Pith. sign in

REVIEW 4 major objections 4 minor 193 references

Scaling Inherently Interpretable Language Models

T0 review · 4 major / 4 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 Scaling interpretable LMs costs a fixed offset, not a growing tax

desk verdict Impressive scaling-law analysis and real engineering, but the human-interpretability claim rests on a thin validation sample and partly circular metrics. read the letter →

arxiv 2608.07594 v1 pith:UJOWAE5M submitted 2026-08-06 cs.CL cs.AI

Guide Labs Team , Andreas Madsen , Aya Abdelsalam Ismail , Giang Nguyen , Isaac Plant , Muawiz Chaudhary , Nathaniel Monson , Saqib Azim
show 2 more authors
Zhichen Guo Julius Adebayo
This is my paper · ORCID
classification cs.CLcs.AI
keywords inherentinterpretabilityconceptbottleneckscalinglawsdiffusionlanguagemodelsattributionmodelsteeringannotationcompute-optimal
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 challenges the assumption that interpretability is a tax on capability: the standard view is that you train an opaque model and then try to explain it afterwards, paying some capability price for the explanation. Instead, the authors build interpretability into pretraining as a design constraint—a concept bottleneck between the transformer backbone and the output head, supervised by a large automatically-built library of human-meaningful concepts—and measure what it costs. Across autoregressive and causal-diffusion backbones, over three orders of magnitude of compute, they find that adding the concept module shifts the compute-optimal scaling exponents by a small, fixed per-backbone offset, so the penalty does not grow with scale. At the same time, the model's concept predictions, known/unknown disentanglement, concept contribution to predictions, and alignment of concept embeddings with human labels all improve as compute increases. The resulting 8B model, Steerling-8B, remains within roughly 10% of open peers trained on 2–16x more compute, supporting the paper's claim that a model can be both interpretable and competitively performant.

What carries the argument

The load-bearing object is the concept module, an additive concept bottleneck inserted between the transformer backbone and the linear language-modeling head. It decomposes the hidden state as h ≈ k̂ + û + ε, with k̂ a weighted sum of supervised 'known' concept embeddings, û a weighted sum of unsupervised 'unknown' embeddings, and ε an explicit residual; because the head is linear, every output logit decomposes exactly into per-concept contributions, making concept attribution and steering closed-form operations. Two supporting mechanisms carry the argument: Causal Diffusion, a block-causal attention pattern (bidirectional within blocks, causal across them) that keeps diffusion-style training while preserving autoregressive KV caching, and the masked-diffusion objective's [MASK] token, which acts as a trained in-distribution absence baseline for input attribution. The third pillar is Atlas, an automated pipeline that turns millions of documents into a canonical library of 33,732 human-meaningful concepts and a chunk-level annotator, providing the semantic supervision the bottleneck needs. The scaling analysis then fits compute-optimal power laws and an irreducible-loss form, showing a fixed per-backbone offset and favorable interpretability scaling.

What would settle it

Two checks would settle the central claim: (1) rerun the human nameability and label-fit study on a random sample of 1,000 Atlas concepts instead of the 100 and 34 used here; if Atlas labels stop beating taxonomy or embedding distractors under blind comparison, the human-meaningfulness premise fails. (2) Extend the IsoFLOP sweep for a +Concept family to 10x the largest compute budget; if the validation-loss gap to the unconstrained backbone widens beyond the fitted fixed offset, or if concept loss and independence loss stop improving, the 'fixed offset, improves with scale' claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that interpretability does not have to be purchased after training at a growing capability cost. Training with an explicit, additive concept bottleneck—where the hidden state is decomposed as a sum of known-concept, unknown-concept, and residual contributions before the language-modeling head—shifts the compute-optimal scaling exponents by a small, fixed per-backbone offset rather than a penalty that compounds with scale. The evidence is an IsoFLOP sweep across four families (autoregressive, causal diffusion, and each with the concept module) over three orders of magnitude of compute, plus a joint Chinchilla-form fit that predicts the validation loss of the deployed 8B model within 0.11 nats. All four interpretability metrics improve with compute: concept loss declines, known and unknown representations become more linearly independent, the fraction of prediction routed through concepts rises, and concept embeddings become better aligned with their human-assigned labels. The authors interpret this as showing that models do not become harder to understand as they scale; they become easier, and the fixed offset means interpretability can be included in the training recipe without giving up the scaling frontier.

Load-bearing premise

The entire interpretability story rests on the Atlas concept library being genuinely meaningful to humans and on the chunk-level annotations being accurate enough to supervise the concept module; if many of the 33,732 concepts are not human-recognizable or the annotator frequently mislabels chunks, the claimed semantic faithfulness collapses.

Editorial extensions

If this is right

  • If adding an interpretable concept bottleneck only shifts scaling exponents by a fixed per-backbone offset, then the compute-optimal frontier itself is preserved; the same scaling-law machinery used to plan opaque models can be used to plan interpretable ones.
  • Because all four interpretability metrics obey predictable power laws (three within tight bounds, one in the favorable direction), interpretability properties can be extrapolated before a large run, much like validation loss.
  • The trained [MASK] baseline and exact additive decomposition make the model's attribution and steering operations closed-form rather than estimated, so diagnosis and correction do not require extra fitted explanation machinery.
  • An inherently interpretable model can compete with opaque peers trained on 2–16x more compute, which directly contradicts the premise that interpretability must be exchanged for capability.

Reading between the lines

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

  • Editorial extension: if the fixed-offset law holds beyond 8B, interpretability-by-design becomes a practical default for safety-critical or regulated deployments, since the one-time capability cost does not compound with scale.
  • Editorial extension: the scaling results depend on the concept library being held fixed; if the concept library itself were scaled or updated during training, the observed offset and exponent shifts could change, suggesting a testable variant where library size is a controlled variable.
  • Editorial extension: the paper's interpretability metrics all measure the concept module's own behavior; an independent test—e.g., steering toward concepts outside the Atlas library and measuring whether generation moves accordingly—would probe whether the claimed alignment generalizes beyond the supervised vocabulary.
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

4 major / 4 minor

Summary. The paper proposes that interpretability can be designed into language-model pretraining as a fixed rather than growing cost, and that the resulting models become more interpretable as they scale. It formalizes inherent interpretability through five conditions, builds the Atlas concept library (33,732 concepts derived from roughly 14 million tags), introduces a causal-diffusion backbone with an additive concept bottleneck, and reports IsoFLOP sweeps over autoregressive, causal-diffusion, and their +Concept variants. Scaling fits are used to predict the validation loss and four interpretability metrics of the 8B Steerling model; the validation-loss prediction is within 0.11 nats. The paper also describes a 1.2T-token pretraining run followed by 150B-token mid-training with an interleaved steering phase, and compares Steerling-8B with open peer models.

