Pith. sign in

REVIEW 3 major objections 6 minor 62 references

One Adapter Pair per Model: A Universal Activation Interface for Language Models

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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…

desk verdict A solid new framework for cross-model activation tools, but the missing ridge baseline leaves the bus's added value unproven. read the letter →

arxiv 2608.09521 v1 pith:H6VXU55L submitted 2026-08-10 cs.AI

Su-Hyeon Kim , Jiwan Mun , Yo-Sub Han This is my paper · ORCID
classification cs.AI
keywords universalactivationbusinterfacecross-modelalignmentlinearadapterssparseautoencodernaturallanguagemodelonboardingrepresentation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

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.

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 (3)
  1. [§4; Appendix B.2, Table 13; Section 5] 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.
  2. [Appendix C.3, Table 18; Section 5 (NLA)] 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.
  3. [Tables 1, 2, 4, 5, and 19] 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.
minor comments (6)
  1. [Eq. (7) and Table 12] 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.
  2. [Figure 3] 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.
  3. [Table 3 and Appendix C.2] 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.
  4. [Tables 14, 16, and 19] 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.
  5. [Appendix E.2 and Figure 4b] 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.
  6. [Title and Section 4] 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.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; one mild scope-level self-definition via the compatibility screen, while headline tool-transfer results are genuine held-out/OOD evaluations.

  1. self definitional [Section 4 'Model scope and selection'; Appendix B.2 'Label-Free Compatibility Screening' (Table 13)]
    "Because arbitrary LLM pairs do not always admit high-resolution alignment (Oozeer et al. 2025; Kim and Han 2026), we restrict the main study to models with compatible intermediate representations. Before bus training, a label-free screen fits a closed-form ridge map from each candidate's z-scored activations to a fixed Llama-3.2-3B reference space on matched prefixes. Candidates exceeding 50% R@1 in held-out 4,096-way cosine retrieval are retained."

    The universality claim is scoped to models already selected by a screen that requires near-perfect linear ridge maps to a fixed Llama reference (retained models score 98.6-99.6% held-out R@1; rejected models 0.1-4.7%). Because the bus is then trained and evaluated only on this pre-selected pool, the statement that 'compatible models share an interface' is in part a restatement of the screen's compatibility criterion rather than a property demonstrated independently of it. The paper also never benchmarks the simpler alternative of using the screen's ridge maps into Llama's native space as the shared interface.

full rationale

The central derivation chain is otherwise self-contained. The alignment loss (Eq. 7) directly trains matched prefixes to be close, so high retrieval on training positions is by construction, but Table 1 reports held-out 2% split positions and an independent nine-axis OOD pool; probes, the shared SAE, and the NLA are trained only on source models and applied unchanged to onboarded OLMo, with no target-side fitting. The front-back execution (Section 6) uses no execution-specific training at all. The paper's only self-citations (Kim and Han 2026) are contextual prior-work citations for cross-family alignment and the difficulty of arbitrary pairs; the paper's own screen empirically establishes that difficulty, so the self-citation is not load-bearing. The one mild self-referential element is the compatibility screen, which operationally defines the 'compatible models' scope and therefore makes the universality claim scope-limited rather than arbitrary-model universal; this is an evaluation/scope caveat (a ridge baseline is missing) rather than an equation-level reduction of outputs to inputs. Score 2 reflects that minor scope-level self-definition, not a circular derivation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The bus introduces no new physical entities; its shared space is a learned representation. The main axioms are domain assumptions about matched byte prefixes, compatibility screening, and the sufficiency of a single dense space with linear adapters. Hyperparameters including D=3072, cut depth 0.5, loss weights, and temperature are fixed without full ablation, and the R@1 screen threshold determines the model pool.

