as in Appendix B.2, map all model activations into Llama's native space, and repeat the Section 5 evaluations (pooled probes, shared TopK SAE, Qwen2.5 NLA routing) and Section 6 front–back execution using exactly the same data splits and metrics. Compare ID/OOD R@1, macro AUROC, SAE Jaccard, NLA accuracy, and composed-model agreement against the bus results. If the ridge baseline is within a few points on all metrics, the bus's contribution is unestablished; if the bus is substantially better, the screen-inheritance concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the frozen Universal Activation Bus—not the pre-selected pool—is what makes cross-model tool reuse possible. The label-free screen (Section 4; Appendix B.2) already demands a closed-form ridge map from each retained model to a fixed Llama reference, with held-out 4,096-way R@1 of 98.6–99.6% for retained models and ≤5% for rejected ones. The bus is then trained only on this pool, so a much simpler interface exists that the paper never benchmarks: use those ridge maps (one per model to the common reference) and train probes, the shared SAE, and the NLA routing in the reference's native space. If that ridge baseline reproduces the retrieval and tool-transfer numbers in Tables 2, 4, and 5, the reported universality is a property of the screen's linear-alignability filter, not of the bus's shared modules, frozen coordinates, or adapter-only onboarding. If the bus clearly beats the ridge baseline, the concern is resolved. This is load-bearing because every headline evaluation is performed after the screen has restricted the model set to linearly alignable families.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Universal Activation Bus: a shared dense space Z with frozen shared modules, where each model is attached by a per-model linear encoder–decoder adapter pair. After joint source training, a new model is onboarded by fitting only its own adapters on unlabeled matched text. The authors claim that this frozen interface lets shared linear probes, one shared SAE dictionary, and a carrier-specific NLA be reused across compatible models, and that a translated intermediate activation can be executed by another model's frozen upper layers. The manuscript reports near-perfect held-out cross-model retrieval on the ID pool, somewhat lower OOD retrieval, pooled probes that match native probes in AUROC, high shared-SAE feature overlap with no onboarding gap, NLA routing accuracy of 84.9–94.4% against 95.3% native, and cross-model next-token agreement of 72.4–87.7% with the receiving model.","tokens_in":19784,"tokens_out":7233,"duration_ms":72071,"significance":"If the central claim holds, the paper makes a substantial contribution: it replaces pair-specific connectors with a model-wise activation contract, and it demonstrates that probes, an SAE dictionary, and an NLA can survive adapter-only onboarding. The strengths of the paper are its clear three-RQ structure, the detailed appendices with training configurations, the leave-one-out and ten-checkpoint generality experiments, and the explicit admission that the scope is limited to linearly compatible models. The main weakness is that the label-free compatibility screen already demonstrates near-perfect linear maps from each retained model to a fixed reference, and the paper never benchmarks the simpler ridge-map interface against the bus; without that comparison, the added value of the bus's shared modules, frozen coordinates, and contrastive training is not established. The NLA sentiment failure on the main onboarded model and the absence of error bars further temper the strength of the tool-reuse claims.","major_comments":[{"comment":"The manuscript never benchmarks the simpler interface that the compatibility screen already provides. The screen (Section 4, Appendix B.2, Table 13) fits a closed-form ridge map from each candidate's z-scored activations to a fixed Llama-3.2-3B reference and retains models with held-out 4,096-way R@1 of 98.6–99.6%, while the bus is then trained only on this pre-selected pool. A natural baseline is to use those ridge maps as per-model encoders into Llama's native space (with the inverse ridge map as decoder), train the probes and shared SAE in Llama's native space, and onboard new models by fitting the ridge map. If that baseline reproduces the retrieval, probe, SAE, and NLA numbers in Tables 2, 4, and 5, then the reported universality is a property of the screen's linear-alignability filter rather than of the bus's shared modules or frozen coordinates. This is load-bearing for the central claim in Section 1, so I request this baseline comparison before acceptance.","section":"§4; Appendix B.2, Table 13; Section 5"},{"comment":"The NLA reuse result has a concrete failure mode that the paper does not explain. On the main onboarded model, OLMo, the routed verbalization classifier obtains only 6% recall on the sentiment axis, below the 11.1% chance level and far below the 100% native-carrier recall, even though the direct shared-space probe identifies OLMo sentiment well (Table 2). The paper attributes this to carrier routing or verbalization, but no diagnostic is provided. Because OLMo is the central onboarded model and NLA reuse is a headline RQ3 claim, this issue should either be