Significance. If the central claims held, the paper would be a substantial contribution: it would provide a recipe-level argument that human-understandable concept bottlenecks can survive at pretraining scale, with attribution and steering as native operations rather than post-hoc estimates. The main empirical asset is the careful IsoFLOP methodology and the successful validation-loss extrapolation to Steerling-8B; the exact additive logit decomposition and the honest reporting of pretraining failures in Sections 9.4 and 10.4 are also strengths. However, the interpretability-scaling claim currently rests on metrics that are partly the optimized losses themselves, and the human-meaningfulness evidence covers a tiny fraction of the concept library. If the authors add non-circular metrics or ablations and broaden the human validation, the contribution would be significant; as presented, the magnitude of the claimed result exceeds the evidential support.

major comments (4)
  1. [§7, §8.3 (Eqs. 10, 12, 14)] The claim that 'all interpretability metrics improve with compute' (Section 8.3.2) is weakened by the fact that three of the four metrics are the training losses themselves or direct consequences of loss weights: Concept Loss is Eq. (10), Concept Independence Loss is Eq. (14), and Concept Contribution is pushed toward 1 by the reconstruction loss (Eq. 12) and residual dropout p_epsilon. The scaling fits therefore partly describe optimization of the objective rather than an emergent property of larger models. Known Concept Alignment is the only metric not directly supervised, and it is also the metric closest to the paper's semantic-faithfulness claim. I recommend adding at least one non-loss-based behavioral metric (e.g., the steering benchmark of Section 10.2.4) across the small-scale sweep, or ablating the interpretability losses to show that favorable scaling is not simply inherited from the loss schedule.
  2. [§4.4] Condition 4 of Definition 3.3 (Interpretation) and the abstract's 'human-understandable concepts' are supported by a human study covering only 100 concepts in Phase 1 and 34 concepts in Phase 2 out of 33,732, with the Stage 3 annotator validated on 50 human-rated chunks and an LLM judge at a permissive 'at least 2' threshold (95.5% per-chunk). The paper explicitly states that this 'does not certify every concept.' Since the concept bottleneck is trained with positive-only labels from this annotator, a substantial minority of non-human-meaningful or misannotated concepts could break the semantic faithfulness of the concept attribution interface while leaving all reported numbers unchanged. A larger stratified sample with a coverage bound on the fraction of concepts that fail human recognition, or an explicit restriction of the central claim to the validated subset, is needed.
  3. [§8.2.2, Table 6] The 'small, fixed per-backbone offset' in compute-optimal exponents is not yet established. In the AR pair the exponent shift is +0.093 (0.528 to 0.621) and the bootstrap intervals do not overlap with the baseline; in the CDLM pair the shift is -0.129 (0.632 to 0.503) and the intervals overlap. The sign change across backbones is compatible only with a per-backbone statement, not a universal 'interpretability tax' story. More importantly, L_infinity moves downward in both pairs (1.857 to 1.193; 2.658 to 1.942) with wide confidence intervals (e.g., AR+Concept L_infinity 90% CI [-0.703, +0.394]), so the joint fit's offset is not identified at the small scales used. A paired-difference bootstrap on the implied loss offset at the deployment compute budget would be a more direct test of the claim.
  4. [§8.3, Tables 7–8] The metric extrapolations are presented as 'within tight bounds' without prediction intervals. For the compute-only fits, R-squared ranges from 0.49 to 0.75 and several asymptotes e are pinned to natural bounds (e.g., Concept Loss e = 0.000 with CI [-0.002, +0.000]; Concept Contribution e = 1.000; Known Concept Alignment e = 5.000). From Table 7, CDLM+Concept Concept Contribution is predicted 0.915 versus actual 0.876, and Known Concept Alignment is predicted 3.87 versus actual 3.77; the differences are small, but so are the natural ranges, and a trivial predictor that always outputs the nearest bound would produce comparable errors for some metrics. Please provide prediction intervals for the extrapolation (not only parameter CIs) and report the error of a constant-at-bound baseline.
minor comments (4)
  1. [§1] The sentence 'each concept is reprsentation as a direction' contains a typo; it should read 'represented'.
  2. [§4.4, Table 4] The Phase 2 sample is described as 34 concepts from 8 annotators, but Table 4 reports n = 205 for the Atlas label and n = 402 for human labels; the relationship between these counts and the number of concepts is not explained in the main text and should be clarified.
  3. [§8.2.1, Eq. (28)] The Huber loss is applied to log-residuals with delta = 10^{-3}; this is an unusually small threshold for Huber loss and the choice deserves a one-sentence justification or a sensitivity check.
  4. [§9.4.1] The masking-curriculum description says the moving Gaussian center 'rose from 0.2 to 0.8' while Section 9.2 says the curriculum shifts 'from low to high masking'; the text later refers to 'reaching 0.5 on its climb to its endpoint of 0.8.' Please align these statements and specify whether the center or the mean of the Gaussian is being reported.

Circularity Check

1 steps flagged · score 5.0 of 10

Interpretability-scaling claim is partially circular: three of four 'interpretability metrics' are the training losses themselves, so their improvement with compute is partly true by construction.

  1. self definitional [Section 7 (Interpretability Metrics); Section 8.3.2 (Scaling with compute)]
    "The concept module was trained to detect a set of concepts in context, via the concept loss. The model was trained to route its predictions through those concepts rather than through the residual, via the reconstruction loss. The known and unknown concepts were trained to be disentangled, so they do not encode the same information, via the independence loss. Each of these properties has a corresponding metric, and each metric tests a necessary condition for the architecture to be working as built."