free parameters (4)
  • Compatibility screen R@1 threshold = 50%
    Models are retained only if a closed-form ridge map from their z-scored activations to Llama-3.2-3B achieves >50% R@1 on 4,096-way retrieval; this choice defines the compatible models scope and affects all downstream claims.
  • Cut layer depth = 0.5 of model depth (l_m = L_m/2)
    All main models use the residual-stream activation at normalized depth 0.5; the interface is defined at this layer, so results depend on this architectural choice. A sensitivity sweep for OLMo shows mild variation, but other models are not swept.
  • Shared dimension D = 3072
    The bus dimension is chosen by hand; OLMo's native width (4096) exceeds D, and its lower self-FVE is attributed to this compression.
  • Loss weights and temperature = lambda_cross=1.0, lambda_nce=0.1, lambda_align=1.0, tau=0.07
    Training objective hyperparameters; no ablation is reported for their values, so the central results are conditional on this fixed setting.
assumptions (4)
  • domain assumption Matched byte prefixes that end on token boundaries in all source tokenizers provide semantically paired activations across models.
    Stated in Section 3 and Appendix A.3; boundary coverage is measured (99.83-100%), but the semantic equivalence of final-token states after the same byte prefix is assumed, not proven.
  • domain assumption There exists a class of compatible models for which simple linear maps to a Llama reference space indicate alignability, and this class is representative of the intended scope.
    The label-free screen in Section 4 and Appendix B.2 defines compatibility by ridge-map R@1; the framework is only claimed for models passing this screen.
  • domain assumption A single shared dense space with linear adapters and small shared MLPs can simultaneously preserve native information (reconstruction) and align cross-model semantics (contrastive) without collapsing.
    This is the core architectural premise; it is tested empirically on five models, but not proven in general.
  • standard math InfoNCE with in-batch negatives and reconstruction losses are standard, well-founded objectives.
    Uses van den Oord et al. 2019; standard in representation learning, no need to rederive.

how reviews work

0 comments
Cite this review

Pith. "Pith review of One Adapter Pair per Model: A Universal Activation Interface for Language Models." pith.science (2026). https://pith.science/paper/H6VXU55L

@misc{pith2026260809521,
  author       = {Pith},
  title        = {Pith review of: One Adapter Pair per Model: A Universal Activation Interface for Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H6VXU55L}},
  note         = {Machine review of arXiv:2608.09521}
}
read the original abstract

Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model. We present a Universal Activation Bus, a framework that provides a common activation interface across compatible language models. Using a small set of source models, we learn a shared dense space together with one lightweight linear encoder--decoder adapter pair per model. After source training, the interface is frozen; a new model joins by fitting only its adapter pair on unlabeled matched text. The resulting interface allows activation-based tools to be shared across connected models, including common probes and SAE features as well as access to an NLA originally trained for a different model. Across five models, semantically related texts form consistent neighborhoods in the shared space, and an onboarded model reuses these tools effectively without retraining them. We further show that an intermediate activation from one model can be used by another model's frozen upper layers to produce predictions. These results establish a stable, model-wise activation contract for reusable tools across compatible language models.

Figures

Figures reproduced from arXiv: 2608.09521 by the authors.

Figure 1
Figure 1. Overview of the Universal Activation Bus. Source models first learn model-specific encoder–decoder adapter pairs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture and training of the shared activation interface. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Behavioral directions in the shared activation space. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Generality and calibration efficiency of frozen onboarding. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 43 canonical work pages

  1. [1]

    Proceedings of the 36th International Conference on Machine Learning (ICML) , year =

    Kornblith, Simon and Norouzi, Mohammad and Lee, Honglak and Hinton, Geoffrey , title =. Proceedings of the 36th International Conference on Machine Learning (ICML) , year =. 1905.00414 , archivePrefix =

    arXiv 1905

  2. [2]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , year =

    Huh, Minyoung and Cheung, Brian and Wang, Tongzhou and Isola, Phillip , title =. Proceedings of the 41st International Conference on Machine Learning (ICML) , year =. 2405.07987 , archivePrefix =

    arXiv

  3. [3]

    Relative Representations Enable Zero-Shot Latent Space Communication , booktitle =

    Moschella, Luca and Maiorca, Valentino and Fumero, Marco and Norelli, Antonio and Locatello, Francesco and Rodol. Relative Representations Enable Zero-Shot Latent Space Communication , booktitle =. 2023 , eprint =

    work page 2023

  4. [4]

    , title =

    Jha, Rishi and Zhang, Collin and Shmatikov, Vitaly and Morris, John X. , title =. 2025 , eprint =

    work page 2025

  5. [5]

    2026 , eprint =

    Kim, Su-Hyeon and Han, Yo-Sub , title =. 2026 , eprint =

    work page 2026

  6. [7]

