Pith. sign in

REVIEW 3 major objections 4 minor 151 references

Interpreting Language Model Hidden States at Scale

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read OmniLens claims that one low-rank, identity-residual lens family can attach to any model-width hookpoint, making dense trained-lens coverage practical on models as large as LLaMA-3-70B.

desk verdict A solid scaling contribution whose headline interpretability claim needs better statistics before it is treated as established. read the letter →

arxiv 2608.10260 v1 pith:JEXUFH6M submitted 2026-08-10 cs.AI

classification cs.AI
keywords trainedlenseslow-ranktranslatorsSubset-KLLLMinterpretabilityhookpoint-agnosticdenselenscoveragevisibility-actiondecouplingLLaMA-3-70B
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

OmniLens claims that one lens family, an identity-based low-rank translator combined with vocabulary-subset KL training, can attach to any model-width hookpoint, making dense trained-lens coverage practical on models as large as LLaMA-3-70B. The paper demonstrates 482 lenses at 70B with 90.5% fewer trainable parameters than a full-rank residual-only stack, and up to 70% lower peak memory on GPT-2. It reproduces three published interpretability studies, prompt-injection detection, multi-hop memory injection, and toxicity localization, at substantially lower cost. Model-wide coverage yields a new empirical constraint: the hookpoint where a behavior is most visible is often not the hookpoint where intervening changes the behavior most, and the best intervention sites lie outside attention heads. If right, trained-lens interpretability stops being a sparse, component-specific tool and becomes a model-wide instrument for near-frontier models.

What carries the argument

The load-bearing mechanism is the identity-residual low-rank translator $L_{\ell,u}=I+\frac{\alpha}{r}B_{\ell,u}A_{\ell,u}$, which turns each hookpoint's activation into residual-stream space before the frozen final normalization and unembedding; because it starts as the identity and only needs to agree on activations the model actually produces, rank $r=64$ captures most of a full-rank lens's fidelity at a fraction of the parameters. The second mechanism is Subset-KL training: Top-$k$ truncation renormalizes both distributions over the teacher's $k$ most probable tokens, while Top-$k$+IS computes the head exactly and importance-samples the tail with a lens-independent proposal, so the estimated objective and its gradients are unbiased for the full KL (Theorem 1). A fused CUDA kernel computes only the selected logits without materializing the gathered unembedding rows, removing the memory blow-up that subset objectives would otherwise reintroduce. Together these two mechanisms make per-hookpoint parameter growth linear in width and keep vocabulary-side memory under control, which is what allows 482 lenses on a 70B model.

What would settle it

At 8B, train full-rank full-KL lenses on all six hookpoint types and rerun the toxicity audit; if MLP outputs no longer give the largest toxicity reduction or the negative Spearman correlation between detection and intervention disappears, the visibility-action decoupling is an artifact of the low-rank or subset-KL approximation.

Watch

Extended reading notes

Core claim

OmniLens establishes that a single low-rank lens family can decode any model-width activation—residual, attention, or MLP—and that with this universal family dense coverage is tractable at 70B scale. The translator is an identity-residual low-rank map $L_{\ell,u}=I+\frac{\alpha}{r}BA$, trained against the model's frozen final readout, with parameters growing linearly rather than quadratically in width. Two Subset-KL objectives supply the memory lever: Top-$k$ truncation renormalizes over the teacher's most probable tokens, and Top-$k$+IS keeps those head tokens exact while importance-sampling the tail, giving unbiased stochastic gradients for the full KL. On a 482-lens LLaMA-3-70B ensemble the paper reproduces published prompt-injection detection, multi-hop memory injection, and toxicity localization results at lower cost. The new finding that model-wide coverage exposes is that visibility and causal actionability decouple: the components where toxic language is most detectable are negatively correlated (Spearman $-0.43$ at 8B) with the components where intervention reduces toxicity, and the most effective intervention targets lie outside the attention heads studied previously.

Load-bearing premise

The paper's central new finding—that visible hookpoints differ from effective intervention sites—rests on the assumption that its rank-64 and subset-vocabulary approximations preserve the relative ordering of hookpoints; the paper's own appendix shows the approximations resolve predictions several hidden states later than the full-KL reference, so if the approximations shift rankings, the decoupling could be an artifact.

Editorial extensions

If this is right

  • Trained-lens interpretability scales to near-frontier models: a 482-hookpoint ensemble trains on LLaMA-3-70B within A100-40GB budgets, and eight optimizer steps execute on LLaMA-3-405B.
  • Component-agnostic coverage makes cross-component comparison routine: the same lens family reads residual, attention, and MLP activations, so rankings of visibility and intervention efficacy can be compared without retraining separate lens families.
  • Visibility and actionability can diverge: at 8B, the components with the strongest detected toxicity signal and the components whose modification most reduces toxicity are negatively correlated (Spearman $-0.43$), so localization studies that rely on readout visibility alone may pick the wrong intervention site.
  • Established lens applications survive the approximations: prompt-injection detection reaches mean AUROC 0.997 on the five classification tasks at every scale, memory injection captures 61–82% of achievable gain at $\tau=4$, and toxicity localization reproduces the published head concentration.
  • Full-rank references are not needed for task-level utility: at 70B, where no full-rank reference is trainable, the Subset-KL lenses detect knowledge-task injections the logit lens misses.

Reading between the lines

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

  • Beyond the paper, a user of the released framework can test the decoupling directly on any 8B model: select intervention sites by measured causal effect rather than by lens visibility and compare; if visibility-selected sites continue to underperform, the paper's constraint becomes a general design rule for interpretability tools.
  • Beyond the paper, the rank-64 default makes early-layer analyses the fragile case; the paper's own ablations show layer-0 top-1 agreement of only 56.7%, so component or layer rankings in the first few layers should be read with wider error bars than late-layer rankings.
  • Beyond the paper, the identity-residual translator plus a learned input map could extend the same family to individual attention heads, which the paper lists as future work; if that holds, per-head interpretability at 70B becomes a matter of engineering rather than method design.
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.