    The three metrics offered as evidence that 'interpretability scales with capability' are, by the paper's own description, the training losses themselves: Concept Loss is L_concept (Eq. 10), Concept Independence Loss is L_indep (Eq. 14), and Concept Contribution is the property targeted by L_rec (Eq. 12), which drives the residual toward zero. All three appear with weights in the final objective (Eq. 15). Consequently, the finding in Section 8.3.2—'Concept Loss and Concept Independence Loss decrease with compute, while Concept Contribution ... increase'—is a restatement that models trained to minimize those objectives achieve lower values of the same objectives on held-out data.

full rationale

The paper's scaling analysis is not circular in its capability component: RQ1 compares compute-optimal validation-loss exponents across backbones with and without the concept module, and the Steerling-8B validation-loss prediction (within 0.11 nats) is an internal extrapolation, not a restatement of fitted inputs. The circularity is partial and confined to the interpretability-scaling claim. Three of the four interpretability metrics are training losses (or direct consequences of a training loss) evaluated on held-out data, so their improvement with compute is largely self-definitional: the optimizer directly minimizes these quantities. The one metric not directly supervised, Known Concept Alignment, is scored by an LLM judge rather than humans, and the human validation of the underlying concept library is limited to a stratified pilot of 100 concepts (Phase 1) and 34 concepts (Phase 2), which the paper itself concedes 'does not certify every concept.' This is a missing-support issue rather than an internal contradiction. Self-citations (Madsen et al. 2024b, Ismail et al. 2024, Andersson et al. 2026) are used for building blocks such as masking-based attribution and the independence loss, but the core scaling measurements are the paper's own experiments, so no load-bearing self-citation chain is present. Overall, the central claim that representations become more interpretable with scale is not fully independent of the training objectives, but the capability results and the single unsupervised alignment metric give it partial independent content; hence a moderate score of 5 rather than a higher one.

Assumptions & free parameters 7 free parameters · 6 assumptions · 2 invented entities

The central claims rest on a chain of modeling choices: the concept library must be meaningful, the additive bottleneck must not destroy capability, the [MASK] baseline must be a valid absence operation, and the scaling forms must apply to the new architecture. Each of these is an assumption with varying support; the paper provides training losses and small-scale validation for several, but none are proven outside the paper's own pipeline.

free parameters (7)
  • Scaling law exponents and asymptotes (L∞, A_P, A_D, α, β for loss; A, β, e for each metric) = e.g., AR αP=0.528, AR+Concept αP=0.621, CDLM+Concept L∞=1.942, per-metric e values in Tables 7-8
    Fitted to small-scale IsoFLOP checkpoints via nonlinear least squares; used to extrapolate Steerling-8B.
  • Number of unknown concepts m = m = 3n (n=33,732 known), rank R=256
    Chosen by ablation; no measurable gain at 5n; factorized to reduce parameters.
  • Teacher forcing floors α_known, α_unknown = 0.5 for both during pretraining; annealed to 0 in mid-training
    Set by sweeping; lower floors hurt known concept alignment for no capability gain.
  • Masking schedule parameters = Moving Gaussian center 0.2 to 0.8 pretrain; uniform in mid-training
    Curriculum chosen by ablation at 1B scale; later found over-aggressive at scale.
  • Atlas clustering hyperparameters = k=80,000; coherence thresholds 9/8/7; merge τ=0.95/0.9/0.85
    Selected by cluster coherence and LLM-judge evaluation; hand-set.
  • Residual dropout p_ε = 0.1 pretrain, 0.3 mid-training
    Higher dropout pushes predictions through concepts; chosen by ablation.
  • Steering strength calibration τ = global τ; per-concept γ = τ/peak(e_c)
    Calibrated so all concepts have same peak logit shift; τ chosen by experiment.
assumptions (6)
  • domain assumption The additive bottleneck h ≈ k^hat + u^hat + ε, with linear LM head, yields an exact and meaningful decomposition of the output logit.
    Architecture enforces this by construction (Eq. 8), but the assumption that the residual ε is small and that known/unknown concepts capture the information relevant to prediction is trained for, not guaranteed. Invoked throughout Section 5.
  • domain assumption Chunk-level positive-only concept labels, aggregated by OR over tokens, provide sufficient supervision to teach token-level concept representations.
    Section 5.4.1: concept loss uses OR-aggregation; assumes a concept present in a chunk can be detected at any masked token, and that no negative labels are needed.
  • domain assumption [MASK] token is a valid and meaningful absence baseline for input attribution.
    Section 6.1.1: integrated gradients from [MASK] assume the path is in-distribution and that 'absence' semantics is captured by the mask; the model was trained on masked inputs, so validity holds by construction, but the semantic meaning of 'absence' is assumed.
  • standard math The ELBO estimator with t~U(10^-3,1-10^-3) is a comparable validation loss for scaling-law fits and differs from the training interval only by a constant.
    Appendix H: they note the validation estimator and training loss sample different t intervals and therefore 'will not match numerically', yet treat the ELBO as a valid loss for power-law fitting. This is a standard practice but an approximation.
  • domain assumption LLM judges provide a valid proxy for human judgment of concept relevance, coherence, and alignment.
    Used throughout Section 4 and for Known Concept Alignment (Section 7). Only a small human study (100 concepts) validates label fit, not the alignment metric itself; the judge is another LLM (Mistral-Small), so there is a risk of LLM-fluent agreement.
  • domain assumption The concept library is fixed and human-meaningful across all 33,732 concepts.
    Section 4.4 validates only 100 concepts; the library is built by LLM clustering and labeling, so the assumption that all concepts are human-meaningful is load-bearing for the interpretability claims.
invented entities (2)
  • Unknown concept embeddings (m approximately 101,196 learned directions)
    purpose: Provide an unsupervised channel that, together with known concepts, reconstructs the hidden state exactly so that the logit decomposition in Eq. 8 is exact and residuals are small.
    No external falsifiable prediction; the only evidence is internal reconstruction and independence losses. The unknown concepts have no human-meaningful labels, so they do not contribute to semantic interpretability.
  • Atlas concept library (33,732 concepts)
    purpose: Fixed vocabulary of human-understandable concepts used to supervise the concept module and to serve as the attribution interface.
    The library is derived from LLM annotations and clustering; only 100 of 33,732 concepts were tested with humans. It is a new constructed resource with partial internal validation only; no external benchmark confirms the full library is human-meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scaling Inherently Interpretable Language Models." pith.science (2026). https://pith.science/paper/UJOWAE5M

