Causal Structure is Inducible but Functionally Decoupled: The Routing/Readout Boundary of a Typed Mechanism Library

Xining Xun

REVIEW 2 major objections 4 minor 21 references

Causal Structure is Inducible but Functionally Decoupled: The Routing/Readout Boundary of a Typed Mechanism Library

T0 review · 2 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A transformer's induced slot-by-type structure scaffolds routing but stays functionally decoupled from answer readout, so explicit typed state is an index, not the computation.

desk verdict Strong protocol and a genuinely useful moving-null result, but the central decoupling claim may be an architectural artifact; worth refereeing if the authors resolve the wiring question. read the letter →

arxiv 2608.11767 v1 pith:TQ7DHOPJ submitted 2026-08-12 cs.CL

Xining Xun This is my paper · ORCID

classification cs.CL

keywords causalreasoninginterpretabilitymodeleditingtypedmechanismlibraryrouting/readoutboundarypreregistrationmovingnulltransformer

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

This paper asks whether a discrete, human-auditable causal structure inside a language model actually drives the model's answers — the assumption behind circuit-board-style interpretability. Using a typed mechanism library of discrete slots grouped by evidence type inside a transformer, trained on synthetic causal worlds with exact ground truth, it reports that the slot-by-type organization is real and induced by type-level supervision, but that it acts only as a routing index: structural edits move the answer readout by at most $3.4\times10^{-6}$ with exactly zero collateral, at two scales and across three seeds. The paper therefore claims that the circuit-board intuition fails for explicit-state architectures: exposed structure can be auditable and exactly revertible as state while remaining functionally decoupled from answer production. If the result holds, model editing aimed at explicit state must target the routing/readout connection, not the slot contents.

What carries the argument

The central object is the typed mechanism library: $N$ discrete slots ($N=600$ at 125M, $N=200$ at 22.6M) statically partitioned over evidence types such as identity, child, relation, sign, and confidence, with a gating head trained by a type-classification auxiliary loss. The library is the model's only writable causal state, and the backbone consumes library-mediated representations, so the architecture gives the structure every condition to participate in readout. The H-$\alpha$ boundary is measured by paired structural flip edits: changing slot contents changes routing logits but leaves the answer readout unchanged at $|\Delta\hat{y}| \le 3.4\times10^{-6}$. The moving null is identified through permutation-based mutual-information tests with a frozen pipeline gate; the blocks arm, a structural-prior control, sits on the null at 22.6M but carries weak organization at 125M, forcing a paired-permutation attribution estimand.

What would settle it

Train the same typed-library architecture with the answer readout head given the slot codes as explicit concatenated input, keeping everything else identical. If $|\Delta\hat{y}|$ after structural slot edits remains at or below $3.4\times10^{-6}$, the boundary is a wiring property; if it jumps above the preregistered bar, the decoupling is trainable and the paper's claim that this is not an architectural triviality is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is a measured decoupling inside a transformer trained on causal worlds: explicit slot-by-type structure is induced by type-level supervision, faithfully organizes routing, and yet is functionally decoupled from answer readout. On the paper's own criteria, the structural codes are not on the readout pathway: 150/150 structural edits leave the readout unchanged at $|\Delta\hat{y}| \le 3.4\times10^{-6}$ with exactly zero off-path collateral, stable across three seeds and a 5.6x scale window from 22.6M to 125M parameters. The induced organization is statistically attributable to the type supervision signal, measured by permutation-tested mutual information and replicated at 125M under a powered preregistered protocol with all nine criterion cells passing. The structure costs nothing measurable in LM quality, with a gap of at most 0.0082 nats versus a parameter-matched monolith, and the library state is exactly local and bit-exactly revertible under edit, with 250 single edits and 1,000 stacked reverts per seed and zero failures. The paper also reports that the unsupervised null itself moves with scale, so organization claims calibrated at one scale cannot silently be inherited by another. It explicitly makes no behavioral-editability claim: the library is a typed routing index with exact state semantics, not the computation that produces answers.

Load-bearing premise

The sharp routing/readout boundary is load-bearing only if the answer readout genuinely received a path to use the slot codes; if the architecture never wired the readout to consume slot-level features, the measured decoupling is a design choice rather than something training discovered.

