Pith. sign in

REVIEW 2 major objections 5 minor 41 references

Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read In-context conditional rules split into a separable predicate test and a non-separable router.

desk verdict The TEST localization is solid and new; the route negative is a well-hedged but finite-probe result, and the paper's own limitations are admirably honest. read the letter →

arxiv 2608.04183 v1 pith:2MYAF65E submitted 2026-08-04 cs.CL

Luxshan Thavarasa , Sivasuthan Sukumar This is my paper · ORCID
classification cs.CL
keywords activationpatchingin-contextlearningconditionalrulesmechanisticinterpretabilityresidualstreammultilinguallanguagemodelssubspacetransfer
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

The paper asks whether a language model following an in-context rule like "if $P(x)$ then $A$ else $B$" builds a runtime circuit with one module that tests the predicate and another that routes the answer. By patching activations from four donor prompts that make the condition and the answer word disagree, the authors read off, layer by layer, whether a residual-stream site carries the truth value or the answer token. They find a mid-stack band that carries the predicate's truth value and can be overwritten almost perfectly in isolation, across three open models, six languages, and five predicate families. The routing step behaves oppositely: learned subspaces flip the trained label pair perfectly but do not transfer to a new pair, so no separable, transferable router was found at the probed sites. The paper concludes that testing is modular, while routing, under these probes, is not.

What carries the argument

The carrying object is the four-donor disagreement design: four prompts cross rule mapping (standard versus swapped) with query outcome (true versus false), so that in two of them the predicate's truth value and the answer word point to opposite labels. Patching each donor's last-token residual vector into a base run, layer by layer, reveals whether each patched output follows the condition or the answer. A second mechanism, a low-rank subspace probe called DAS-lite (a simplified distributed alignment search) trained on an interchange objective, is used with a transfer control to ask whether the routing step has a separable subspace. The transfer control is what separates a genuine router from a token-bound readout direction.

What would settle it

Run the same transfer test with the learned subspace trained jointly on two or more readout-band sites, with a nonlinear alignment, or at rank above 8; any transfer above chance to an unrelated label pair in any of the three models would overturn the non-separability claim. Conversely, a cross-threshold patch with the answer mapping held fixed but the predicate truth reversed can test whether the mid-stack band carries the evaluated boolean rather than the specific digit: if the answer follows the donor's truth value exactly, the abstract-truth-value reading is confirmed, and if not, it fails.

Watch

Extended reading notes

Core claim

The central discovery is an asymmetry in how an in-context conditional is factored across the residual stream. The predicate test is a causally separable module: at a single mid-stack band, patching in a donor's truth value reroutes the answer with outcome flip near 1.0 and mapping flip near 0.0, meeting a strict isolation criterion in 17 of 18 model-language cells. The same localization holds across five predicate families and a three-output rule. The router does not have this structure: a low-rank learned subspace at each model's readout onset flips the trained label pair with interchange accuracy 1.00, yet transfers to an unrelated label pair at roughly 0 in every model; in the one model probed cross-lingually, it transfers almost perfectly to the same pair in other languages. Under the probes used, the answer-routing direction is bound to the answer tokens themselves, not to an abstract routing role.

Load-bearing premise

The negative router claim assumes that the probe family tried—one residual site at each model's readout onset, linear subspaces of rank up to 8, and greedy head groups—covers every way a router could be implemented; if routing lives in a nonlinear, multi-site, higher-rank, or differently positioned mechanism, every probe would miss it and "route is not modular" would be an artifact of the search.

Editorial extensions

If this is right

  • Predicate-level interventions should be portable: steering or auditing the truth-value band should survive changes of language, predicate family, and even a three-output rule.
  • Answer-level interventions are not portable across vocabulary: a direction that flips one label pair fails on a new pair, so steering built on the answer side will not survive a change of action words.
  • For a fixed label pair, the answer-token direction does transfer across languages, so language change alone is not what blocks routing transfer.
  • In deployed guardrail and tool-routing patterns that instantiate if-then rules, the judgment half is a stable monitoring target located mid-stack before the action is emitted.
  • Without a transfer control, the within-pair interchange accuracy of 1.00 would have certified a spurious router in all three models.