@misc{pith2026260807594,
  author       = {Pith},
  title        = {Pith review of: Scaling Inherently Interpretable Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UJOWAE5M}},
  note         = {Machine review of arXiv:2608.07594}
}
read the original abstract

Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective. Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpretability scales with capability rather than against it. Surprisingly, model representations become more disentangled and aligned with human-understandable concepts with scale. We instantiate the training-time recipe with Steerling-8B, a diffusion language model with a causal attention mask. For any group of generated tokens, Steerling-8B attributes the output to relevant input tokens, human-understandable concepts, and training data. This enables closed-loop intervention: diagnose an output through its concept or feature attribution, retrieve similar training data, and correct the behavior through concept steering without retraining. Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

193 extracted references · 9 canonical work pages

  1. [1]

    Alignment Without Retraining: Auditing and Controlling Steerling-8B , year =

  2. [2]

    Second Conference on Language Modeling , year=

    2 OLMo 2 Furious (COLM’s Version) , author=. Second Conference on Language Modeling , year=

  3. [3]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Large Language Diffusion Models , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  4. [4]

    arXiv preprint arXiv:1912.02164 , year=

    Plug and play language models: A simple approach to controlled text generation , author=. arXiv preprint arXiv:1912.02164 , year=

    arXiv 1912

  5. [5]

    Findings of the Association for Computational Linguistics: ACL 2022 , pages=

    Extracting latent steering vectors from pretrained language models , author=. Findings of the Association for Computational Linguistics: ACL 2022 , pages=

    2022

  6. [6]

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

    Nemotron-cc: Transforming common crawl into a refined long-horizon pretraining dataset , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  7. [7]

    arXiv preprint arXiv:1803.05457 , year=

    Think you have solved question answering? try arc, the ai2 reasoning challenge , author=. arXiv preprint arXiv:1803.05457 , year=

    arXiv

  8. [8]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Piqa: Reasoning about physical commonsense in natural language , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

Show all 193 references
  1. [9]

    Communications of the ACM , volume=

    Winogrande: An adversarial winograd schema challenge at scale , author=. Communications of the ACM , volume=. 2021 , publisher=

    2021

  2. [10]

    arXiv preprint arXiv:2203.15556 , volume=

    Training compute-optimal large language models , author=. arXiv preprint arXiv:2203.15556 , volume=

    arXiv

  3. [11]

    arXiv preprint arXiv:2412.01505 , year=

    Scaling law for language models training considering batch size , author=. arXiv preprint arXiv:2412.01505 , year=

    arXiv

  4. [12]

    Findings of the Association for Computational Linguistics: EMNLP 2021 , pages=

    Gedi: Generative discriminator guided sequence generation , author=. Findings of the Association for Computational Linguistics: EMNLP 2021 , pages=

    2021

  5. [13]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    FUDGE: Controlled text generation with future discriminators , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  6. [14]

    arXiv preprint arXiv:2404.06395 , year=

    Minicpm: Unveiling the potential of small language models with scalable training strategies , author=. arXiv preprint arXiv:2404.06395 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2512.10858 , year=

    Scaling behavior of discrete diffusion language models , author=. arXiv preprint arXiv:2512.10858 , year=

  8. [16]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2510.03280 , year=

    Training optimal large diffusion language models , author=. arXiv preprint arXiv:2510.03280 , year=

  10. [18]

    arXiv preprint arXiv:2410.18514 , year=

    Scaling up masked diffusion models on text , author=. arXiv preprint arXiv:2410.18514 , year=

    arXiv

  11. [19]

    arXiv preprint arXiv:2602.15014 , year=

    Scaling Beyond Masked Diffusion Language Models , author=. arXiv preprint arXiv:2602.15014 , year=

  12. [20]

    arXiv preprint arXiv:2401.02954 , year=

    DeepSeek llm: Scaling open-source language models with longtermism , author=. arXiv preprint arXiv:2401.02954 , year=

    arXiv

  13. [21]

    International conference on machine learning , pages=

    Concept bottleneck models , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  14. [22]

    The Thirteenth International Conference on Learning Representations , year=

    Concept bottleneck language models for protein design , author=. The Thirteenth International Conference on Learning Representations , year=

  15. [23]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    Hellaswag: Can a machine really finish your sentence? , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

  16. [24]

    arXiv preprint arXiv:2106.13314 , year=

    Promises and pitfalls of black-box concept learning models , author=. arXiv preprint arXiv:2106.13314 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2411.04430 , year=

    Towards unifying interpretability and control: Evaluation via intervention , author=. arXiv preprint arXiv:2411.04430 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2501.17148 , year=

    Axbench: Steering llms? even simple baselines outperform sparse autoencoders , author=. arXiv preprint arXiv:2501.17148 , year=

    arXiv

  19. [27]

    International Conference on Machine Learning , pages=

    Robust learning with the hilbert-schmidt independence criterion , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  20. [28]

    Proceedings of the 26th annual international conference on machine learning , pages=

    Regression by dependence minimization and its application to causal inference in additive noise models , author=. Proceedings of the 26th annual international conference on machine learning , pages=

  21. [29]

    International conference on machine learning , pages=

    Similarity of neural network representations revisited , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  22. [30]

    The annals of mathematical statistics , volume=

    On information and sufficiency , author=. The annals of mathematical statistics , volume=. 1951 , publisher=

    1951

  23. [31]

    Journal of the Royal statistical society: series B (Methodological) , volume=

    Controlling the false discovery rate: a practical and powerful approach to multiple testing , author=. Journal of the Royal statistical society: series B (Methodological) , volume=. 1995 , publisher=