Referee Report

3 major / 4 minor

Summary. The paper introduces OmniLens, a trained-lens framework that attaches one low-rank, identity-residual translator family to residual-stream, attention, and MLP hookpoints, and trains it with two Subset-KL objectives: a biased Top-k truncation and an unbiased Top-k+IS importance-sampled estimator. The authors report large reductions in trainable parameters and peak memory, train a 482-lens ensemble on LLaMA-3-70B, run eight optimization steps on LLaMA-3.1-405B as a feasibility demonstration, and reproduce three lens applications: prompt-injection detection, multi-hop memory injection, and toxicity localization. The paper's central scientific claim beyond scaling is that model-wide lens coverage reveals a decoupling between where a behavior is most visible in lens readouts and where intervention is most causally effective, with a reported Spearman correlation of -0.43 between detected toxicity and intervention reduction across six component types at 8B.

Significance. If the scaling results hold, OmniLens is a substantial systems contribution: it combines two independent cost reductions (low-rank translators and subset-KL training), ships reproducible code, gives a correct unbiasedness proof for the Top-k+IS estimator in Appendix E, and validates its memory model against measured GPT-2 and 70B runs. Enabling dense trained-lens coverage at 70B and a training path at 405B is genuinely valuable for the interpretability community. The decoupling claim, if confirmed with proper statistical support, would be an important empirical constraint on how lens readouts are used, because it would show that visibility rankings do not generally imply intervention rankings. However, as detailed below, the evidence for that claim is currently not sufficient: it rests on six component-type aggregates with no uncertainty quantification and on approximation choices whose effect on the relevant rankings is not verified against a full-rank componentwise reference.

major comments (3)
  1. [§5, Fig. 6d] The headline decoupling result, Spearman -0.43 between detected toxicity and intervention effectiveness at 8B, is computed from six component-type aggregates with no confidence interval, no permutation test, and no cross-seed spread on the intervention axis. The paper itself notes in Appendix J that the selector comparison in Table 22 is a single-run evaluation and that differences of a few points are within run-to-run noise. As presented, the correlation cannot be distinguished from sampling noise. Please report bootstrap confidence intervals over lens seeds and prompts, and a permutation test over component labels, or explicitly demote the claim to a qualitative observation.
  2. [Appendix K, Table 7] The decoupling claim requires that the low-rank/Subset-KL approximation preserve the relative ordering of hookpoints by visibility and by intervention efficacy. The paper's own fidelity statistics do not establish this. Table 7 reports early-layer Kendall tau@100 of only 0.243 for the recommended Top-k+IS lens against the full-rank reference (and -0.077 for Top-k), and Appendix K states that Subset-KL lenses resolve predictions 4-5 hidden states later than the full-KL reference. Since toxicity detection is a fine-grained, partly early-layer signal, this is a real risk that the component rankings in Fig. 6d are artifacts of the approximation. A concrete remedy is to compute the same detection-vs-intervention Spearman at GPT-2 with the full-rank full-KL reference, which is trainable at that scale, and show that the low-rank and full-rank component orderings agree; at 8B no full-rank componentwise reference exists, so that check is essential rather than optional.
  3. [Abstract, §5, Appendix J] The abstract states that 'the most effective interventions lie outside the attention heads examined by prior lens studies' without qualification, but the supporting evidence is limited to GPT-2 and 8B. At 70B, Appendix J reports that ablating the flagged heads is no more effective than ablating random heads (Fig. 30), so the outside-attention conclusion has no 70B support; the 8B selector comparison in Table 22 is single-run. Please qualify the abstract and Section 6 claims to 'at GPT-2 and 8B', or provide an intervention result at 70B that identifies an effective non-attention site.
minor comments (4)
  1. [Fig. 6d] The six component-type points in the main-text figure are not directly labeled in the rendered figure; please add explicit labels or a legend so the reader can map each point to attn_in, attn_out, resid_mid, mlp_in, mlp_out, and resid_post without cross-referencing the text.
  2. [Appendix L] Appendix L states that the detection captures of Table 16 and the whole-model audit predate the 250-step cosine-annealed schedule and use earlier-schedule checkpoints. Please make this schedule mismatch visible in the main text, since the three case studies are presented as a single coherent evaluation and the inconsistency is currently easy to miss.
  3. [Fig. 6d caption] The caption's statement 'No full-rank reference is trainable at 70B (Section 4)' is ambiguous: Appendix G shows that LoRA+Full-KL fits at 35.5GB, so the missing reference is a full-rank translator, not a full-KL objective. Please rephrase to 'no full-rank translator reference is trainable at 70B'.
  4. [Appendix E] In the proof of Theorem 1, the display for the tail expectation omits the explicit average over the k_tail draws; the surrounding text explains it, but the equation would be clearer if the 1/k_tail sum appeared explicitly before taking the expectation.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the scaling and estimator results are self-contained, and the self-citations in the case studies are replication baselines with external outcomes.

full rationale

The paper's central derivations are not circular. The low-rank translator in Eq. (3) is an explicit parameterization with O(rd) parameters, so the parameter-count scaling is arithmetic rather than a fitted prediction. The Subset-KL estimator's unbiasedness (Theorem 1) is a standard importance-sampling identity, proved in Appendix E and numerically verified against full-KL autograd gradients; it does not assume the conclusion. Fidelity claims for low-rank and Subset-KL lenses are checked against full-rank full-KL baselines on GPT-2 and 8B, and against teacher KL, so the approximations are evaluated against independent references rather than being defined in terms of their own outputs. The three case studies use external task outcomes (detection AUROC on held-out attacks, 2WMH causal lift in the model's own final P(answer), and Toxic-BERT scores on generations) that were not used to fit any lens constant; consequently, no fitted parameter is recycled as a 'prediction.' The visibility-vs-intervention decoupling is a new empirical claim whose robustness is limited by low-rank/Subset-KL fidelity at early layers (Appendix K, Table 7: early-layer Kendall tau 0.243 for Top-k+IS vs. the full-rank reference; 4-5 hidden-state later prediction depth), but this is an approximation-validity concern, not circularity: the intervention measurements are external and the lens was not optimized to produce the decoupling. Self-citations to Pettyjohn (2025) and Sakarvadia et al. (2023, 2024) appear as described prior work and replication baselines; they are not load-bearing because the current paper re-measures those effects on external benchmarks (2WMH, Toxic-BERT, prompt-injection AUROC) rather than relying on the citations for support. These non-load-bearing self-citations justify a score of 2, not 0, but no circular step is present.