    Scaling and Evaluating Sparse Autoencoders , year =

    Gao, Leo and Dupr. Scaling and Evaluating Sparse Autoencoders , year =. 2406.04093 , archivePrefix =

    arXiv

  7. [8]

    and Ameisen, Emmanuel and Chen, James and Kishylau, Dzmitry and Pearce, Adam and Tarng, Julius and Wu, Alex and Wu, Jeff and Zhang, Yang and Ziegler, Daniel M

    Fraser-Taliente, Kit and Kantamneni, Subhash and Ong, Euan and Mossing, Dan and Lu, Christina and Bogdan, Paul C. and Ameisen, Emmanuel and Chen, James and Kishylau, Dzmitry and Pearce, Adam and Tarng, Julius and Wu, Alex and Wu, Jeff and Zhang, Yang and Ziegler, Daniel M. and Hubinger, Evan and Batson, Joshua and Lindsey, Jack and Zimmerman, Samuel and M...

    work page 2026

  8. [9]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

    Thasarathan, Harrish and Forsyth, Julian and Fel, Thomas and Kowal, Matthew and Derpanis, Konstantinos , title =. Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =. 2502.03714 , archivePrefix =

    arXiv

Show all 62 references
  1. [10]

    , title =

    Nasiri-Sarvi, Ali and Rivaz, Hassan and Hosseini, Mahdi S. , title =. 2025 , eprint =

    2025

  2. [11]

    2025 , eprint =

    Puri, Bruno and Berend, Jim and Lapuschkin, Sebastian and Samek, Wojciech , title =. 2025 , eprint =

    2025

  3. [12]

    2026 , eprint =

    Zhao, Haiyan and He, Zirui and Wang, Guanchu and Payani, Ali and Li, Yingcong and Du, Mengnan , title =. 2026 , eprint =

    2026

  4. [13]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Bansal, Yamini and Nakkiran, Preetum and Barak, Boaz , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2106.07682 , archivePrefix =

    arXiv

  5. [14]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Hu, Bodun and Li, Shuozhe and Agarwal, Saurabh and Lee, Myungjin and Jajoo, Akshay and Li, Jiamin and Xu, Le and Kim, Geon-Woo and Kim, Donghyun and Xu, Hong and Zhang, Amy and Akella, Aditya , title =. Proceedings of the 63rd Annual Meeting of the Association for Computationa...

  6. [15]

    2026 , eprint =

    Liu, Xiaoze and Zhang, Ruowang and Yu, Weichen and Xiong, Siheng and He, Liu and Wu, Feijie and Jung, Hoin and Fredrikson, Matt and Wang, Xiaoqian and Gao, Jing , title =. 2026 , eprint =

    2026

  7. [16]

    2025 , eprint =

    Chen, Alan and Merullo, Jack and Stolfo, Alessandro and Pavlick, Ellie , title =. 2025 , eprint =

    2025

  8. [17]

    2026 , eprint =

    Zhang, Peiyan and Xin, Jason , title =. 2026 , eprint =

    2026

  9. [18]

    2025 , eprint =

    Oozeer, Narmeen and Nathawani, Dhruv and Prakash, Nirmalendu and Lan, Michael and Harrasse, Abir and Abdullah, Amirali , title =. 2025 , eprint =

    2025

  10. [19]

    The American Statistician , volume =

    Kessy, Agnan and Lewin, Alex and Strimmer, Korbinian , title =. The American Statistician , volume =. 2018 , doi =

    2018

  11. [20]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Rimsky, Nina and Gabrieli, Nick and Schulz, Julian and Tong, Meg and Hubinger, Evan and Turner, Alexander , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  12. [21]

    2024 , eprint =

    Grattafiori, Aaron and others , title =. 2024 , eprint =

    2024

  13. [22]

    2505.09388 , archivePrefix =

    Qwen3 Technical Report , year =. 2505.09388 , archivePrefix =

    arXiv

  14. [23]

    SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model , year =

    Ben Allal, Loubna and Lozhkov, Anton and Bakouch, Elie and Bl. SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model , year =. 2502.02737 , archivePrefix =

    arXiv

  15. [24]

    2501.00656 , archivePrefix =

    2 OLMo 2 Furious , year =. 2501.00656 , archivePrefix =

    arXiv

  16. [26]

    2019 , howpublished =

    Gokaslan, Aaron and Cohen, Vanya , title =. 2019 , howpublished =

    2019

  17. [27]