analyzed (for example, by comparing per-example AV/AR outputs for sentiment inputs) or explicitly acknowledged as a limitation in the main text. Without this, the statement that the NLA 'remains informative across connected models' overstates the coverage of the result.","section":"Appendix C.3, Table 18; Section 5 (NLA)"},{"comment":"The headline numbers are reported as point estimates or ranges without confidence intervals, bootstrap errors, or repeated-seed variance, while several conclusions rest on small differences: pooled probe AUROC 0.966 vs native 0.963, SAE Jaccard values 0.591–0.614, and NLA accuracy gaps of 1–10 points. With evaluation pools of 3,600–4,096 items, binomial or bootstrap uncertainty is likely non-negligible. For example, a difference of 0.003 in AUROC between pooled and native probes is not meaningful without error bars. I request that at least the central quantitative claims in Tables 2, 4, and 5 report standard errors, bootstrap intervals, or variance over seeds.","section":"Tables 1, 2, 4, 5, and 19"}],"minor_comments":[{"comment":"Equation (7) writes L_bus = L_self + lambda_cross L_cross + lambda_nce L_nce, but lambda_self is omitted from the displayed equation; Table 12 defines lambda_self = 1.0. Please align the notation.","section":"Eq. (7) and Table 12"},{"comment":"The caption says '† marks adapter-only onboarding', but the plot does not visibly show a dagger on the OLMo points; please make the marker explicit or remove the mention.","section":"Figure 3"},{"comment":"The qualitative SAE features in Table 3 were selected from the high-overlap group before inspecting their text examples, as disclosed in Appendix C.2; the main text should state this selection criterion more prominently to avoid the impression of cherry-picking.","section":"Table 3 and Appendix C.2"},{"comment":"All cross-entropy columns are labeled Delta CE, but only Table 16 states the units (nats). Please add units to Tables 14 and 19 as well.","section":"Tables 14, 16, and 19"},{"comment":"The main text says calibration R@1 saturates at 10K positions, but the 3K condition uses a smaller batch size and self-FVE continues to improve from 0.696 at 10K to 0.734 at 100K. Please qualify the saturation claim so it is not read as applying to reconstruction fidelity.","section":"Appendix E.2 and Figure 4b"},{"comment":"The title and abstract use 'universal', while the method applies only to models that pass the label-free linear-alignability screen. Consider adding a qualifier such as 'for linearly compatible models' in the title or abstract to match the scoped claim.","section":"Title and Section 4"}],"recommendation":"major_revision","confidential_remarks":"The missing ridge-map baseline is the key technical gap; I do not view the compatibility screen as circular, but the paper must separate the bus's contribution from the screen's selection. The NLA sentiment failure on the main onboarded model should be resolved or prominently disclosed. The paper is otherwise a solid empirical study with detailed appendices and a broad set of models; the main claims are defensible if the baseline comparison is added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious look, but it needs one more experiment before the central claim is safe. The selling point is a single frozen shared activation space that models attach to with one linear adapter pair each, after which probes, an SAE dictionary, and a carrier-specific NLA transfer without retraining. That combination is new; prior work builds a pool of jointly aligned models or trains pair-specific connectors. The paper does a lot right: matched-prefix construction across tokenizers is careful, leave-one-out onboarding is the right test, the OOD pool is genuinely independent, and the cross-model execution experiment is a nice functional probe beyond geometric similarity.\n\nThe soft spot is the missing ridge baseline. The compatibility screen already fits closed-form ridge maps from every retained model to a fixed Llama-3.2-3B reference, and those maps hit 98.6–99.6% R@1 on held-out 4,096-way retrieval. The bus is then trained only on models that pass this screen. The paper never checks whether the ridge maps themselves—projecting every model's activations into Llama's native space and running probes/SAE/NLA there—reproduce the tool-transfer numbers in Tables 2, 4, and 5. If they do, the universality is inherited from the screen, not from the bus's learned coordinates or adapter-only onboarding. If the bus clearly beats the ridge baseline, the concern disappears. This is load-bearing and it is absent.\n\nOther issues are minor by comparison. The headline numbers are mostly reported as ranges without error bars; the seed sensitivity in Appendix E only covers the source interface. OLMo sentiment NLA recall collapses to 6% in Table 18, which the paper attributes to carrier routing; plausible, but it shows tool transfer is not uniformly reliable. Code is public but not versioned. The model scope is explicitly limited to models that pass a sharp linear-alignability filter, so 'universal' should be read as 'for these compatible families.'