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

The central scaling claims rest on a small set of standard domain assumptions plus a few hand-set hyperparameters (rank, subset budget, alpha). The memory-model coefficient is calibrated from measurements but is auxiliary. No new physical or architectural entities are postulated.

free parameters (4)
  • lens rank r = r = 64
    Default rank selected from GPT-2 ablation sweep (Appendix C.1); at r=64 the lens uses 16.7% of full-rank parameters with 88.8% final-layer top-1 agreement. The same r is used for all LLaMA lenses.
  • Subset-KL budget (k_head, k_tail) = Top-k+IS: (512, 1024); Top-k: k = 512
    Production budgets for LLaMA lenses. GPT-2 sweeps show tradeoffs between final-layer, mean, and early-layer KL. The choice is a hand-set cost/quality control, not derived from first principles.
  • translator scale alpha = alpha = r (unit scaling)
    Fixed scale in Eq. (3); alpha=r follows LoRA practice and was not tuned.
  • memory-model coefficient beta = 7.0 bytes per token times vocabulary slot
    Calibrated from the GPT-2 Full-KL minus Top-k peak-memory gap (Appendix G) and used to project 70B readout aggregates. The projection is validated against a measured 70B run to within 0.4%, so it is a calibrated auxiliary constant rather than a central fitted parameter.
assumptions (5)
  • domain assumption The teacher distribution P(·|x) is the correct self-supervised target for lenses.
    The paper follows prior lens work (Belrose et al. 2023) and does not justify why matching the final next-token distribution is the right notion of an interpretable readout; it is the defining assumption of lens methods.
  • domain assumption The frozen final normalization and unembedding provide a sufficient readout for every hookpoint type.
    Eq. (1) and Eq. (3) assume the model's own readout, applied after a learned translator, is an adequate decoder for residual, attention, and MLP activations.
  • domain assumption The Pile corpus is a sufficient training distribution for lenses.
    Appendix B states all lenses were trained on The Pile with no filtering; a preliminary corpus swap showed negligible differences but was not studied systematically.
  • standard math Importance-sampling support condition R(v|x) > 0 wherever P(v|x) > 0 on the tail.
    Required for Theorem 1; the teacher-tail proposal satisfies it by construction.
  • domain assumption The student's full-vocabulary log-partition can be computed exactly in chunks without materializing all student logits.
    Appendix B.1 describes chunked log-sum-exp accumulation; this is an implementation assumption underpinning the Top-k+IS memory claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpreting Language Model Hidden States at Scale." pith.science (2026). https://pith.science/paper/JEXUFH6M

@misc{pith2026260810260,
  author       = {Pith},
  title        = {Pith review of: Interpreting Language Model Hidden States at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEXUFH6M}},
  note         = {Machine review of arXiv:2608.10260}
}
read the original abstract

Lens methods interpret large language models (LLMs) by mapping intermediate activations to the output vocabulary, revealing how next-token predictions develop through the network. Trained lenses remain expensive: affine-translator parameters grow quadratically with model width, while exact, full-vocabulary Kullback--Leibler (KL) training dominates memory. Consequently, prior trained lenses have been applied to models of at most 20B parameters and remain tied to particular component types. We present OmniLens, which applies a single lens family to any model-width activation, whether residual stream, attention, or MLP, and combines two independent scaling techniques. First, low-rank translators make per-lens parameter growth linear in model width and reduce trainable parameters by up to 98.4%. Second, Subset-KL materializes only selected vocabulary logits: its Top-k mode cuts peak training memory by up to 70%, while its importance-sampled variant retains unbiased stochastic gradients for the full KL. These savings enable a dense ensemble of 482 lenses for LLaMA-3.3-70B, providing 6x the coverage of a residual-stream design at the same depth. Model-wide coverage then reveals what single-component lenses cannot: the components where a behavior is most visible need not be those where intervention is most effective, and the most effective interventions lie outside the attention heads examined by prior lens studies. Across three case studies (prompt-injection detection, multi-hop memory injection, and toxicity localization), OmniLens reproduces key published results at substantially lower cost.

Figures

Figures reproduced from arXiv: 2608.10260 by the authors.