    1995

  24. [32]

    Journal of Pharmacokinetics and Biopharmaceutics , volume=

    A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability , author=. Journal of Pharmacokinetics and Biopharmaceutics , volume=. 1987 , publisher=

    1987

  25. [33]

    Social psychological and personality science , volume=

    Equivalence tests: A practical primer for t tests, correlations, and meta-analyses , author=. Social psychological and personality science , volume=. 2017 , publisher=

    2017

  26. [34]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    arXiv

  27. [35]

    2024 , url =

    Anthropic , title =. 2024 , url =

    2024

  28. [36]

    arXiv preprint arXiv:2403.08295 , year=

    Gemma: Open models based on gemini research and technology , author=. arXiv preprint arXiv:2403.08295 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2412.19437 , year=

    DeepSeek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    arXiv

  30. [38]

    Improving language understanding by generative pre-training , author=

  31. [39]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    arXiv

  32. [40]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    arXiv

  33. [41]

    arXiv preprint arXiv:2512.13961 , year=

    OLMo 3 , author=. arXiv preprint arXiv:2512.13961 , year=

    arXiv

  34. [42]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  35. [43]

    arXiv preprint arXiv:2406.03736 , year=

    Your absorbing discrete diffusion secretly models the conditional distributions of clean data , author=. arXiv preprint arXiv:2406.03736 , year=

    arXiv

  36. [44]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  37. [45]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  38. [46]

    arXiv preprint arXiv:2506.17298 , year=

    Mercury: Ultra-Fast Language Models Based on Diffusion , author=. arXiv preprint arXiv:2506.17298 , year=

    arXiv

  39. [47]

    arXiv preprint arXiv:2508.02193 , year=

    Seed diffusion: A large-scale diffusion language model with high-speed inference , author=. arXiv preprint arXiv:2508.02193 , year=

    arXiv

  40. [48]

    Findings of the Association for Computational Linguistics: EACL 2026 , pages=

    Enabling autoregressive models to fill in masked tokens , author=. Findings of the Association for Computational Linguistics: EACL 2026 , pages=

    2026

  41. [49]

    arXiv preprint arXiv:2503.09573 , year=

    Block diffusion: Interpolating between autoregressive and diffusion language models , author=. arXiv preprint arXiv:2503.09573 , year=

    arXiv

  42. [50]

    arXiv preprint arXiv:2209.10652 , year=

    Toy models of superposition , author=. arXiv preprint arXiv:2209.10652 , year=

    arXiv

  43. [51]

    Transformer Circuits Thread , volume=

    Towards monosemanticity: Decomposing language models with dictionary learning , author=. Transformer Circuits Thread , volume=

  44. [52]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Backpack language models , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  45. [53]

    2025 , journal=

    Concept bottleneck large language models , author=. 2025 , journal=

    2025

  46. [54]

    Andersson, Alma and Ismail, Aya Abdelsalam and De Brouwer, Edward and Haviv, Doron and Biancalani, Tommaso and Cho, Kyunghyun and Scalia, Gabriele and BenTaieb, Aicha and Corrada Bravo, Hector , booktitle =. sc

  47. [55]

    International Conference on Learning Representations , volume=

    Concept bottleneck generative models , author=. International Conference on Learning Representations , volume=

  48. [56]

    International conference on machine learning , pages=

    Axiomatic attribution for deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  49. [57]

    International Conference on Learning Representations , volume=

    Simple guidance mechanisms for discrete diffusion models , author=. International Conference on Learning Representations , volume=

  50. [58]

    arXiv preprint arXiv:2506.10892 , year=

    The diffusion duality , author=. arXiv preprint arXiv:2506.10892 , year=

  51. [59]

    arXiv preprint arXiv:2506.01928 , year=

    Esoteric language models , author=. arXiv preprint arXiv:2506.01928 , year=

    arXiv

  52. [60]

    arXiv preprint arXiv:2308.10248 , year=

    Steering language models with activation engineering , author=. arXiv preprint arXiv:2308.10248 , year=

    arXiv

  53. [61]

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

    Steering llama 2 via contrastive activation addition , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  54. [62]

    arXiv preprint arXiv:1610.01644 , year=

    Understanding intermediate layers using linear classifier probes , author=. arXiv preprint arXiv:1610.01644 , year=

    arXiv

  55. [63]

    International conference on machine learning , pages=

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav) , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  56. [64]

    arXiv preprint arXiv:2310.01405 , year=

    Representation engineering: A top-down approach to ai transparency , author=. arXiv preprint arXiv:2310.01405 , year=

    arXiv

  57. [65]

    2024 , publisher=

    Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet , author=. 2024 , publisher=

    2024

  58. [66]

    Nature machine intelligence , volume=

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead , author=. Nature machine intelligence , volume=. 2019 , publisher=

    2019

  59. [67]

    arXiv preprint arXiv:1702.08608 , year=

    Towards a rigorous science of interpretable machine learning , author=. arXiv preprint arXiv:1702.08608 , year=

    arXiv

  60. [68]

    arXiv preprint arXiv:2206.07682 , year=

    Emergent abilities of large language models , author=. arXiv preprint arXiv:2206.07682 , year=

    arXiv

  61. [69]

    International Conference on Learning Representations , volume=

    Beyond autoregression: Discrete diffusion for complex reasoning and planning , author=. International Conference on Learning Representations , volume=

  62. [70]

    arXiv preprint arXiv:2311.03658 , year=

    The linear representation hypothesis and the geometry of large language models , author=. arXiv preprint arXiv:2311.03658 , year=

    arXiv

  63. [71]

    A is B” fail to learn “B is A

    The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A” , author=. International Conference on Learning Representations , volume=

  64. [72]

    International Conference on Learning Representations , volume=

    Scaling and evaluating sparse autoencoders , author=. International Conference on Learning Representations , volume=

  65. [73]

    International conference on machine learning , pages=

    Understanding black-box predictions via influence functions , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  66. [74]

    Advances in Neural Information Processing Systems , volume=

    Attention is all you need , author=. Advances in Neural Information Processing Systems , volume=