Reading between the lines

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

  • The paper's negative router result is bounded by the probe family it searched: a router that is nonlinear, multi-site, higher-rank, or positioned at other token sites would evade every probe, and the paper itself explicitly limits the conclusion with "under these probes."
  • A natural follow-up is a cross-threshold patch, which the paper notes was not run: holding the answer mapping identical while reversing only the predicate's truth value would confirm whether the mid-stack band carries the evaluated boolean rather than the specific digit.
  • If the test/route asymmetry holds in deployed conditional policies, monitoring a rule's judgment half should be easier than steering its action: a mid-stack truth-value readout could flag judgment/action mismatches before the answer is emitted, whereas changing action vocabulary would require learning a new direction.
  • The Latin-script answer tokens are a deliberate control; repeating the cross-lingual probe with native-script labels would test whether the observed language invariance of the routing direction is due to shared tokens or to a deeper language-agnostic mechanism.
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

2 major / 5 minor

Summary. This paper introduces a four-donor activation-patching design to separate the predicate-testing and answer-routing components of in-context conditional rules in LLMs. Across three open models and six languages, the authors report that a mid-stack residual band carries the predicate's truth value (patching it reroutes the answer with outcome-flip near 1.0 and mapping-flip near 0.0), while a learned low-rank subspace that flips the trained label pair fails to transfer to a new label pair but transfers across languages in Gemma-3-4B. The paper concludes that the predicate test is a separable, localizable module, whereas no separable, transferable router was found under the probes used.

Significance. If the central claims hold, this is a valuable contribution to mechanistic interpretability: it provides a causal decomposition of a runtime-assembled conditional computation, with a novel donor design that separates condition from answer word. The study is unusually well controlled: it includes donor-correctness gating, a label-mass gate, a coherent null, a subspace-illusion guard, a cross-label-pair transfer control, joint two-pair training, and per-family predicate generalization. The multilingual shared-bank design is also a strength. The main limitation is that the router negative is an existential claim based on a finite probe family, and some evidence is uneven across models; these issues weaken but do not destroy the core TEST-localization result.

major comments (2)
  1. [§4.2, Table 2] The router conclusion is an existential negative: "no separable, transferable router" (Table 2 caption) and "The ROUTE, though, shows no separable, transferable handle" (Section 5). The evidence is a finite search over one residual site at each model's readout onset, ranks k∈{1,2,4,8}, a linear orthogonal parametrization, and greedy head groups over the readout band; cross-pair transfer is English-only except for Gemma-3-4B, and cross-lingual transfer was run on only that model. The paper itself concedes in the Limitations that "wider multi-pair banks, other sites, and nonlinear alignments remain unprobed." Because the claimed Test/ROUTE asymmetry depends on this negative, the abstract's unqualified "route is not" and the Table 2 caption overstate what the probes can establish. I recommend either adding a positive control—for example, constructing a synthetic transferable subspace at the same site and showing that the DAS-lite/transfer pipeline recovers it—or consistently restating the claim as "no separable transferable subspace was found under these probes." The final sentence of the abstract already contains the qualifier "under these probes," but it does not appear in the title or in several body statements.
  2. [§4.2, Table 2] The route negative is supported by uneven evidence across models. In Gemma-3-12B, the full-band head patch is nearly null (Φ=0.09), so the transfer negative rests entirely on the DAS-lite subspace method. In Qwen3-8B, the learned subspace has near-zero cosine with the readout direction (0.01–0.09), and there is no independent evidence that the trained direction corresponds to a routing computation rather than a local optimum of the linear interchange objective. The claim that the router is "pair-specific" in Qwen is therefore not fully supported. I would like to see either repeated-seed DAS training with variance reporting, an alternative nonlinear probe, or an explicit downgrade of the cross-model router claim to "not found in Gemma-3-4B and not found by the subspace method alone in the other two models."
minor comments (5)
  1. [Limitations] The paper states "each cell reports a single seed" but uses McNemar p-values and transfer values (e.g., 0.00, ≤0.013) as if they were stable quantities. A repeated-seed or bootstrap analysis of the DAS-lite transfer numbers would strengthen the quantitative zero claims; the current single-seed reporting is a presentation gap even if the effect sizes are large.
  2. [§3.4] The thresholds for the verdict criteria are described as "pre-specified" and "fixed before the grid was run," but there is no external preregistration or timestamp. Please clarify how and when the thresholds were set, or soften the wording to "fixed in advance of the grid runs" without implying external preregistration.
  3. [Table 1] The dagger on the Qwen3-8B Tamil row is informative but slightly confusing: the table reports the O-DONOR-selected layer, and the footnote says the cell passes at the condition peak. Consider adding the condition-peak flip values for that cell (or in an appendix) so the reader can verify the isolation criterion directly.
  4. [Figure 3 caption] The caption says "All patches are applied at the last token of the prompt (the Output: position), which is also the readout site," but the same figure shows an upstream patch at the query-digit position. Please clarify that the readout-site patches are at the last token while the upstream localization patch is at the query-digit token.
  5. [Abstract/introduction] The phrase "route is not" in the title-adjacent framing is catchy but reads as an unconditional existential negative. Since the abstract's final sentence already contains "under these probes," I suggest carrying that qualifier into the earlier abstract sentence "Under every probe we ran, the router direction is token-bound and non-transferable" to avoid overstatement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the TEST/ROUTE asymmetry is supported by held-out transfer controls and pre-registered isolation thresholds; acknowledged probe-coverage limits are not circularity.