\n\nNet: a solid empirical demonstration of a useful idea, with one missing control that could change the interpretation. Send it to peer review. The referee should require the ridge baseline and confidence intervals on at least the headline retrieval and probe numbers before publication.","headline":"A solid new framework for cross-model activation tools, but the missing ridge baseline leaves the bus's added value unproven.","tokens_in":20302,"tokens_out":3547,"would_cite":false,"duration_ms":33651,"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":"The Universal Activation Bus creates a frozen, shared activation space that compatible language models enter through one linear encoder–decoder adapter pair each, so probes, sparse-autoencoder features, and natural-language interpreters…","keywords":["universal activation bus","activation interface","cross-model alignment","linear adapters","sparse autoencoder","natural language autoencoder","model onboarding","representation alignment"],"falsifier":"Train the full Universal Activation Bus on a pool that includes a candidate the screen rejects (e.g., Phi-3.5-mini, at 1.7% R@1) together with the compatible sources, and measure cross-model retrieval and shared-SAE feature Jaccard on held-out text; if the rejected model aligns and transfers tools, the screen is not load-bearing, and if it collapses, the interface's reach is exactly the screen's reach.","tokens_in":19363,"feed_emoji":"🔌","tokens_out":12377,"duration_ms":87626,"temperature":0.7,"pith_summary":"Activation-based interpretability tools—linear probes, sparse autoencoders, and natural-language autoencoders—are usually tied to one model's hidden space, so they must be rebuilt for each new language model. This paper proposes a Universal Activation Bus: a shared dense space that compatible models enter through one lightweight linear encoder–decoder adapter pair per model. The bus is trained jointly on a small set of source models and then frozen; a new model joins by fitting only its own adapter pair on unlabeled matched text. Across five main models and ten additional checkpoints, semantically related texts form consistent neighborhoods in the shared space, and one probe, one SAE dictionary, and a carrier-specific NLA remain usable across connected models without retraining. If true, the framework replaces pair-specific connectors with a stable, extensible activation contract.","feed_headline":"Frozen activation bus lets new LLMs reuse interpretability tools","feed_subtitle":"One linear adapter pair per model; probes, SAE features, and an NLA transfer with no retraining.","key_machinery":"The central object is the Universal Activation Bus: a shared dense space $Z=\\mathbb{R}^{D}$ with $D=3072$, a shared encoder MLP $\\Phi_{\\rm in}$ and decoder MLP $\\Phi_{\\rm out}$, and per-model linear adapters $A^{\\rm enc}_m:\\mathbb{R}^{d_m}\\to\\mathbb{R}^{D}$ and $A^{\\rm dec}_m:\\mathbb{R}^{D}\\to\\mathbb{R}^{d_m}$. Each model's residual-stream activation is normalized by per-dimension $z$-scoring followed by regularized ZCA whitening, then mapped through the encoder adapter and the shared bus; the decoder inverts the path. Training minimizes a self-reconstruction loss (round-trip fidelity), a cross-reconstruction loss (a state written by model $m$ decoded by model $n$), and a contrastive InfoNCE loss that pulls matched prefixes together in $Z$. The load-bearing identity is the composition $D_n(E_m(h_m(x)))$, which defines a directed translation between any two connected models with no pair-specific connector. Onboarding freezes the shared modules and all source adapters, and fits only the new model's adapter pair against the source consensus representation.","core_discovery":"The central claim is that after source training, the Universal Activation Bus is frozen, and any compatible new model—identified by a label-free ridge-map screen against a fixed reference—can onboard by fitting only its linear adapter pair on unlabeled text. Once connected, the model's shared-space coordinates are stable: a probe trained on pooled source states transfers unchanged, a single TopK sparse autoencoder trained in the shared space yields feature indices with cross-model Jaccard overlap above 0.59, and an NLA trained for a different carrier model produces consistent verbalizations after decoding through the carrier's adapter. The paper further shows that an intermediate activation translated from one model can be inserted into another model's frozen upper layers, which then completes the next-token prediction with 72–88% agreement with the receiving model, exceeding native front–back model agreement in every direction. The result is presented as a model-wise activation contract in which each model connects once and all tools share the same coordinates.","pith_inferences":["The label-free compatibility screen may be doing a large share of the alignment work: the paper excludes models whose ridge-map retrieval falls below 50% R@1 before training the bus, and the reported near-perfect shared-space retrieval comes from this preselected pool; a stronger test would train the bus on an unscreened model mix and see whether compatible and incompatible models still separate.","If the frozen contract holds across scale and checkpoint variants, activation tools become a durable asset that survives model upgrades within a family, so probe re-fitting and SAE retraining could be skipped for new releases.","The