  67. [75]

    2023 , howpublished =

    Thomas Chaton and Lightning AI , title =. 2023 , howpublished =

    2023

  68. [76]

    GitHub repository , howpublished =

    Nanotron: a minimalistic library for pretraining transformer models , year =. GitHub repository , howpublished =

  69. [77]

    Cryptographers’ track at the RSA conference , pages=

    Ciphers with arbitrary finite domains , author=. Cryptographers’ track at the RSA conference , pages=. 2002 , organization=

    2002

  70. [78]

    Scientific american , volume=

    Cryptography and computer privacy , author=. Scientific american , volume=. 1973 , publisher=

    1973

  71. [79]

    arXiv preprint arXiv:2305.06161 , year=

    Starcoder: may the source be with you! , author=. arXiv preprint arXiv:2305.06161 , year=

    arXiv

  72. [80]

    Luca Soldaini and Kyle Lo , year = 2023, title =

    2023

  73. [81]

    International Conference on Learning Representations , volume=

    Openwebmath: An open dataset of high-quality mathematical web text , author=. International Conference on Learning Representations , volume=

  74. [82]

    arXiv preprint arXiv:2310.10631 , year=

    Llemma: An open language model for mathematics , author=. arXiv preprint arXiv:2310.10631 , year=

    arXiv

  75. [83]

    2026 , month =

    Held, Will , title =. 2026 , month =

    2026

  76. [84]

    doi:10.5281/zenodo.12608602 , url =

    Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and...

    doi

  77. [85]

    arXiv preprint arXiv:2101.00027 , year=

    The pile: An 800gb dataset of diverse text for language modeling , author=. arXiv preprint arXiv:2101.00027 , year=

    arXiv

  78. [86]

    arXiv preprint arXiv:2009.03300 , year=

    Measuring massive multitask language understanding , author=. arXiv preprint arXiv:2009.03300 , year=

    2009 arXiv

  79. [87]

    Nelson and Ku

    Francis, W. Nelson and Ku. Computational Analysis of Present-Day American English , publisher =

  80. [88]

    International Conference on Learning Representations , volume=

    Sparse autoencoders find highly interpretable features in language models , author=. International Conference on Learning Representations , volume=

  81. [89]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Network dissection: Quantifying interpretability of deep visual representations , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  82. [90]

    Advances in Neural Information Processing Systems , volume=

    Pace: Parsimonious concept engineering for large language models , author=. Advances in Neural Information Processing Systems , volume=

  83. [91]

    arXiv preprint arXiv:2304.06129 , year=

    Label-free concept bottleneck models , author=. arXiv preprint arXiv:2304.06129 , year=

    arXiv

  84. [92]

    Advances in Neural Information Processing Systems , volume=

    Datacomp-lm: In search of the next generation of training sets for language models , author=. Advances in Neural Information Processing Systems , volume=

  85. [93]

    arXiv preprint arXiv:2411.06068 , year=

    Zyda-2: a 5 trillion token high-quality dataset , author=. arXiv preprint arXiv:2411.06068 , year=

    arXiv

  86. [94]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    S2ORC: The semantic scholar open research corpus , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  87. [95]

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

    Dolma: An open corpus of three trillion tokens for language model pretraining research , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  88. [96]

    URL https://arxiv

    Training verifiers to solve math word problems, 2021 , author=. URL https://arxiv. org/abs/2110.14168 , volume=

    2021 arXiv

  89. [97]

    The stack: 3 tb of permissively licensed source code, 2022 , author=

    2022

  90. [98]

    International conference on machine learning , pages=

    The flan collection: Designing data and methods for effective instruction tuning , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  91. [99]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 embedding: Advancing text embedding and reranking through foundation models, 2025 , author=. arXiv preprint arXiv:2506.05176 , year=

    2025 arXiv

  92. [100]

    IEEE transactions on big data , volume=

    Billion-scale similarity search with GPUs , author=. IEEE transactions on big data , volume=. 2019 , publisher=

    2019

  93. [101]

    arXiv preprint arXiv:2010.14701 , year=

    Scaling laws for autoregressive generative modeling , author=. arXiv preprint arXiv:2010.14701 , year=

    2010 arXiv

  94. [102]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling laws for neural language models , author=. arXiv preprint arXiv:2001.08361 , year=

    2001 arXiv

  95. [103]

    Journal of computational and applied mathematics , volume=

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis , author=. Journal of computational and applied mathematics , volume=. 1987 , publisher=

    1987

  96. [104]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    A cluster separation measure , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 1979 , publisher=

    1979

  97. [105]

    Journal of statistical mechanics: theory and experiment , volume=

    Fast unfolding of communities in large networks , author=. Journal of statistical mechanics: theory and experiment , volume=

  98. [106]

    arXiv preprint arXiv:1802.03426 , year=

    Umap: Uniform manifold approximation and projection for dimension reduction , author=. arXiv preprint arXiv:1802.03426 , year=

    arXiv

  99. [107]

    Advances in Neural Information Processing Systems , volume=

    Estimating training data influence by tracing gradient descent , author=. Advances in Neural Information Processing Systems , volume=

  100. [108]

    arXiv preprint arXiv:2308.03296 , year=

    Studying large language model generalization with influence functions , author=. arXiv preprint arXiv:2308.03296 , year=

    arXiv

  101. [109]

    Advances in Neural Information Processing Systems , volume=

    If influence functions are the answer, then what is the question? , author=. Advances in Neural Information Processing Systems , volume=

  102. [110]

    RedPajama: an Open Dataset for Training Large Language Models , author =

  103. [111]

    The Twelfth International Conference on Learning Representations , year=

    OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text , author=. The Twelfth International Conference on Learning Representations , year=

  104. [112]

    2023 , eprint=

    Llemma: An Open Language Model For Mathematics , author=. 2023 , eprint=

    2023

  105. [113]

    arXiv preprint arXiv:2605.23080 , year=

    The Attribution Contract: Feature Attribution for Generative Language Models , author=. arXiv preprint arXiv:2605.23080 , year=

    arXiv

  106. [114]