full rationale

The paper's central claims are not derived from fitted constants or self-citations. The four-donor patching design is a controlled causal intervention: the TEST layer is selected as the argmax of the O-DONOR flip curve, but the decisive isolation claim requires independent thresholds on MAP-NULL and MAP-DONOR flips at that layer, and 17 of 18 cells passing this pre-registered criterion is an empirical outcome, not a construction. The router analysis trains a DAS subspace on one label pair and tests transfer to held-out pairs and to a jointly trained pair set; the within-pair IIA = 1.00 is explicitly presented as a spurious-certification caution, while non-transfer (about 0 to a new pair, about 0.98 cross-lingually to the same tokens) is evaluated on data not used in fitting. There are no self-citations, and the Limitations section candidly states that other sites, multi-pair banks, and nonlinear alignments remain unprobed, which is a probe-coverage caveat rather than a circular reduction. The only mildly self-referential feature is that reporting the O-DONOR flip at the argmax layer is partly a selection property, but the independent isolation and transfer controls carry the evidential load.

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

The central claims rest on standard interpretability assumptions (residual-stream patching as causal readout, donor design disentangling condition from answer) and on hand-chosen thresholds. No new physical entities are introduced; the TEST band and ROUTE stage are interpretive labels for measured layer behavior. The negative router conclusion additionally assumes the probed subspace family is representative, a limitation the paper acknowledges in its phrase 'under these probes'.

free parameters (4)
  • Isolation thresholds for TEST verdict = MAP-NULL flip >= 0.6 and MAP-DONOR flip <= 0.2
    These pre-specified cutoffs define PASS_TEST_ISOLATED and therefore the headline 17 of 18 cells result. They are hand-chosen, not fitted, but the verdict count is directly determined by them.
  • Label-mass gate = 0.30
    A patch counts as a flip only when the combined probability mass on the two answer labels is at least 0.30. This hand-chosen threshold prevents destructive patches from registering as flips and shapes all flip probabilities.
  • Behavioral gate = 0.85
    A language-model cell enters the analysis only if base accuracy on the shared bank is at least 0.85. This inclusion criterion defines the population over which the localization grid is reported.
  • Readout onset threshold = 0.9
    The router probe is placed at the first layer where the MAP-DONOR flip probability reaches 0.9. The choice of this threshold affects where each model's router analysis is performed.
assumptions (4)
  • domain assumption Residual-stream activation patching at the last-token position is a valid causal intervention that reveals which information a layer carries.
    The entire localization argument depends on copying donor residual vectors and reading flip probabilities. If residual-stream patching confounds or misses distributed effects, the TEST and ROUTE boundary could be an artifact. This is a standard assumption in mechanistic interpretability.
  • domain assumption The four-donor crossing of rule mapping and query outcome fully disentangles the condition from the answer word.
    The inference that a layer follows the condition rather than the answer word assumes MAP-DONOR and MAP-NULL differ only in predicate truth while matching in answer, and vice versa. Any hidden confound in tokenization or surface form would break the logic, especially across six languages.
  • domain assumption The native-speaker checked translations faithfully express the intended predicate in each language.
    Cross-lingual comparability assumes the rule text in Tamil, Sinhala, Chinese, Hindi, and Indonesian carries the same predicate semantics as the English template. For Tamil and Sinhala, which are new to this kind of localization, translation error could change what computation is being patched.
  • ad hoc to paper A single run per cell is representative enough for the McNemar p-values to support the flip probabilities.
    Appendix A states that every experiment reports a single seed per cell. The paper implicitly assumes that one run estimates cell-level behavior well enough, but seed variance is not reported, so near-1.0 flips may overstate stability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages." pith.science (2026). https://pith.science/paper/2MYAF65E