    , title =

    Taori, Rohan and Gulrajani, Ishaan and Zhang, Tianyi and Dubois, Yann and Li, Xuechen and Guestrin, Carlos and Liang, Percy and Hashimoto, Tatsunori B. , title =. 2023 , howpublished =

    2023

  18. [28]

    2021 , eprint =

    Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and Hesse, Christopher and Schulman, John , title =. 2021 , eprint =

    2021

  19. [29]

    Competition-Level Code Generation with

    Li, Yujia and Choi, David and Chung, Junyoung and Kushman, Nate and Schrittwieser, Julian and Leblond, R. Competition-Level Code Generation with. Science , volume =. 2022 , doi =

    2022

  20. [30]

    2018 , eprint =

    van den Oord, Aaron and Li, Yazhe and Vinyals, Oriol , title =. 2018 , eprint =

    2018

  21. [31]

    Kornblith, Simon and Norouzi, Mohammad and Lee, Honglak and Hinton, Geoffrey , title =

  22. [32]

    Huh, Minyoung and Cheung, Brian and Wang, Tongzhou and Isola, Phillip , title =

  23. [33]

    Relative Representations Enable Zero-Shot Latent Space Communication , booktitle = ICLR, year =

    Moschella, Luca and Maiorca, Valentino and Fumero, Marco and Norelli, Antonio and Locatello, Francesco and Rodol. Relative Representations Enable Zero-Shot Latent Space Communication , booktitle = ICLR, year =

  24. [34]

    Jha, Rishi and Zhang, Collin and Shmatikov, Vitaly and Morris, John , title =

  25. [35]

    Gao, Leo and Dupre la Tour, Tom and Tillman, Henk and Goh, Gabriel and Troll, Rajan and Radford, Alec and Sutskever, Ilya and Leike, Jan and Wu, Jeffrey , title =

  26. [36]

    2026 , eprint=

    Characterizing Linear Alignment Across Language Models , author=. 2026 , eprint=

    2026

  27. [37]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

    Thasarathan, Harrish and Forsyth, Julian and Fel, Thomas and Kowal, Matthew and Derpanis, Konstantinos , title =. Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

  28. [38]

    , title =

    Nasiri-Sarvi, Ali and Rivaz, Hassan and Hosseini, Mahdi S. , title =. 2026 , journal =

    2026

  29. [39]

    2025 , booktitle = NEURIPS, doi =

    Chen, Alan and Merullo, Jack and Stolfo, Alessandro and Pavlick, Ellie , title =. 2025 , booktitle = NEURIPS, doi =

    2025

  30. [40]

    Oozeer, Narmeen and Nathawani, Dhruv and Prakash, Nirmalendu and Lan, Michael and Harrasse, Abir and Abdullah, Amirali , title =

  31. [41]

    2019 , eprint =

    Aaron van den Oord and Yazhe Li and Oriol Vinyals , title =. 2019 , eprint =

    2019

  32. [42]

    S.; Barsbey, M.; Rodol \`a , E.; and Crisostomi, D

    Achara, A.; Gaintseva, T.; Mahaut, M.; Chakraborty, P.; Johansson, V. S.; Barsbey, M.; Rodol \`a , E.; and Crisostomi, D. 2026. Multi-Way Representation Alignment. arXiv:2602.06205

    2026 arXiv

  33. [43]

    Bansal, Y.; Nakkiran, P.; and Barak, B. 2021. Revisiting Model Stitching to Compare Neural Representations. In Advances in Neural Information Processing Systems (NeurIPS)

    2021

  34. [44]

    Chen, A.; Merullo, J.; Stolfo, A.; and Pavlick, E. 2025. Transferring Linear Features Across Language Models With Model Stitching. In Advances in Neural Information Processing Systems (NeurIPS)

    2025

  35. [45]

    C.; Ameisen, E.; Chen, J.; Kishylau, D.; Pearce, A.; Tarng, J.; Wu, A.; Wu, J.; Zhang, Y.; Ziegler, D

    Fraser-Taliente, K.; Kantamneni, S.; Ong, E.; Mossing, D.; Lu, C.; Bogdan, P. C.; Ameisen, E.; Chen, J.; Kishylau, D.; Pearce, A.; Tarng, J.; Wu, A.; Wu, J.; Zhang, Y.; Ziegler, D. M.; Hubinger, E.; Batson, J.; Lindsey, J.; Zimmerman, S.; and Marks, S. 2026. Natural Language A...