Figure 1
Figure 1. Full-rank lens parameter grows as O(Layer×d 2 ) exceeding 200B for LLaMA-3-405B. be trained in a self-supervised manner: the model’s own fi￾nal distribution supplies the target, so no curated labels are needed. A single trained lens can therefore support many downstream analyses. This capability is most valuable when applied densely across a model, that is, with multiple hook￾points at every layer. However, training… view at source ↗
Figure 2
Figure 2. OmniLens at a glance. Hooks are placed at arbitrary user-defined points in the model; each lens applies the low-rank [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Fidelity-parameter tradeoff on GPT-2 Small at step [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (25 more)
Figure 4
Figure 4. Figure 4: Low rank is a constraint, not a compression (GPT [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Layerwise KL on GPT-2 Small (r = 64, step 1,000) at matched budgets of 512 (dashed) and 1024 (solid) scored tokens. Shading marks layers 0–3; envelopes span three seeds. of 512 it reaches final-layer KL 0.194, against 0.067 for Top-k+IS, and its training loss deteriora…
Figure 6
Figure 6. Figure 6: Interpretability applications. (a) Prompt-injection AUROC across ten tasks (bars: means; points: individual tasks). (b) Fraction of achievable 2WMH intervention gain captured at τ = 4. (c) GPT-2 toxicity reduction by component; the dashed line marks the prior attention…
Figure 7
Figure 7. Figure 7: Dynamic hook configurations in hookbox. Any hook can be toggled independently: a macro-scale configuration (left) reads only block outputs, a micro-scale configuration (center) instruments every projection of every head, and mixed-scale configurations (right) target sp…
Figure 8
Figure 8. Figure 8: Layerwise KL difference (baseline KL minus LoRA [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Top-1 agreement (left) and top-10 overlap (right) between LoRA and full-rank lens outputs by layer, over every evaluated checkpoint grouped by run family (white separators). weighted draw from R, so Eti∼R  P(ti |x) R(ti |x) log P(ti |x) Qℓ,u(ti |x)  = X v /∈H R(v|x) …
Figure 10
Figure 10. Figure 10: Token-level Pearson ρ (left) and Kendall τ@100 (right) against the full-rank baseline, by layer and LoRA rank, GPT-2 Small. posal probabilities are held fixed during differentiation rather than differentiated through. F Indexed Logits CUDA Implementation This appendix…
Figure 12
Figure 12. Figure 12: The Top-k+IS decomposition: the contribution of the khead most probable teacher tokens is computed ex￾actly (blue), and the remaining mass is estimated from ktail importance-weighted tail samples (green). Under the teacher￾tail proposal, every importance weight equals…
Figure 13
Figure 13. Figure 13: Measured single-GPU peak memory for an 8B [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]
Figure 14
Figure 14. Figure 14: The persistent optimizer bucket, exact from pa [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Prompt-injection detection AUROC by lens (bars: [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]
Figure 16
Figure 16. Figure 16: Causal lift maxℓ Eℓ/Pobs from memory injection, by model and dataset. The dotted line at 1.0 marks no effect; the GPT-2 2WMH bar is the null case described in the main text. eficial memory vector exists. On the 70B control set both the lens and the depth heuristic sel…
Figure 17
Figure 17. Figure 17: Fraction of achievable causal gain captured [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: Lens-read P(answer) at LLaMA-3-8B as a func￾tion of injection layer (y) and tweak factor τ (x), on the control set (top) and 2WMH (bottom); the star marks the peak (2WMH: layer 17 at τ=4). This is the lens’s view of the injected state, not the model’s output; causal e…
Figure 20
Figure 20. Figure 20: Lens-read answer recovery as a function of the [PITH_FULL_IMAGE:figures/full_fig_p028_20.png]
Figure 22
Figure 22. Figure 22: Collateral damage at the selected layer (2WMH): [PITH_FULL_IMAGE:figures/full_fig_p029_22.png]
Figure 23
Figure 23. Figure 23: Fraction of examples whose top-1 token survives [PITH_FULL_IMAGE:figures/full_fig_p029_23.png]
Figure 24
Figure 24. Figure 24: ToxIn soft subtraction at 8B trades toxicity against [PITH_FULL_IMAGE:figures/full_fig_p029_24.png]
Figure 26
Figure 26. Figure 26: Share of the total toxic signal carried by the top-5 [PITH_FULL_IMAGE:figures/full_fig_p030_26.png]
Figure 25
Figure 25. Figure 25: The whole-model ablation audit (GPT-2 top, [PITH_FULL_IMAGE:figures/full_fig_p030_25.png]
Figure 27
Figure 27. Figure 27: DART per-head toxic-token counts at LLaMA-3- [PITH_FULL_IMAGE:figures/full_fig_p030_27.png]
Figure 28
Figure 28. Figure 28: As Figure 27, for LLaMA-3-70B: the same late [PITH_FULL_IMAGE:figures/full_fig_p031_28.png]
Figure 30
Figure 30. Figure 30: ToxIn zero-ablation: removing the DART-flagged [PITH_FULL_IMAGE:figures/full_fig_p031_30.png]
Figure 29
Figure 29. Figure 29: DART toxic signal aggregated by layer (LLaMA [PITH_FULL_IMAGE:figures/full_fig_p031_29.png]
Figure 31
Figure 31. Figure 31: The classic prediction-trajectory grid (top-1 token of the lens readout at every layer and position; shading = probability) [PITH_FULL_IMAGE:figures/full_fig_p032_31.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

151 extracted references · 12 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems , volume=

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

  2. [2]

    PLOS ONE , publisher =

    On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation , year =. PLOS ONE , publisher =. doi:10.1371/journal.pone.0130140 , author =

  3. [3]

    1905.05950 , archivePrefix=

    Ian Tenney and Dipanjan Das and Ellie Pavlick , year=. 1905.05950 , archivePrefix=

  4. [4]

    2021 , eprint=

    Transformer Interpretability Beyond Attention Visualization , author=. 2021 , eprint=

  5. [5]

    Interpreting

    Nostalgebraist , howpublished=. Interpreting. 2020 , note=

  6. [6]

    2023 , eprint=

    Eliciting Latent Predictions from Transformers with the Tuned Lens , author=. 2023 , eprint=

  7. [7]

    2023 , eprint=

    Attention Lens: A Tool for Mechanistically Interpreting the Attention Head Information Retrieval Mechanism , author=. 2023 , eprint=

  8. [8]

    2023 , eprint=

    Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. 2023 , eprint=

Show all 151 references
  1. [9]

    and Ducharme, R and Vincent, Pascal and Mathematiques, Centre , year =

    Bengio, Y. and Ducharme, R and Vincent, Pascal and Mathematiques, Centre , year =

  2. [10]

    2013 , eprint=

    Efficient Estimation of Word Representations in Vector Space , author=. 2013 , eprint=

  3. [11]