@misc{pith2026260804183,
  author       = {Pith},
  title        = {Pith review of: Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MYAF65E}},
  note         = {Machine review of arXiv:2608.04183}
}
read the original abstract

When a language model follows an in-context conditional rule such as "if P(x) then A else B," does it assemble a runtime circuit with one module that tests the predicate and another that routes the answer? We probe this with activation patching under a four-donor design whose two swapped-rule donors make the condition and the answer word disagree, so each layer reveals which of the two it carries. Across three open models from two families and six languages sharing one fixed item bank, a mid-stack residual band carries the predicate's truth value: patching it reroutes the answer with predicate-outcome flip near 1.0 and mapping flip near 0.0, meeting a strict pre-specified isolation criterion in 17 of 18 cells, and the same localization holds across five predicate families. The router shows the opposite profile. A learned subspace flips A and B near-perfectly within the trained pair yet transfers to a new pair at approximately 0 in every model, while in Gemma-3-4B (the only model probed cross-lingually) it transfers at approximately 0.98 to the same pair in other languages. Under every probe we ran, the router direction is token-bound and non-transferable (largely answer-readout in Gemma, pair-specific in Qwen) rather than an abstract routing module. Test is modular; under these probes, route is not.

Figures

Figures reproduced from arXiv: 2608.04183 by the authors.

Figure 1
Figure 1. The patching operation. The donor’s last [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The four donors, crossing rule mapping (standard vs. swapped) with query outcome (TRUE vs. FALSE). In MAP-DONOR and MAP-NULL the predicate truth and the answer word disagree, exposing per layer whether a patched site follows the condition (TEST) or the answer (ROUTE). turbation whose answer nonetheless equals BASE. The methodological core is that MAP-DONOR and MAP-NULL make the condition and the answer disagree (tru… view at source ↗
Figure 3
Figure 3. The TEST→ROUTE stage boundary in Gemma-3-4B (English). All patches are applied at the last token of the prompt (the Output: position), which is also the readout site. The predicate is first computed at the query-digit position (L14–15), read out as a truth value at the mid-stack TEST band (L21–22, amber), where the O-DONOR reroutes the answer (1.00) but the MAP-DONOR does not (0.00); patches there follow the conditi… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: The router transfer pattern (k=1). In every model the learned subspace flips the trained pair per￾fectly (teal) but transfers to a new pair at ≈0 (red); for Gemma-3-4B the same subspace transfers near-perfectly to the same pair in Tamil and Chinese (light teal): token￾…
Figure 5
Figure 5. Figure 5: English prompts for each predicate family [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 7
Figure 7. Figure 7: Multi-token answers (Gemma-3-4B, En￾glish; sub-word pair and two-word-phrase pair pooled, sequence-margin metrics): per-layer flip for the O￾DONOR (outcome), MAP-NULL (condition) and MAP￾DONOR (answer) donors. The shaded band (same crite￾rion as [PITH_FULL_IMAGE:figur…
Figure 6
Figure 6. Figure 6: Policy-action labels (Gemma-3-4B, English; [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 9
Figure 9. Figure 9: Three-output rule (Gemma-3-4B): per￾layer flip for the O-DONOR (outcome), MAP-NULL (condition) and MAP-DONOR (answer) donors of the dog/cat/cow variant. The shaded band (L18–22, same criterion as [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 8
Figure 8. Figure 8: IOI-contrast head map (Gemma-3-4B): IOI mover heads (black, sized by Φ) never fall in the TEST band (amber); three coincide with our ROUTE heads (teal circles; shared heads ringed red). D The Three-Output Rule The three-output variant (§4.5) uses three numeric buckets …
Figure 10
Figure 10. Figure 10: Greedy head-group patching: fraction-of [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 12
Figure 12. Figure 12: Experimental design. One fixed 250-item bank feeds every cell of the [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Predicate generality (Gemma-3-4B): per￾layer O-DONOR flip for all five predicate families, each reaching 1.0 in the mid-stack band; triangles mark se￾lected TEST layers. Bottom strip: Qwen3-8B spot-check (set-membership, vowel). Max null flip 0.062. 16 [PITH_FULL_IMA…
Figure 14
Figure 14. Figure 14: Two probes of the argument ROUTE step, shown on the primary cell (Gemma-3-4B, English; see [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Gemma-3-4B: per-layer flip probability for all six languages. Left: [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Gemma-3-12B: as above; TEST band L27–28. The MAP-DONOR rise varies by language downstream of the boundary while the O-DONOR curves coincide. Qwen3-8B (36 layers) 0 10 20 30 0 0.5 1 layer flip probability O-donor patch (flips the outcome) en zh hi id ta si 0 10 20 30 0…
Figure 17
Figure 17. Figure 17: Qwen3-8B: as above; TEST band L23–27. The MAP-DONOR rise is displaced several layers past the band: the staggered condition→readout structure. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 23 canonical work pages