Editorial extensions

If this is right

Locate-and-edit methods aimed at explicit causal state should expect structural edits to be behaviorally inert and should target teaching the readout to consume slot codes instead.
Organization claims about internal structure must be calibrated at the same scale where they are evaluated, because a moving null can make a null calibrated at one scale invalid at another.
Auditable state management with bit-exact rollback and zero collateral is achievable for a typed library at zero measured LM cost, providing a substrate for reversible model maintenance.
The circuit-board intuition is false for this explicit-state architecture: exposing structure buys understanding but not control.
The paired-permutation attribution protocol with frozen intersection gates offers a reusable template for attributing induced structure to supervision rather than to unsupervised priors.

Reading between the lines

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

The decoupling may explain why localization does not imply editing efficacy in distributed-weight models: if the optimizer keeps discrete, sparse addressing off the dense readout path as an implicit regularizer, editing addresses or localized weights will not change behavior; the paper hints at this account but labels it speculation.
A testable extension the paper does not run: add a small auxiliary loss forcing the readout to consume slot codes; under the gradient-noise account, training stability should visibly degrade.
The moving-null result suggests mechanistic interpretability claims about emergent specialization need scale-matched controls, not inherited ones, even when the architecture family is unchanged.
If slot codes are high-precision, low-bandwidth addresses, then making the library denser or continuous, for example with soft slots, might blur the routing/readout boundary; this is an inference, not a claim of the paper.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The paper is worth your time, though its headline claim is less bulletproof than the reporting suggests. The methodology is the strongest part: preregistered criteria, a failed gate reported rather than hidden, a powered replication that passed all nine cells, and bit-exact rollback audits. The moving-null finding is a real contribution to how we compare organization claims across scales. The paper also does the right thing by explicitly not claiming behavioral editability.

The new result is the routing/readout boundary: induced slot-by-type structure organizes routing but doesn't move the answer readout, with |Δŷ| ≤ 3.4e-6. If true, that is an important boundary for model editing. The problem is that the paper's own description makes it hard to tell whether the boundary is learned or designed. Figure 1 labels structural codes as NOT on the readout path, and the text restates that as a finding. That could simply mean the readout head was never wired to receive slot-level codes. The assertion that the architecture gave the structure every condition to sit on the readout pathway is not backed by a wiring audit. If the readout input was constructed without slot-code projections, then 150/150 flips cannot move it by construction, and the zero collateral and scale invariance are automatic. The ROME-style row fine-tuning control suggests some pathway exists, but not specifically for the structural codes used in routing.

This ambiguity is load-bearing, because the broader conclusion—that the circuit-board intuition fails for explicit-state architectures—depends on the decoupling being a property of training dynamics, not of wiring. The authors should specify the readout head's input construction and show that slot codes were accessible and the optimizer chose to suppress them. That could be a minor clarification or a major re-interpretation; the paper needs to make it visible.

Everything else is careful and well-executed. The self-citation to Xun 2026 is fine given the continuity. I would send this to a serious referee, but I would push the authors to resolve the wiring question before acceptance. For you: worth a reading-group slot for the protocol design and the moving-null caution, and citeable on that point. I would not cite H-alpha until the wiring is settled.

Referee Report

2 major / 4 minor

Summary. The paper introduces a transformer with a typed mechanism library (discrete slots partitioned by evidence type) and uses it to measure whether induced slot-by-type structure participates in answer readout. On synthetic causal worlds with exact interventional ground truth, at 22.6M and 125M scales and under a heavily preregistered protocol, the authors report four findings: (i) slot-by-type organization is induced by type-level supervision and is not present in unsupervised or content-free-label controls; (ii) the induced structure is a routing index that passes the routing/readout boundary, with structural edits leaving the answer readout essentially unchanged (|Δŷ| ≤ 3.4×10^{-6}, zero collateral); (iii) the library imposes no measurable LM cost (gap ≤ 0.0082 nats); and (iv) the library state is bit-exactly revertible and exactly local under edit. The paper also reports that the unsupervised null itself moves with scale, and it documents one failed preregistered gate and the frozen protocol that handled it.