    G lo V e: Global Vectors for Word Representation

    Pennington, Jeffrey and Socher, Richard and Manning, Christopher. G lo V e: Global Vectors for Word Representation. Conference on Empirical Methods in Natural Language Processing. 2014. doi:10.3115/v1/D14-1162

  4. [12]

    2018 , eprint=

    Deep contextualized word representations , author=. 2018 , eprint=

  5. [13]

    C Users J

    Gage, Philip , title =. C Users J. , month = feb, pages =. 1994 , issue_date =

  6. [14]

    2021 , eprint=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. 2021 , eprint=

  7. [15]

    2018 , eprint=

    Self-Attention with Relative Position Representations , author=. 2018 , eprint=

  8. [16]

    Jianlin Su and Yu Lu and Shengfeng Pan and Ahmed Murtadha and Bo Wen and Yunfeng Liu , year=. Ro. 2104.09864 , archivePrefix=

  9. [17]

    Fu and Stefano Ermon and Atri Rudra and Christopher Ré , year=

    Tri Dao and Daniel Y. Fu and Stefano Ermon and Atri Rudra and Christopher Ré , year=. Flash. 2205.14135 , archivePrefix=

  10. [18]

    1810.04805 , archivePrefix=

    Jacob Devlin and Ming-Wei Chang and Kenton Lee and Kristina Toutanova , year=. 1810.04805 , archivePrefix=

  11. [19]

    and Zipser, David , journal=

    Williams, Ronald J. and Zipser, David , journal=. A Learning Algorithm for Continually Running Fully Recurrent Neural Networks , year=

  12. [20]

    2001 , eprint=

    A Bit of Progress in Language Modeling , author=. 2001 , eprint=

  13. [21]

    OpenAI Blog , year =

    Improving Language Understanding by Generative Pre-Training , author =. OpenAI Blog , year =

  14. [22]

    2017 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2017 , eprint=

  15. [23]

    1954 , url=

    Distributional Structure , author=. 1954 , url=

  16. [24]

    2017 , eprint=

    The Mythos of Model Interpretability , author=. 2017 , eprint=

  17. [25]

    2017 , eprint=

    Towards A Rigorous Science of Interpretable Machine Learning , author=. 2017 , eprint=

  18. [26]

    Why Should I Trust You?

    "Why Should I Trust You?": Explaining the Predictions of Any Classifier , author=. 2016 , eprint=

  19. [27]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , year=

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , year=. 2106.09685 , archivePrefix=

  20. [28]

    R eal T oxicity P rompts: Evaluating Neural Toxic Degeneration in Language Models

    Gehman, Samuel and Gururangan, Suchin and Sap, Maarten and Choi, Yejin and Smith, Noah A. R eal T oxicity P rompts: Evaluating Neural Toxic Degeneration in Language Models. Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. doi:10.18653/v1/2020.findin...

  21. [29]

    2021 , eprint=

    Ethical and social risks of harm from Language Models , author=. 2021 , eprint=

  22. [30]

    2022 , eprint=

    Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models , author=. 2022 , eprint=

  23. [31]

    and Gebru, Timnit and McMillan-Major, Angelina and Shmitchell, Shmargaret , title =

    Bender, Emily M. and Gebru, Timnit and McMillan-Major, Angelina and Shmitchell, Shmargaret , title =. 2021 , isbn =. doi:10.1145/3442188.3445922 , booktitle =

  24. [32]

    2024 , eprint=

    Memory Injections: Correcting Multi-Hop Reasoning Failures during Inference in Transformer-Based Language Models , author=. 2024 , eprint=

  25. [33]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Transformer Interpretability Beyond Attention Visualization , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2021 , doi =

  26. [34]

    2402.05602 , archivePrefix=

    Reduan Achtibat and Sayed Mohammad Vakilzadeh Hatefi and Maximilian Dreyer and Aakriti Jain and Thomas Wiegand and Sebastian Lapuschkin and Wojciech Samek , year=. 2402.05602 , archivePrefix=

  27. [35]

    2018 , eprint=

    Understanding intermediate layers using linear classifier probes , author=. 2018 , eprint=

  28. [36]

    2018 , eprint=

    What you can cram into a single vector: Probing sentence embeddings for linguistic properties , author=. 2018 , eprint=

  29. [37]

    2024 , url =

    Monte MacDiarmid and Timothy Maxwell and Nicholas Schiefer and Jesse Mu and Jared Kaplan and David Duvenaud and Sam Bowman and Alex Tamkin and Ethan Perez and Mrinank Sharma and Carson Denison and Evan Hubinger , title =. 2024 , url =

  30. [38]

    2021 , journal=

    A Mathematical Framework for Transformer Circuits , author=. 2021 , journal=

  31. [39]

    How does

    Michael Hanna and Ollie Liu and Alexandre Variengien , year=. How does. 2305.00586 , archivePrefix=

  32. [40]

    Interpretability in the Wild: A Circuit for Indirect Object Identification in

    Kevin Wang and Alexandre Variengien and Arthur Conmy and Buck Shlegeris and Jacob Steinhardt , year=. Interpretability in the Wild: A Circuit for Indirect Object Identification in. 2211.00593 , archivePrefix=

  33. [41]

    2023 , eprint=

    Progress measures for grokking via mechanistic interpretability , author=. 2023 , eprint=

  34. [42]

    GitHub repository , howpublished =

    Nova Patch and Emmanuel Rosa and others , title =. GitHub repository , howpublished =. 2020 , publisher =

  35. [43]

    Orthrus Toxic Dictionary implementation , howpublished =

    Dawid Grad, Orthrus-Lexicon , year =. Orthrus Toxic Dictionary implementation , howpublished =

  36. [44]

    The IEEE International Conference on Computer Vision (ICCV) , month =

    Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books , author =. The IEEE International Conference on Computer Vision (ICCV) , month =

  37. [45]