    Attention is not explanation , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  107. [115]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages=

    OLMoTrace: Tracing language model outputs back to trillions of training tokens , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages=

  108. [116]

    Advances in Neural Information Processing Systems , volume=

    The effectiveness of feature attribution methods and its correlation with automatic evaluation scores , author=. Advances in Neural Information Processing Systems , volume=

  109. [117]

    arXiv preprint arXiv:2009.07896 , year=

    Captum: A unified and generic model interpretability library for pytorch , author=. arXiv preprint arXiv:2009.07896 , year=

    2009 arXiv

  110. [118]

    Proceedings of the 39th International Conference on Machine Learning , year=

    Datamodels: Predicting Predictions from Training Data , author=. Proceedings of the 39th International Conference on Machine Learning , year=

  111. [119]

    Proceedings of the 40th International Conference on Machine Learning , pages=

    TRAK: attributing model behavior at scale , author=. Proceedings of the 40th International Conference on Machine Learning , pages=

  112. [120]

    Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    Towards tracing knowledge in language models back to the training data , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , pages=

    2022

  113. [121]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  114. [122]

    Mass-Editing Memory in a Transformer , author=

  115. [123]

    ICLR , year =

    Leask, Patrick and Bussmann, Bart and Pearce, Michael and Bloom, Joseph and Tigges, Curt and Al Moubayed, Noura and Sharkey, Lee and Nanda, Neel , title =. ICLR , year =

  116. [124]

    and Oseledets, Ivan and Tutubalina, Elena , title =

    Korznikov, Anton and Galichin, Andrey and Dontsov, Alexey and Rogov, Oleg Y. and Oseledets, Ivan and Tutubalina, Elena , title =. arXiv preprint arXiv:2602.14111 , year =

  117. [125]

    Sparse Shift Autoencoders for Identifying Concepts from Large Language Model Activations , journal =

    Joshi, Shruti and Dittadi, Andrea and Lachapelle, S\'. Sparse Shift Autoencoders for Identifying Concepts from Large Language Model Activations , journal =

  118. [126]

    arXiv preprint arXiv:2506.15963 , year =

    Cui, Jingyi and Zhang, Qi and Wang, Yifei and Wang, Yisen , title =. arXiv preprint arXiv:2506.15963 , year =

  119. [127]

    arXiv preprint arXiv:2511.11711 , year =

    (authors) , title =. arXiv preprint arXiv:2511.11711 , year =

  120. [128]

    GitHub repository , howpublished =

    Microsoft , title =. GitHub repository , howpublished =. 2019 , publisher =

    2019

  121. [129]

    Advances in neural information processing systems , volume=

    Concept embedding models: Beyond the accuracy-explainability trade-off , author=. Advances in neural information processing systems , volume=

  122. [130]

    Advances in Neural Information Processing Systems , volume=

    Learning to receive help: Intervention-aware concept embedding models , author=. Advances in Neural Information Processing Systems , volume=

  123. [131]

    arXiv preprint arXiv:2205.15480 , year=

    Post-hoc concept bottleneck models , author=. arXiv preprint arXiv:2205.15480 , year=

    arXiv

  124. [132]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Language in a bottle: Language model guided concept bottlenecks for interpretable image classification , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  125. [133]

    arXiv preprint arXiv:2401.14142 , year=

    Energy-based concept bottleneck models: Unifying prediction, concept intervention, and probabilistic interpretations , author=. arXiv preprint arXiv:2401.14142 , year=

    arXiv

  126. [134]

    Advances in Neural Information Processing Systems , volume=

    Glancenets: Interpretable, leak-proof concept-based models , author=. Advances in Neural Information Processing Systems , volume=

  127. [135]

    Advances in Neural Information Processing Systems , volume=

    Relational concept bottleneck models , author=. Advances in Neural Information Processing Systems , volume=

  128. [136]

    International conference on learning representations , volume=

    Causal concept graph models: Beyond causal opacity in deep learning , author=. International conference on learning representations , volume=

  129. [137]

    arXiv preprint arXiv:2506.04877 , year=

    There Was Never a Bottleneck in Concept Bottleneck Models , author=. arXiv preprint arXiv:2506.04877 , year=

  130. [138]

    Advances in Neural Information Processing Systems , volume=

    Bayesian concept bottleneck models with llm priors , author=. Advances in Neural Information Processing Systems , volume=

  131. [139]

    Nature Machine Intelligence , volume=

    Concept whitening for interpretable image recognition , author=. Nature Machine Intelligence , volume=. 2020 , publisher=

    2020

  132. [140]

    URL https://arxiv

    Phi-3 technical report: A highly capable language model locally on your phone, 2024 , author=. URL https://arxiv. org/abs/2404.14219 , volume=

    2024 arXiv

  133. [141]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  134. [142]

    Mistral Small 24B Instruct (Version 2501) , year =

  135. [143]

    2023 , howpublished =

    2023

  136. [144]

    2024 , howpublished =

    Tseng, Alan , title =. 2024 , howpublished =

    2024

  137. [145]

    2021 , url =

    all-mpnet-base-v2 Model Card , author =. 2021 , url =

    2021

  138. [146]

    Advances in neural information processing systems , volume=

    Positive-unlabeled learning with non-negative risk estimator , author=. Advances in neural information processing systems , volume=

  139. [147]

    International conference on algorithmic learning theory , pages=

    PAC learning from positive statistical queries , author=. International conference on algorithmic learning theory , pages=. 1998 , organization=

    1998

  140. [148]

    International conference on algorithmic learning theory , pages=

    Positive and unlabeled examples help learning , author=. International conference on algorithmic learning theory , pages=. 1999 , organization=

    1999

  141. [149]

    Theoretical Computer Science , volume=

    Learning from positive and unlabeled examples , author=. Theoretical Computer Science , volume=. 2005 , publisher=

    2005

  142. [150]

    2026 , eprint=

    Neuron Populations Exhibit Divergent Selectivity with Scale , author=. 2026 , eprint=

    2026

  143. [151]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    arXiv

  144. [152]