Significance. If the routing/readout boundary result holds, it is significant for interpretability and model editing: it would show that an explicitly induced, typed structure can serve as an addressable routing index without being on the causal path to the answer, thereby challenging the 'circuit-board' intuition that exposed structure must drive behavior. The paper's methodological contributions are considerable: machine-checkable preregistered criteria, archived decision trees, a powered replication with fresh seeds, and transparent reporting of a failed gate and a moving null. These practices raise the bar for interpretability claims. However, the central boundary claim currently rests on an unresolved wiring question, which substantially tempers the significance as written.

major comments (2)

[§2.4 and Figure 1] The paper's own architecture diagram labels 'structural codes NOT on readout path (H-α, scale-invariant)' as an architectural fact, while §1 and §4.2 claim that 'training decoupled a structure that was built to participate' in readout. These statements are contradictory. If the answer readout head's input is constructed without any projection from the slot-level structural codes, then 150/150 structural edits cannot move the readout by construction, and 'exactly zero collateral' is a wiring consequence rather than a learned decoupling. The manuscript must provide a wiring audit showing that the readout had a viable path to consume slot-level codes, or add a control in which the readout is given direct access to those codes and training still produces decoupling. Without this, H-α is a design choice, not a measured property of training dynamics, and the paper's headline conclusion is unsupported.
[§2.2 and §4.1] The 'origin' finding is partially a restatement of the training objective: the type-supervised arm is trained with λg = 0.1 on the gating head to classify evidence type, so slot-by-type alignment is directly encouraged. The unsupervised controls and paired-attribution tests convincingly show that this alignment is not emergent and not buyable with content-free labels, but the causal phrasing 'induced by type-level supervision' should be more careful. The correct claim is that the alignment is attributable to the type-classification loss beyond what unsupervised routing priors produce, not that the supervision 'induces' a structure that the loss is already explicitly forcing. A sentence clarifying that the MI measurement is taken on a quantity directly optimized by the auxiliary loss would prevent over-reading.

minor comments (4)

[§4.2 and Abstract] The numbers for the maximal readout change after structural edits are inconsistent: the abstract and §4.3 report |Δŷ| ≤ 3.4×10^{-6}, while §4.2 states '≤2×10^{-4}'. If these come from different protocols (raw max vs criterion protocol), the distinction should be stated explicitly.
[§2.1] The abbreviation 'MM' is introduced as 'typed mechanism library (MM)' but never expanded; this is likely a typo for 'ML' or an intended acronym that should be defined at first use.
[Figure 1] The figure label '×structural codes NOT on readout path' is presented as a fixed architectural annotation, which reinforces the wiring concern in the first major comment. Recasting this label as a measured result, or clearly indicating that it is an architectural constraint, would make the paper's claims more consistent.
[§4.1] The moving-null discussion correctly notes that scale and type-count changed together, but the phrase 'the null itself moves with scale' is repeated in the abstract and conclusion without the confound. Consider adding a one-sentence reminder that the scale increase was accompanied by an increase in slot and type counts, so the cause of the null shift is not isolated.

Circularity Check

1 steps flagged · score 3.0 of 10

Origin finding partly restates the type-supervision objective; the central routing/readout boundary is not circular.

self definitional [§2.2 (Typed Routing and Gating); §4.1 (The Moving Null, and the Origin of Typed Organization)]
"type (supervised by the type-classification auxiliary loss λg = 0.1), emergent (λg = 0, free routing), and blocks (λg = 0, block-structured routing prior without type labels). . . . Type-supervised routing produces slot×type organization that replicates across three seeds: per-seed z = 4.47 / 5.99 / 7.02 . . . debiased MI excess 0.072–0.124 nats (mean 0.0998)."