    2017 , howpublished =

    cjadams and Jeffrey Sorensen and Julia Elliott and Lucas Dixon and Mark McDonald and nithum and Will Cukierski , title =. 2017 , howpublished =

  38. [46]

    2016 , eprint=

    Pointer Sentinel Mixture Models , author=. 2016 , eprint=

  39. [47]

    Aaron Grattafiori and Abhimanyu Dubey and Abhinav Jauhri and Abhinav Pandey and Abhishek Kadian and Ahmad Al-Dahle and et al. , year=. The. 2407.21783 , archivePrefix=

  40. [48]

    Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and Presser, Shawn and Leahy, Connor , journal=. The

  41. [49]

    AAAI Conference on Artificial Intelligence , year=

    On Scalar Embedding of Relative Positions in Attention Models , author=. AAAI Conference on Artificial Intelligence , year=

  42. [50]

    2023 , eprint=

    Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer , author=. 2023 , eprint=

  43. [51]

    Le and Ruslan Salakhutdinov , year=

    Zihang Dai and Zhilin Yang and Yiming Yang and Jaime Carbonell and Quoc V. Le and Ruslan Salakhutdinov , year=. Transformer-. 1901.02860 , archivePrefix=

  44. [52]

    2021 , eprint=

    Challenges in Detoxifying Language Models , author=. 2021 , eprint=

  45. [53]

    2024 , eprint=

    Full Parameter Fine-tuning for Large Language Models with Limited Resources , author=. 2024 , eprint=

  46. [54]

    2021 , eprint=

    Overcoming Catastrophic Forgetting by Generative Regularization , author=. 2021 , eprint=

  47. [55]

    What's in a Name?

    Alejandro Salinas and Amit Haim and Julian Nyarko , year=. What's in a Name?. 2402.14875 , archivePrefix=

  48. [56]

    2020 , isbn =

    Rasley, Jeff and Rajbhandari, Samyam and Ruwase, Olatunji and He, Yuxiong , title =. 2020 , isbn =. doi:10.1145/3394486.3406703 , booktitle =

  49. [57]

    arXiv preprint arXiv:2311.04897 , year=

    Future lens: Anticipating subsequent tokens from a single hidden state , author=. arXiv preprint arXiv:2311.04897 , year=

  50. [58]

    arXiv preprint arXiv:2403.05846 , year=

    Diffusion lens: Interpreting text encoders in text-to-image pipelines , author=. arXiv preprint arXiv:2403.05846 , year=

  51. [59]

    arXiv preprint arXiv:2502.16570 , year=

    Entropy-lens: The information signature of transformer computations , author=. arXiv preprint arXiv:2502.16570 , year=

  52. [60]

    Joint International Conference on Computational Linguistics, Language Resources and Evaluation , pages=

    Jump to Conclusions: Short-Cutting Transformers with Linear Transformations , author=. Joint International Conference on Computational Linguistics, Language Resources and Evaluation , pages=

  53. [61]

    59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pages=

    Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning , author=. 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pages=

  54. [62]

    Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke , journal=

  55. [63]

    Zhao, Jiawei and Zhang, Zhenyu and Chen, Beidi and Wang, Zhangyang and Anandkumar, Anima and Tian, Yuandong , booktitle=. Ga

  56. [64]

    D. G. Horvitz and D. J. Thompson , journal =. A Generalization of Sampling Without Replacement From a Finite Universe , urldate =

  57. [65]

    Empirical

    Andreas Maurer and Massimiliano Pontil , year=. Empirical. 0907.3740 , archivePrefix=

  58. [66]

    Constructing A Multi-hop

    Xanh Ho and Anh-Khoa Duong Nguyen and Saku Sugawara and Akiko Aizawa , year=. Constructing A Multi-hop. 2011.01060 , archivePrefix=

  59. [67]

    2020 , howpublished =

    Schulman, John , title =. 2020 , howpublished =

  60. [68]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, YK and Wu, Yang and others , journal=

  61. [69]

    Zhang, Lunjun and Ba, Jimmy , journal=

  62. [70]

    Better Estimation of the

    Amini, Afra and Vieira, Tim and Cotterell, Ryan , booktitle=. Better Estimation of the

  63. [71]

    A Comedy of Estimators: On

    Shah, Vedant and Obando-Ceron, Johan and Jain, Vineet and Bartoldson, Brian and Kailkhura, Bhavya and Mittal, Sarthak and Berseth, Glen and Castro, Pablo Samuel and Bengio, Yoshua and Malkin, Nikolay and others , journal=. A Comedy of Estimators: On

  64. [72]

    arXiv preprint arXiv:2408.01416 , year=

    The quest for the right mediator: Surveying mechanistic interpretability through the lens of causal mediation analysis , author=. arXiv preprint arXiv:2408.01416 , year=

  65. [73]

    Interpretability Can Be Actionable , author=

  66. [74]

    2026 , eprint=

    Agents of Chaos , author=. 2026 , eprint=

  67. [75]

    Feature Visualization , volume =

    Olah, Chris and Mordvintsev, Alexander and Schubert, Ludwig , year =. Feature Visualization , volume =. Distill , doi =

  68. [76]

    Distill , year =

    Cammarata, Nick and Goh, Gabriel and Carter, Shan and Voss, Chelsea and Schubert, Ludwig and Olah, Chris , title =. Distill , year =

  69. [77]

    What Does BERT Look at? A n Analysis of BERT ' s Attention

    Clark, Kevin and Khandelwal, Urvashi and Levy, Omer and Manning, Christopher D. What Does BERT Look at? A n Analysis of BERT ' s Attention. ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. 2019. doi:10.18653/v1/W19-4828

  70. [78]

    Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned

    Voita, Elena and Talbot, David and Moiseev, Fedor and Sennrich, Rico and Titov, Ivan. Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned. 57th Annual Meeting of the Association for Computational Linguistics. 2019. doi:10.18653/v...