cross-model execution result is matched-context state injection, not full-sequence stitching: the receiving model keeps its own context and key-value cache, so full-sequence knowledge transfer across models would need evidence on longer generations and swapped full contexts."],"forward_implications":["A single linear probe fitted on pooled source-model shared states transfers to an onboarded model (OLMo) with macro AUROC 0.959, against native-probe reference 0.963, and nine-way axis identification at 95.4% accuracy without target-side calibration.","One TopK sparse autoencoder trained in the shared space yields Top-32 feature-index Jaccard overlap of 0.59–0.61 across all model pairs, including the unseen onboarded model, while retaining 93.8–98.8% of native-SAE utility.","An NLA trained for Qwen2.5-7B remains informative for other connected models: a nine-way classifier trained only on native-carrier verbalizations reaches 88–94% accuracy on routed sources and 84.9% on the onboarded model, with reconstruction cosine retention above 91%.","Composed front–back execution, where a translated cut-layer state is inserted into another model's frozen upper layers, achieves 72–88% next-token agreement with the receiving model, exceeding native front–back agreement by 7.5–27.2 percentage points in every evaluated direction.","Onboarding saturates at roughly 10K matched calibration positions, and ten additional checkpoints across scales (1.24B–14.8B) and base/instruction-tuned variants all reach minimum ID R@1 of 99.9–100%."],"supporting_citations":[{"why":"supplies the TopK sparse-autoencoder method used to train the single shared dictionary.","marker":"Gao et al. 2025"},{"why":"defines the natural-language autoencoder (NLA) and the released Qwen2.5-7B system that the paper reuses unchanged as a carrier-specific tool.","marker":"Fraser-Taliente et al. 2026"},{"why":"provides the InfoNCE contrastive objective that aligns matched prefixes across models in the shared space.","marker":"van den Oord, Li, and Vinyals 2019"},{"why":"provides the regularized ZCA whitening transform used in each model's frozen normalization.","marker":"Kessy, Lewin, and Strimmer 2018"},{"why":"establishes the model-stitching setting that the cross-model state-execution experiment extends.","marker":"Bansal, Nakkiran, and Barak 2021"},{"why":"shows activation-space interventions can transfer across LLMs and supplies the caution that some model pairs resist alignment.","marker":"Oozeer et al. 2025"},{"why":"provides the mean-difference direction method used to evaluate behavioral geometry in the shared space.","marker":"Rimsky et al. 2024"},{"why":"supplies the Platonic Representation Hypothesis, the background claim that similar geometric structure exists across independently trained models and motivates the shared bus.","marker":"Huh et al. 2024"}],"fun_headline_variants":["Universal bus + one adapter per model: LLMs share interpretability tools","Frozen activation bus makes probes, SAEs, and NLAs model-agnostic","One adapter per model for a common activation space across LLMs","Cross-model activation bus lets tools transfer without retraining","Stable activation contract: new models reuse probes, SAEs, NLAs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the set of models retained by the label-free compatibility screen (R@1 above 50% in ridge retrieval against Llama-3.2-3B) is a meaningful and representative scope for the claimed interface.","fun_headline_variants_meta":{"raw":{"variants":["Universal bus + one adapter per model: LLMs share interpretability tools","Frozen activation bus makes probes, SAEs, and NLAs model-agnostic","One adapter per model for a common activation space across LLMs","Cross-model activation bus lets tools transfer without retraining","Stable activation contract: new models reuse probes, SAEs, NLAs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000694,"raw_usage":{"total_tokens":3132,"prompt_tokens":932,"completion_tokens":2200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2106}},"tokens_in":548,"tokens_out":2200,"duration_ms":13299,"temperature":1.0,"reasoning_tokens":2106,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:43:44.665865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full Universal Activation Bus on a pool that includes a candidate the screen rejects (e.g., Phi-3.5-mini, at 1.7% R@1) together with the compatible sources, and measure cross-model retrieval and shared-SAE feature Jaccard on held-out text; if the rejected model aligns and transfers tools, the screen is not load-bearing, and if it collapses, the interface's reach is exactly the screen's reach.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the TopK sparse-autoencoder method used to train the single shared dictionary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes the model-stitching setting that the cross-model state-execution experiment extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the mean-difference direction method used to evaluate behavioral geometry in the shared space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Platonic Representation Hypothesis, the background claim that similar geometric structure exists across independently trained models and motivates the shared bus."}],"review_version":1}