The treatment arm is defined by an auxiliary loss whose target is the evidence type, and the outcome metric is mutual information between slot routing and the same evidence-type labels. Slot-type alignment in the type arm is therefore partly the optimization objective restated as a measured effect, not an independently discovered phenomenon. The paper's unsupervised controls, content-free-label checks, and paired-attribution design provide genuine comparative content, so this is a partial, not a definitional, circularity; it does not extend to the H-alpha boundary or the cost/trust findings.

full rationale

The only step with a definitional component is Finding (i), the origin of typed organization. The type arm differs from controls by adding a type-classification auxiliary loss (λg = 0.1, §2.2), and the paper measures slot×type mutual information using those same type labels (§4.1). High MI in the supervised arm is thus partly a measure of the gating head fitting its own training labels. This would be a serious circularity if the paper stopped there, but it does not: the unsupervised emergent and blocks arms sit on the null at 22.6M, content-free gating labels are unlearnable, and the 125M attribution uses paired permutations against fresh unsupervised controls, so the comparison carries independent empirical content. The central H-alpha claim does not reduce to its inputs. The paper states the architecture gave the structure 'every condition to sit on the readout pathway' because the backbone consumes library-mediated representations; the measured |Δŷ| ≤ 3.4×10^-6 is an empirical outcome, not an equation forcing zero. Figure 1's 'structural codes NOT on readout path' is a label of the measured boundary, not a wiring diagram, since the dashed orange path is described as 'the measured H-α boundary' (§4.2). Whether the readout head actually had access to slot-level codes is a legitimate evidence gap worth a wiring audit, but that is a correctness concern, not a circular derivation. The only self-citation (Xun 2026) appears in Related Work and is not load-bearing. No uniqueness theorem, ansatz-smuggling, or renaming-known-result pattern is present. Overall, the paper's main boundary result is self-contained; the origin result is partially circular but rescued by its control structure, so the appropriate score is low rather than severe.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The quantitative claims rest on a small set of hand-chosen training weights and frozen thresholds, plus domain assumptions about the benchmark and the readout wiring. No new physical entities are postulated; the typed mechanism library is a model component, not an external entity.

free parameters (5)

Type-supervision weight λg = 0.1
Hand-chosen; the Origin claim is measured on models trained with this gating loss weight, and the slot-type organization depends on it.
Per-type floor βfloor = 0.3
Prevents slot-type collapse and shapes the routing; hand-chosen before data.
Load-balancing weight λlb = 0.01
Hand-chosen regularizer that could influence slot-type alignment.
Slot count and type count at each scale = N=200/3 types at 22.6M; N=600/6 types at 125M
The moving-null result is confounded because scale and slot/type counts change together, as the paper admits in §4.1.
Preregistered criterion thresholds = z≥3, excess≥0.10, |Δŷ|≤1e-3, gap≤0.02, battery≥95%
Hand-set frozen bars; every PASS verdict depends on them, and the 22.6M strong-grade miss is 0.0002 below the 0.10 excess bar.

assumptions (4)

standard math Permutation tests yield valid p-values for MI excess under the null of no type-slot association.
Invoked in §3 and §4.1 for all MI claims; relies on exchangeability of permutations (Pesarin and Salmaso 2010).
domain assumption Synthetic causal worlds with exact interventional ground truth are a faithful testbed for how transformers organize causal knowledge.
Stated in §1 and acknowledged as a limitation in §7; the generality of the routing/readout boundary depends on it.
domain assumption The answer readout consumes library-mediated representations, giving slot codes a real opportunity to affect readout.
Entered in §2.4 and Figure 1; if the readout never sees slot-level codes, the H-alpha decoupling is partly by architecture.
domain assumption The debiased raw-pair permutation MI estimator measures the intended slot-type organization.
Described only briefly in §3 and Appendix A; exact estimator and null construction are not fully specified in the main text.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Causal Structure is Inducible but Functionally Decoupled: The Routing/Readout Boundary of a Typed Mechanism Library." pith.science (2026). https://pith.science/paper/TQ7DHOPJ

@misc{pith2026260811767,
  author       = {Pith},
  title        = {Pith review of: Causal Structure is Inducible but Functionally Decoupled: The Routing/Readout Boundary of a Typed Mechanism Library},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQ7DHOPJ}},
  note         = {Machine review of arXiv:2608.11767}
}