  71. [79]

    Analyzing the Structure of Attention in a Transformer Language Model

    Vig, Jesse and Belinkov, Yonatan. Analyzing the Structure of Attention in a Transformer Language Model. ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. 2019. doi:10.18653/v1/W19-4808

  72. [80]

    36th International Conference on Neural Information Processing Systems , articleno =

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , title =. 36th International Conference on Neural Information Processing Systems , articleno =. 2022 , isbn =

  73. [81]

    What do Neural Machine Translation Models Learn about Morphology?

    Belinkov, Yonatan and Durrani, Nadir and Dalvi, Fahim and Sajjad, Hassan and Glass, James. What do Neural Machine Translation Models Learn about Morphology?. 55th Annual Meeting of the Association for Computational Linguistics. 2017. doi:10.18653/v1/P17-1080

  74. [82]

    A Structural Probe for Finding Syntax in Word Representations

    Hewitt, John and Manning, Christopher D. A Structural Probe for Finding Syntax in Word Representations. 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies. 2019. doi:10.18653/v1/N19-1419

  75. [83]

    2024 , eprint=

    The Linear Representation Hypothesis and the Geometry of Large Language Models , author=. 2024 , eprint=

  76. [84]

    Emergent Linear Representations in World Models of Self-Supervised Sequence Models

    Nanda, Neel and Lee, Andrew and Wattenberg, Martin. Emergent Linear Representations in World Models of Self-Supervised Sequence Models. 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2023. doi:10.18653/v1/2023.blackboxnlp-1.2

  77. [85]

    2023 , eprint=

    Sparse Autoencoders Find Highly Interpretable Features in Language Models , author=. 2023 , eprint=

  78. [86]

    Daniel and Sumers, Theodore R

    Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and Cunningham, Hoagy and Turner, Nicholas L and McDougall, Callum and MacDiarmid, Monte and Fre...

  79. [87]

    AI Alignment Forum , volume=

    Taking features out of superposition with sparse autoencoders , author=. AI Alignment Forum , volume=

  80. [88]

    2024 , eprint=

    Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning , author=. 2024 , eprint=

  81. [89]

    Adaptive learning rate optimization algorithms with dynamic bound based on

    Zhi-Jun Wang and He-Bei Gao and Xiang-Hong Wang and Shuai-Ye Zhao and Hong Li and Xiao-Qin Zhang , keywords =. Adaptive learning rate optimization algorithms with dynamic bound based on. Information Sciences , volume =. 2023 , issn =. doi:https://doi.org/10.1016/j.ins.2023.03....

  82. [90]

    2015 , eprint=

    Distilling the Knowledge in a Neural Network , author=. 2015 , eprint=

  83. [91]

    Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf , year=. Distil. 1910.01108 , archivePrefix=

  84. [92]

    2019 , eprint=

    Not All Samples Are Created Equal: Deep Learning with Importance Sampling , author=. 2019 , eprint=

  85. [93]

    35th International Conference on Machine Learning , pages =

    Adaptive Sampled Softmax with Kernel Based Sampling , author =. 35th International Conference on Machine Learning , pages =. 2018 , editor =

  86. [94]

    2022 , eprint=

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback , author=. 2022 , eprint=

  87. [95]

    2022 , eprint=

    Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned , author=. 2022 , eprint=

  88. [96]

    2020 , eprint=

    Scaling Laws for Neural Language Models , author=. 2020 , eprint=

  89. [97]

    International Conference for High Performance Computing, Networking, Storage and Analysis , articleno =

    Rajbhandari, Samyam and Rasley, Jeff and Ruwase, Olatunji and He, Yuxiong , title =. International Conference for High Performance Computing, Networking, Storage and Analysis , articleno =. 2020 , isbn =

  90. [98]

    2101.06840 , archivePrefix=

    Jie Ren and Samyam Rajbhandari and Reza Yazdani Aminabadi and Olatunji Ruwase and Shuangyan Yang and Minjia Zhang and Dong Li and Yuxiong He , year=. 2101.06840 , archivePrefix=

  91. [99]

    2024 , eprint=

    Steering Language Models With Activation Engineering , author=. 2024 , eprint=

  92. [100]

    Building Production-Ready Probes For

    Kram. Building Production-Ready Probes For. arXiv preprint arXiv:2601.11516 , year=

  93. [101]

    UniReps: 1st Workshop on Unifying Representations in Neural Models , pages=

    Linearly structured world representations in maze-solving transformers , author=. UniReps: 1st Workshop on Unifying Representations in Neural Models , pages=. 2024 , organization_x=

  94. [102]

    arXiv preprint arXiv:2506.20790 , year=

    Stochastic parameter decomposition , author=. arXiv preprint arXiv:2506.20790 , year=

  95. [103]

    Claas Beger and Ryan Yi and Shuhao Fu and Arsenii Kirillovich Moskvichev and Sarah Tsai and Sivasankaran Rajamanickam and Melanie Mitchell , year=. Do

  96. [104]

    Correlations without causation do not support claims of human--

    Vankov, Ivan I and Adolfi, Federico and Heaton, Rachel F and Puebla, Guillermo and Bowers, Jeffrey S , journal=. Correlations without causation do not support claims of human--. 2026 , publisher=

  97. [105]

    33rd ACM International Conference on Information and Knowledge Management , pages =

    Amirizaniani, Maryam and Martin, Elias and Sivachenko, Maryna and Mashhadi, Afra and Shah, Chirag , title =. 33rd ACM International Conference on Information and Knowledge Management , pages =. 2024 , isbn =. doi:10.1145/3627673.3679832 , abstract =

  98. [106]

    On a few pitfalls in

    Tang, Yunhao and Munos, R. On a few pitfalls in. arXiv preprint arXiv:2506.09477 , year=

  99. [107]

    2025 , howpublished =

    Pettyjohn, Jordan , title =. 2025 , howpublished =

  100. [108]