  1. [1]

    Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 37:136037--136083

    work page 2024

  2. [2]

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2023. Discovering latent knowledge in language models without supervision. In The Eleventh International Conference on Learning Representations

    2023

  3. [3]

    Danchun Chen, Qiyao Yan, and Liangming Pan. 2026. Towards a mechanistic understanding of propositional logical reasoning in large language models. arXiv preprint arXiv:2601.04260

    work page arXiv 2026

  4. [4]

    Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri \`a Garriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36:16318--16352

    2023

  5. [5]

    Xander Davies, Max Nadeau, Nikhil Prakash, Tamar Rott Shaham, and David Bau. 2023. Discovering variable binding circuitry with desiderata. arXiv preprint arXiv:2307.03637

    arXiv 2023

  6. [6]

    Cl \'e ment Dumas, Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2025. https://doi.org/10.18653/v1/2025.acl-long.1536 Separating tongue from thought: Activation patching reveals language-agnostic concept representations in T ransformers . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol...

    doi:10.18653/v1/2025.acl-long.1536 2025

  7. [7]

    Jiahai Feng and Jacob Steinhardt. 2024. How do language models bind entities in context? In International Conference on Learning Representations, volume 2024, pages 36391--36413

    work page 2024

  8. [8]

    Costa-juss \`a

    Javier Ferrando and Marta R. Costa-juss \`a . 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.591 On the similarity of circuits across languages: a case study on the subject-verb agreement task . In Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 10115--10125

    doi:10.18653/v1/2024.findings-emnlp.591 2024

Show all 41 references
  1. [9]

    Atticus Geiger, Zhengxuan Wu, Christopher Potts, Thomas Icard, and Noah Goodman. 2024. Finding alignments between interpretable causal variables and distributed neural representations. In Causal Learning and Reasoning, pages 160--187. PMLR

    2024

  2. [10]

    Nicholas Goldowsky-Dill, Chris MacLeod, Lucas Sato, and Aryaman Arora. 2023. Localizing model behavior with path patching. arXiv preprint arXiv:2304.05969

    2023 arXiv

  3. [11]

    Yoav Gur-Arieh, Mor Geva, and Atticus Geiger. 2026. Mixing mechanisms: How language models retrieve bound entities in-context. In The Fourteenth International Conference on Learning Representations

    2026

  4. [12]

    Stefan Heimersheim and Neel Nanda. 2024. How to use and interpret activation patching. arXiv preprint arXiv:2404.15255

    2024 arXiv

  5. [13]

    Roee Hendel, Mor Geva, and Amir Globerson. 2023. In-context learning creates task vectors. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9318--9333

    2023

  6. [14]

    Guan Zhe Hong, Nishanth Dikkala, Enming Luo, Cyrus Rashtchian, Xin Wang, and Rina Panigrahy. 2025. A implies B : Circuit analysis in LLM s for propositional logical reasoning. In Advances in Neural Information Processing Systems ( NeurIPS )

    2025

  7. [15]

    Geonhee Kim, Marco Valentino, and Andre Freitas. 2025. https://doi.org/10.18653/v1/2025.findings-acl.525 Reasoning circuits in language models: A mechanistic interpretation of syllogistic inference . In Findings of the Association for Computational Linguistics: ACL 2025 , page...