abstract

When a language model answers an interventional question, the computation it must perform depends on the type of evidence the query requires. We report a decoupling in how a transformer organizes causal knowledge: slot-by-type structure induced by type-level supervision organizes routing, yet remains functionally decoupled from answer readout. We establish this with a typed mechanism library -- discrete mechanism slots partitioned by evidence type, auditable at the state level -- on a causal-world benchmark with exact interventional ground truth, under a frozen protocol, at two scales (22.6M and 125M). Four preregistered findings. (i) Origin. Slot-by-type organization is induced by type-level supervision: absent in architecturally identical unsupervised controls, not buyable by content-free gating labels, and statistically attributable to the supervision signal, replicating at 125M under a powered preregistered protocol (all nine cells passed). (ii) Boundary. The induced structure is a typed routing index with a sharp routing/readout boundary: slot codes scaffold routing but do not drive answer readout ($|\Delta\hat{y}| \le 3.4\times10^{-6}$, zero collateral, three seeds, stable across a 5.6x scale window) -- we therefore make no behavioral-editability claim. (iii) Cost. The structure is free: LM quality matches a parameter-matched monolith within 0.0082 nats. (iv) Trust. The library state is exactly local under edit and bit-exactly revertible -- 250 single-edit and 1,000 stacked reverts per seed, zero failures. We further find that the unsupervised null itself moves with scale, so comparisons reusing a null calibrated at one scale may be confounded at another. Every claim is tied to a preregistered, machine-checkable criterion archived before the data it governs; the full audit trail, including one criterion we failed and how the frozen protocol handled it, is released as an appendix.

Figures

Figures reproduced from arXiv: 2608.11767 by the authors.

**Figure 2.** Organization at 125M, archived (a,b) and replication (c,d). Type seeds clear both bars in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The pipeline is calibrated and the offending measurement is stable — the basis on which the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Attribution contrasts, archived (a) and replication (b), with the frozen [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Negative results, kept prominent: the library is editable [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Locality and cost. (a) Readout change is [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Descriptive pooled view across collections (no confirmatory pooling). Type-arm excess [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: The preregistered decision tree, as executed (Case A20). Every node is an archived event [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

**Figure 9.** Figure 9: Revert-toolchain audit results. Every preregistered check passes on every seed. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages

[1]

Fine-grained analysis of sentence embeddings using auxiliary prediction tasks

Yossi Adi, Einat Kermany, Yonatan Belinkov, Ofer Lavi, and Yoav Goldberg. Fine-grained analysis of sentence embeddings using auxiliary prediction tasks. InThe Fifth International Conference on Learning Representations (ICLR 2017),

work page 2017
[5]

Categorical reparameterization with Gumbel-Softmax

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In The Fifth International Conference on Learning Representations (ICLR 2017),

work page 2017
[8]

Causal reasoning and large language models: Opening a new frontier for causality.arXiv preprint arXiv:2305.00050,

Emre Kıcıman, Robert Ness, Amit Sharma, and Chenhao Tan. Causal reasoning and large language models: Opening a new frontier for causality.arXiv preprint arXiv:2305.00050,

arXiv
[9]

Zero-shot relation extraction via reading comprehension

Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. Zero-shot relation extraction via reading comprehension. InProceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pp. 333–342,

work page 2017
[10]

Object-centric learning with slot attention

12 Preprint Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. InAdvances in Neural Information Processing Systems 33 (NeurIPS 2020), pp. 11525–11538,

work page 2020
[12]

Locating and editing factual associations in GPT

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. InAdvances in Neural Information Processing Systems 35 (NeurIPS 2022), pp. 17359–17372,

work page 2022
[14]

Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Fast model editing at scale. InThe Tenth International Conference on Learning Representations (ICLR 2022),

work page 2022
[17]

Interpretability in the wild: A circuit for indirect object identification in GPT-2 small

Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: A circuit for indirect object identification in GPT-2 small. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

work page 2023

Show all 21 references

[18]

Evidence-type competition: How typed supervision shapes causal mechanism allocation in language models.arXiv preprint arXiv:2607.29484,

Xining Xun. Evidence-type competition: How typed supervision shapes causal mechanism allocation in language models.arXiv preprint arXiv:2607.29484,

arXiv
[20]

editable

A PREREGISTRATION ARCHIVE(SUMMARY) A-series protocol chain (each entry archived before the data it governs; md5-chained scripts; full text in repository): • A12: pipeline-integrity incident — a leaky aggregation fabricated significance; remediation: raw-pair permutation MI ( m...