    13th International Conference on Artificial Intelligence and Statistics , pages=

    Understanding the difficulty of training deep feedforward neural networks , author=. 13th International Conference on Artificial Intelligence and Statistics , pages=. 2010 , organization=

  101. [109]

    Probing for semantic evidence of composition by means of simple classification tasks

    Ettinger, Allyson and Elgohary, Ahmed and Resnik, Philip. Probing for semantic evidence of composition by means of simple classification tasks. 1st Workshop on Evaluating Vector-Space Representations for NLP. 2016. doi:10.18653/v1/W16-2524

  102. [110]

    arXiv preprint arXiv:2101.00027 , year=

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

  103. [111]

    Advances in Neural Information Processing Systems , volume=

    Deep reinforcement learning from human preferences , author=. Advances in Neural Information Processing Systems , volume=

  104. [112]

    Advances in Neural Information Processing Systems , volume=

    Learning to summarize with human feedback , author=. Advances in Neural Information Processing Systems , volume=

  105. [113]

    Sparse Logit Sampling: Accelerating Knowledge Distillation in LLM s

    Anshumann and Zaidi, Mohd Abbas and Kedia, Akhil and Ahn, Jinwoo and Kwon, Taehwak and Lee, Kangwook and Lee, Haejun and Lee, Joohyung. Sparse Logit Sampling: Accelerating Knowledge Distillation in LLM s. 63rd Annual Meeting of the Association for Computational Linguistics. 20...

  106. [114]

    The Eleventh International Conference on Learning Representations , year =

    Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning , author=. The Eleventh International Conference on Learning Representations , year =

  107. [115]

    34th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning , pages =

    Trimigno, Giuseppe and Lombardo, Gianfranco and Cagnoni, Stefano , title =. 34th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning , pages =. 2026 , address =

  108. [116]

    arXiv preprint arXiv:2607.15495 , year=

    Verbalizable representations form a global workspace in language models , author=. arXiv preprint arXiv:2607.15495 , year=

  109. [117]

    2026 , eprint=

    Verbalizable Representations Form a Global Workspace in Language Models , author=. 2026 , eprint=

  110. [118]

    Conference on Empirical Methods in Natural Language Processing , pages=

    Backward lens: Projecting language model gradients into the vocabulary space , author=. Conference on Empirical Methods in Natural Language Processing , pages=

  111. [119]

    arXiv preprint arXiv:2401.06102 , year=

    Patchscopes: A unifying framework for inspecting hidden representations of language models , author=. arXiv preprint arXiv:2401.06102 , year=

  112. [120]

    Anthropic Research , year=

    Circuit Tracing: Revealing Computational Graphs in Language Models , author=. Anthropic Research , year=

  113. [121]

    Lieberum, Tom and Conerly, Tom and Sonnerat, Nicolas and others , journal=. Gemma

  114. [122]

    Ko, Jongwoo and Kim, SungnyuSe-Young , booktitle=. Distil

  115. [123]

    Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie , booktitle=. Mini

  116. [124]

    International Conference on Learning Representations , volume=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. International Conference on Learning Representations , volume=

  117. [125]

    2025 , url=

    Adam Karvonen and Can Rager and Johnny Lin and Curt Tigges and Joseph Isaac Bloom and David Chanin and Yeu-Tong Lau and Eoin Farrell and Callum Stuart McDougall and Kola Ayonrinde and Demian Till and Matthew Wearden and Arthur Conmy and Samuel Marks and Neel Nanda , booktitle=...

  118. [126]

    Transcoders Find Interpretable

    Dunefsky, Jacob and Chlenski, Philippe and Nanda, Neel , journal=. Transcoders Find Interpretable

  119. [127]

    Wang, Ziqian , journal=. Logit

  120. [128]

    2023 , eprint=

    Mass-Editing Memory in a Transformer , author=. 2023 , eprint=

  121. [129]

    2024 , eprint=

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author=. 2024 , eprint=

  122. [130]

    2024 , eprint=

    A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity , author=. 2024 , eprint=

  123. [131]

    2024 , eprint=

    Refusal in Language Models Is Mediated by a Single Direction , author=. 2024 , eprint=

  124. [132]

    2025 , eprint=

    Representation Engineering: A Top-Down Approach to AI Transparency , author=. 2025 , eprint=

  125. [133]

    2018 , eprint=

    Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge , author=. 2018 , eprint=

  126. [134]

    2019 , eprint=

    BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions , author=. 2019 , eprint=

  127. [135]

    ``Going on a vacation'' takes longer than ``Going for a walk'': A Study of Temporal Commonsense Understanding

    Zhou, Ben and Khashabi, Daniel and Ning, Qiang and Roth, Dan. ``Going on a vacation'' takes longer than ``Going for a walk'': A Study of Temporal Commonsense Understanding. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter...

  128. [136]

    A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference

    Williams, Adina and Nangia, Nikita and Bowman, Samuel R. A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference. Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technolog...

  129. [137]

    GLUE : A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding

    Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R. GLUE : A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. Proceedings of the 2018 EMNLP Workshop B lackbox NLP : Analyzing and Interpreting Ne...

  130. [138]

    and Gardner, Matt

    Welbl, Johannes and Liu, Nelson F. and Gardner, Matt. Crowdsourcing Multiple Choice Science Questions. Proceedings of the 3rd Workshop on Noisy User-generated Text. 2017. doi:10.18653/v1/W17-4413

  131. [139]

    and Ng, Andrew and Potts, Christopher

    Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D. and Ng, Andrew and Potts, Christopher. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. Proceedings of the 2013 Conference on Empirical Methods in Natur...

  132. [140]

    2020 , eprint=

    LogiQA: A Challenge Dataset for Machine Reading Comprehension with Logical Reasoning , author=. 2020 , eprint=

  133. [141]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  134. [142]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  135. [143]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

  136. [144]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  137. [145]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

  138. [146]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  139. [147]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  140. [148]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  141. [149]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  142. [150]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

  143. [151]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.