    Computational Linguistics , volume=

    Probing classifiers: Promises, shortcomings, and advances , author=. Computational Linguistics , volume=

  145. [153]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume , pages=

    Probing the probing paradigm: Does probing accuracy entail task relevance? , author=. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume , pages=

  146. [154]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    What you can cram into a single \ &!\#* vector: Probing sentence embeddings for linguistic properties , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  147. [155]

    A structural probe for finding syntax in word representations , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  148. [156]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Information-theoretic probing with minimum description length , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  149. [157]

    Transactions of the Association for Computational Linguistics , volume=

    Amnesic probing: Behavioral explanation with amnesic counterfactuals , author=. Transactions of the Association for Computational Linguistics , volume=. 2021 , publisher=

    2021

  150. [158]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Do syntactic probes probe syntax? experiments with jabberwocky probing , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  151. [159]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  152. [160]

    arXiv preprint arXiv:2112.00114 , year=

    Show your work: Scratchpads for intermediate computation with language models , author=. arXiv preprint arXiv:2112.00114 , year=

    arXiv

  153. [161]

    arXiv preprint arXiv:1312.6034 , year=

    Deep inside convolutional networks: Visualising image classification models and saliency maps , author=. arXiv preprint arXiv:1312.6034 , year=

    arXiv

  154. [162]

    arXiv preprint arXiv:1611.07450 , year=

    Grad-CAM: Why did you say that? , author=. arXiv preprint arXiv:1611.07450 , year=

    arXiv

  155. [163]

    arXiv preprint arXiv:2607.00510 , year=

    Prototype Language Models , author=. arXiv preprint arXiv:2607.00510 , year=

    arXiv

  156. [164]

    Advances in neural information processing systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in neural information processing systems , volume=

  157. [165]

    arXiv preprint arXiv:2204.08945 , year=

    Missingness bias in model debugging , author=. arXiv preprint arXiv:2204.08945 , year=

    arXiv

  158. [166]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Are self-explanations from Large Language Models faithful? , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  159. [167]

    arXiv preprint arXiv:2310.07819 , year=

    Faithfulness measurable masked language models , author=. arXiv preprint arXiv:2310.07819 , year=

    arXiv

  160. [168]

    arXiv preprint arXiv:2405.05386 , year=

    Interpretability needs a new paradigm , author=. arXiv preprint arXiv:2405.05386 , year=

    arXiv

  161. [169]

    arXiv preprint arXiv:2505.05410 , year=

    Reasoning models don't always say what they think , author=. arXiv preprint arXiv:2505.05410 , year=

    arXiv

  162. [170]

    arXiv preprint arXiv:2507.11473 , year=

    Chain of thought monitorability: A new and fragile opportunity for ai safety , author=. arXiv preprint arXiv:2507.11473 , year=

    arXiv

  163. [171]

    arXiv preprint arXiv:2512.18792 , year=

    The Dead Salmons of AI Interpretability , author=. arXiv preprint arXiv:2512.18792 , year=

  164. [172]

    Advances in Neural Information Processing Systems , volume=

    Efficient training of low-curvature neural networks , author=. Advances in Neural Information Processing Systems , volume=

  165. [173]

    Advances in neural information processing systems , volume=

    Which models have perceptually-aligned gradients? an explanation via off-manifold robustness , author=. Advances in neural information processing systems , volume=

  166. [174]

    Advances in neural information processing systems , volume=

    Explanations can be manipulated and geometry is to blame , author=. Advances in neural information processing systems , volume=

  167. [175]

    European conference on computer vision , pages=

    Visualizing and understanding convolutional networks , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  168. [176]

    Advances in neural information processing systems , volume=

    A benchmark for interpretability methods in deep neural networks , author=. Advances in neural information processing systems , volume=

  169. [177]

    International conference on machine learning , pages=

    Problems with Shapley-value-based explanations as feature importance measures , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  170. [178]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  171. [179]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    arXiv

  172. [180]

    Journal of Machine Learning Research , volume=

    Underspecification presents challenges for credibility in modern machine learning , author=. Journal of Machine Learning Research , volume=

  173. [181]

    Conference on Uncertainty in Artificial Intelligence , pages=

    On counterfactual explanations under predictive multiplicity , author=. Conference on Uncertainty in Artificial Intelligence , pages=. 2020 , organization=

    2020

  174. [182]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Predictive multiplicity in probabilistic classification , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  175. [183]

    Proceedings of the 2022 ACM conference on fairness, accountability, and transparency , pages=

    Model multiplicity: Opportunities, concerns, and solutions , author=. Proceedings of the 2022 ACM conference on fairness, accountability, and transparency , pages=

    2022

  176. [184]

    arXiv preprint arXiv:2407.04846 , year=

    Amazing things come from having many good models , author=. arXiv preprint arXiv:2407.04846 , year=

    arXiv

  177. [185]

    Advances in Neural Information Processing Systems , volume=

    Implications of model indeterminacy for explanations of automated decisions , author=. Advances in Neural Information Processing Systems , volume=

  178. [186]

    Statistical science , volume=

    Statistical modeling: The two cultures (with comments and a rejoinder by the author) , author=. Statistical science , volume=. 2001 , publisher=

    2001

  179. [187]

    Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    On the existence of simpler machine learning models , author=. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2022

  180. [188]

    arXiv preprint arXiv:1801.01489 , volume=

    All models are wrong but many are useful: variable importance for black-box, proprietary, or misspecified prediction models, using model class reliance , author=. arXiv preprint arXiv:1801.01489 , volume=. 2018 , publisher=

    2018 arXiv

  181. [189]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness? , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  182. [190]

    Computational Linguistics , volume=

    Towards faithful model explanation in nlp: A survey , author=. Computational Linguistics , volume=

  183. [191]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Faithfulness tests for natural language explanations , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  184. [192]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    ERASER: A benchmark to evaluate rationalized NLP models , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  185. [193]

    2026 , eprint =

    The Standard Interpretable Model: A general theory of interpretable machine learning to deductively design interpretable methods using Lagrangian mechanics , author =. 2026 , eprint =

    2026

Pith tools

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