2026
[150]

S2/S3 per-seed values appear in §4.3–§4.4; the A20-R replication table is Table 4 (§4.1)

is constructive sampling, not a criterion gap: in three worlds the edge constructed to be absent actually exists, making the edit legal and hence not a failure-semantics instance. S2/S3 per-seed values appear in §4.3–§4.4; the A20-R replication table is Table 4 (§4.1). 0% 25% ...

2026
[1949]

Gomez, Łukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems 30 (NeurIPS 2017),

2017
[2010]

Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In The Fifth International Conference on Learning Representations (ICLR 2017),

2017
[2014]

arXiv:1410.5401. Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-Barwi´nska, Sergio G´omez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, Adri`a Puigdom`enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam ...

arXiv
[2017]

CLadder: Assessing causal reasoning in language models

Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, Ojas Kamal, Zhiheng Lyu, Kevin Blin, Fernando Gonzalez Adauto, Max Kleiman-Weiner, Mrinmaya Sachan, and Bernhard Sch¨olkopf. CLadder: Assessing causal reasoning in language models. InAdvances in Neural Information Processing S...

2023
[2019]

Alex Graves, Greg Wayne, and Ivo Danihelka

arXiv:1904.10922. Alex Graves, Greg Wayne, and Ivo Danihelka. Neural Turing machines,

1904 arXiv
[2020]

Maddison, Andriy Mnih, and Yee Whye Teh

Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. The Concrete distribution: A continuous relaxation of discrete random variables. InThe Fifth International Conference on Learning Representations (ICLR 2017),

2017
[2022]

Mass-editing memory in a Transformer

Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a Transformer. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

2023
[2023]

Can large language models infer causation from correlation? InThe Twelfth International Conference on Learning Representations (ICLR 2024),

Zhijing Jin, Jiarui Liu, Zhiheng Lyu, Spencer Poff, Mrinmaya Sachan, Rada Mihalcea, Mona Diab, and Bernhard Sch¨olkopf. Can large language models infer causation from correlation? InThe Twelfth International Conference on Learning Representations (ICLR 2024),

2024
[2024]

Designing and interpreting probes with control tasks

John Hewitt and Percy Liang. Designing and interpreting probes with control tasks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP 2019), pp. 2733–2743,

2019
[2026]

Dai, Zhifeng Chen, Quoc V

Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M. Dai, Zhifeng Chen, Quoc V . Le, and James Laudon. Mixture-of-experts with expert choice routing. InAdvances in Neural Information Processing Systems 35 (NeurIPS 2022), pp. 7103–7114,

2022

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Fine-grained analysis of sentence embeddings using auxiliary prediction tasks

Yossi Adi, Einat Kermany, Yonatan Belinkov, Ofer Lavi, and Yoav Goldberg. Fine-grained analysis of sentence embeddings using auxiliary prediction tasks. InThe Fifth International Conference on Learning Representations (ICLR 2017),

work page 2017

[2] [5]

Categorical reparameterization with Gumbel-Softmax

Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In The Fifth International Conference on Learning Representations (ICLR 2017),

work page 2017

[3] [8]

Causal reasoning and large language models: Opening a new frontier for causality.arXiv preprint arXiv:2305.00050,

Emre Kıcıman, Robert Ness, Amit Sharma, and Chenhao Tan. Causal reasoning and large language models: Opening a new frontier for causality.arXiv preprint arXiv:2305.00050,

arXiv

[4] [9]

Zero-shot relation extraction via reading comprehension

Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. Zero-shot relation extraction via reading comprehension. InProceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pp. 333–342,

work page 2017

[5] [10]

Object-centric learning with slot attention

12 Preprint Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. InAdvances in Neural Information Processing Systems 33 (NeurIPS 2020), pp. 11525–11538,

work page 2020

[6] [12]

Locating and editing factual associations in GPT

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT. InAdvances in Neural Information Processing Systems 35 (NeurIPS 2022), pp. 17359–17372,

work page 2022

[7] [14]

Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. Fast model editing at scale. InThe Tenth International Conference on Learning Representations (ICLR 2022),

work page 2022

[8] [17]

Interpretability in the wild: A circuit for indirect object identification in GPT-2 small

Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: A circuit for indirect object identification in GPT-2 small. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

work page 2023