    2025 doi

  8. [16]

    Bruce W Lee, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Erik Miehling, Pierre Dognin, Manish Nagireddy, and Amit Dhurandhar. 2025. Programming refusal with conditional activation steering. In International conference on learning representations, volume 2025, pages 90960--90985

    2025

  9. [17]

    Samuel Lippl, Thomas McGee, Kimberly Lopez, Ziwen Pan, Pierce Zhang, Salma Ziadi, Oliver Eberle, and Ida Momennejad. 2026. Algorithmic primitives and compositional geometry of reasoning in language models. In Proceedings of the 43rd International Conference on Machine Learning

    2026

  10. [18]

    Aleksandar Makelov, Georg Lange, Atticus Geiger, and Neel Nanda. 2024. Is this the subspace you are looking for? A n interpretability illusion for subspace activation patching. In The Twelfth International Conference on Learning Representations

    2024

  11. [19]

    Samuel Marks and Max Tegmark. 2024. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. In First Conference on Language Modeling

    2024

  12. [20]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in GPT . Advances in neural information processing systems, 35:17359--17372

    2022

  13. [21]

    Jack Merullo, Carsten Eickhoff, and Ellie Pavlick. 2024. Circuit component reuse across tasks in transformer language models. In International Conference on Learning Representations, volume 2024, pages 18349--18377

    2024

  14. [22]

    Mohammed Suhail B Nadaf. 2026. Steerable but not decodable: Function vectors operate beyond the logit lens. arXiv preprint arXiv:2604.02608

    2026 arXiv

  15. [23]

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. 2022. In-context learning and induction heads. arXiv preprint arXiv:2209.11895

    2022 arXiv

  16. [24]

    Gustaw Opie ka, Hannes Rosenbusch, and Claire E Stevenson. 2026. Causality invariance: Function and concept vectors in LLM s. In The Fourteenth International Conference on Learning Representations

    2026

  17. [25]

    Philip Quirke. 2026. Ablation-reversible heads don't transfer: A stress test for mechanistic role claims in T ransformers. arXiv preprint arXiv:2606.08292

    2026 arXiv

  18. [26]

    Arnab Sen Sharma, Giordano Rogers, Natalie Shapira, and David Bau. 2026. LLM s process lists with general filter heads. In The Fourteenth International Conference on Learning Representations

    2026

  19. [27]

    Denis Sutter, Julian Minder, Thomas Hofmann, and Tiago Pimentel. 2025. The non-linear representation dilemma: Is causal abstraction enough for mechanistic interpretability? In Advances in Neural Information Processing Systems ( NeurIPS )

    2025

  20. [28]

    Wei Tang, Xinyan Jiang, Fakhri Karray, and Lijie Hu. 2026. In-context learning operates as concept subspace learning. arXiv preprint arXiv:2605.18830

    2026 arXiv

  21. [29]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, ...

    2025 arXiv

  22. [30]

    Hinata Tezuka and Naoya Inoue. 2025. The transfer neurons hypothesis: An underlying mechanism for language latent space transitions in multilingual LLM s. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 31742--31792

    2025

  23. [31]

    Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron Wallace, and David Bau. 2024. Function vectors in large language models. In International conference on learning representations, volume 2024, pages 17282--17333

    2024

  24. [32]

    Alexandre Variengien and Eric Winsor. 2025. Look before you leap: Universal emergent mechanism for retrieval in language models. In The Thirteenth International Conference on Learning Representations ( ICLR )

    2025

  25. [33]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems, 33:12388--12401

    2020

  26. [34]

    Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2023. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations

    2023

  27. [35]

    Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do L lamas work in E nglish? O n the latent language of multilingual transformers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pag...

    2024

  28. [36]

    Yiwei Wu, Atticus Geiger, and Rapha \"e l Milli \`e re. 2025. How do transformers learn variable binding in symbolic programs? In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 67284--67299. PMLR

    2025

  29. [37]

    Zhengxuan Wu, Atticus Geiger, Thomas Icard, Christopher Potts, and Noah Goodman. 2023. Interpretability at scale: Identifying causal mechanisms in A lpaca. Advances in neural information processing systems, 36:78205--78226

    2023

  30. [38]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. https://arxiv.org/abs/2505.09388 Qw...

    2025 arXiv

  31. [39]

    Qinyuan Ye, Robin Jia, and Xiang Ren. 2025. Function induction and task generalization: An interpretability study with off-by-one addition. arXiv preprint arXiv:2507.09875

    2025

  32. [40]

    Fred Zhang and Neel Nanda. 2024. Towards best practices of activation patching in language models: Metrics and methods. In The Twelfth International Conference on Learning Representations

    2024

  33. [41]

    Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing. 2024. How do large language models handle multilingualism? Advances in Neural Information Processing Systems, 37:15296--15319

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.