    2026

  36. [46]

    Gao, L.; Dupre la Tour, T.; Tillman, H.; Goh, G.; Troll, R.; Radford, A.; Sutskever, I.; Leike, J.; and Wu, J. 2025. Scaling and Evaluating Sparse Autoencoders. In International Conference on Learning Representations (ICLR)

    2025

  37. [47]

    Gorbett, M.; and Jana, S. 2026. Characterizing Linear Alignment Across Language Models. arXiv:2603.18908

    2026 arXiv

  38. [48]

    Hu, B.; Li, S.; Agarwal, S.; Lee, M.; Jajoo, A.; Li, J.; Xu, L.; Kim, G.-W.; Kim, D.; Xu, H.; Zhang, A.; and Akella, A. 2025. StitchLLM: Serving LLMs, One Block at a Time. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2025

  39. [49]

    Huh, M.; Cheung, B.; Wang, T.; and Isola, P. 2024. Position: The Platonic Representation Hypothesis. In Proceedings of the International Conference on Machine Learning (ICML)

    2024

  40. [50]

    Jha, R.; Zhang, C.; Shmatikov, V.; and Morris, J. 2025. Harnessing the Universal Geometry of Embeddings. In Advances in Neural Information Processing Systems (NeurIPS)

    2025

  41. [51]

    Kessy, A.; Lewin, A.; and Strimmer, K. 2018. Optimal Whitening and Decorrelation. The American Statistician, 72(4): 309--314

    2018

  42. [52]

    Kim, S.-H.; and Han, Y.-S. 2026. Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations. arXiv:2605.09875

    2026 arXiv

  43. [53]

    Kornblith, S.; Norouzi, M.; Lee, H.; and Hinton, G. 2019. Similarity of Neural Network Representations Revisited. In Proceedings of the International Conference on Machine Learning (ICML)

    2019

  44. [54]

    Liu, X.; Zhang, R.; Yu, W.; Xiong, S.; He, L.; Wu, F.; Jung, H.; Fredrikson, M.; Wang, X.; and Gao, J. 2026. The Vision Wormhole: Latent-Space Communication in Heterogeneous Multi-Agent Systems. arXiv:2602.15382

    2026 arXiv

  45. [55]

    Moschella, L.; Maiorca, V.; Fumero, M.; Norelli, A.; Locatello, F.; and Rodol \`a , E. 2023. Relative Representations Enable Zero-Shot Latent Space Communication. In International Conference on Learning Representations (ICLR)

    2023

  46. [56]

    Nasiri-Sarvi, A.; Rivaz, H.; and Hosseini, M. S. 2026. SPARC : Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability. Transactions on Machine Learning Research

    2026

  47. [57]

    Oozeer, N.; Nathawani, D.; Prakash, N.; Lan, M.; Harrasse, A.; and Abdullah, A. 2025. Activation Space Interventions Can Be Transferred Between Large Language Models. In Proceedings of the International Conference on Machine Learning (ICML)

    2025

  48. [58]

    Puri, B.; Berend, J.; Lapuschkin, S.; and Samek, W. 2025. Atlas-Alignment: Making Interpretability Transferable Across Language Models. arXiv:2510.27413

    2025 arXiv

  49. [59]

    Qwen Team . 2024. Qwen2.5 Technical Report. arXiv:2412.15115

    2024 arXiv

  50. [60]

    Rimsky, N.; Gabrieli, N.; Schulz, J.; Tong, M.; Hubinger, E.; and Turner, A. 2024. Steering Llama 2 via Contrastive Activation Addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 15504--15522

    2024

  51. [61]

    Thasarathan, H.; Forsyth, J.; Fel, T.; Kowal, M.; and Derpanis, K. 2025. Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment. In Proceedings of the 42nd International Conference on Machine Learning (ICML)

    2025

  52. [62]

    van den Oord, A.; Li, Y.; and Vinyals, O. 2019. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748

    2019 arXiv

  53. [63]

    Zhang, P.; and Xin, J. 2026. A Negative Result on Cross-Model Activation Transfer in a Pythia Multi-Hop Setting. arXiv:2606.03280

    2026 arXiv

  54. [64]

    Zhao, H.; He, Z.; Wang, G.; Payani, A.; Li, Y.; and Du, M. 2026. Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation. arXiv:2605.25903

    2026 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.