Pith. sign in

REVIEW 2 major objections 5 minor 47 references

A reconstruction test for AI explanations does not certify individual claims: the paper shows high reconstruction is achieved while most specific claims are ungrounded, and proposes RECAP, which trains the target model to keep designated co

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 09:58 UTC pith:UWTXBUR3

load-bearing objection The audit is the real contribution; the scale-transfer fix is promising but conditional on a target-design rule the paper only states retrospectively. the 2 major comments →

arxiv 2607.20379 v1 pith:UWTXBUR3 submitted 2026-07-22 cs.AI cs.CL

Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations

Hiskias Dingeto This is my paper
classification cs.AI cs.CL
keywords natural-language autoencodersactivation explanationsreconstruction scorefaithfulnessinterpretabilityprobe-based monitoringRECAPprivate codes
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the reconstruction test used by natural-language autoencoders—regenerating a hidden activation from an explanation and treating the similarity as faithfulness—is structurally insensitive to individual false claims. A flipped claim is penalized only if the reconstruction drops, and in practice most specific claims are not reconstruction-dependent: on a released Qwen-2.5-7B verbalizer, explanations reconstruct at r~0.84 while only about 2% of audited specific claims are grounded at the standard threshold. Under exact synthetic ground truth, the standard co-training recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, so the test can be passed by gist and by private conventions rather than faithful content. The paper's repair is RECAP: during training, linear heads predict external targets from hidden states, making designated content decodable. In a sandbox this yields claim-faithful fresh verbalizers in 5/5 runs at negligible cost, and on a pretrained Pythia-160M it makes the content probe-decodable, raising fresh-verbalizer truth from near zero to 0.44–0.46 and letting a probe catch lie edits at AUC 0.95.

Core claim

The paper's central discovery is that the reconstruction score measures gist and co-adaptation, not claim-level truth. On the released system, meaning-preserving paraphrase keeps the score while masking content words destroys it, and an append-and-rank control recovers a known-encoded topic at only 28 percent raw, showing the score tracks overall meaning. In the sandbox, the co-trained pair uses grounded-but-false wording (private codes) that an independent evaluator cannot read; the grounded-minus-true gap is 0.30–0.80 in 5/5 standard runs and collapses to 0.08–0.23 under an evaluator swap. The author's proposed fix, RECAP, changes the target model rather than the reader: linear heads co-tr

What carries the argument

The central object is the reconstruction objective J_read = max cos(AR(AV(h)), h), which rewards sufficiency but not per-claim truth. The paper's audit instruments are the flip audit (per-claim grounding = reconstruction drop under a minimal-pair edit) and the grounded-vs-true cross with an evaluator swap (an independent reconstructor), which exposes co-adapted private codes. RECAP replaces J_read-side training with a model-side objective J_model = min LLM(theta) + w L_phi(h(theta), y), where linear heads predict external targets; fresh probes then verify decodability. The scaling condition is a target-design rule: auxiliary targets must be chosen so their loss cannot be minimized without ge

Load-bearing premise

The method's transfer to scale rests on the rule that auxiliary targets must be chosen so their loss cannot be minimized without genuinely discriminating, and the paper identifies this rule only after a naive target failed, without giving an a priori criterion for picking such targets.

What would settle it

Run RECAP on a pretrained model using a target selected purely prospectively from the stated rule (before seeing whether probes read it); if a fresh probe's AUC on the designated content is no better than the 0.77–0.80 control baseline, or if a fresh verbalizer's truth rate on that content is not above the control's 0.06, the claim that RECAP makes designated content verifiable at scale is falsified. In the sandbox, a single run in which a fresh verbalizer shows a grounded-minus-true gap comparable to the standard recipe's 0.30–0.80 would falsify the claim that RECAP eliminates co-adapted priv

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Individual claims in reconstruction-scored explanations cannot be treated as verified by a high reconstruction number; claim-level audits or probe-based checks are required.
  • The released Qwen-2.5-7B verbalizer's high score should be interpreted as gist-level informativeness, not evidence that its specific statements are grounded in the activation.
  • For safety monitoring, RECAP provides a designated-content readout that an independent probe can check; false claims about the designated content are flagged at high AUC even when the reconstruction score is gamed.
  • Because decodability erodes within a few hundred steps of head-free fine-tuning, interpretability guarantees from such supervision must be maintained continuously, not set once.
  • The grounded-vs-true cross and evaluator swap give future deployments a cheap way to test whether an NLA's explanations carry private codes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The claim-level audit standard the paper introduces is portable: any reconstruction-scored explainer, not just the released Qwen-2.5-7B pair, could be audited with the same flip protocol and evaluator swap, and the ~2% grounded-fraction figure is likely a lower bound because the instrument is sensitivity-limited.
  • The target-design rule is the method's open condition. If it cannot be stated prospectively—what property makes a target's loss impossible to minimize without discrimination—then RECAP's success at scale depends on post-hoc validation, and a principled selection rule would be needed before relying on it in new settings.
  • The paper's distinction between storage (decodability) and verbalizability suggests that even successful interpretation pipelines may need probes rather than prose as the trustworthy interface; a testable extension is whether a generation-time objective that directly rewards truth of generated tokens, rather than reconstruction, would close the 0.44–0.46 gap.
  • For frontier models, the open scaling question is whether external self-supervised targets (like the offset-8 token) remain sufficient to keep designated content decodable as model scale grows; the paper did not vary model scale beyond 160M.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper argues that the reconstruction test used by natural-language autoencoders (NLAs) is structurally insensitive to individual false claims in generated explanations, and that passing the test is therefore not evidence of claim-level faithfulness. It supports this with two empirical demonstrations: on a released Qwen-2.5-7B verbalizer/reconstructor pair, explanations reconstruct well above chance while only a small fraction of audited specific claims are reconstruction-dependent (the score tracks gist rather than specifics); and in an exact-ground-truth synthetic sandbox, co-trained verbalizer/reconstructor pairs develop private codes—grounded but false wording—in 5/5 standard-recipe runs. The paper then proposes RECAP, a method that co-trains linear auxiliary heads on external targets during target-model training, so that designated content remains linearly decodable from chosen hidden states. In the sandbox, RECAP gives perfect fresh-probe decodability and propagates to truthful fresh verbalizers at negligible cost. On continued pretraining of Pythia-160M, designated content becomes probe-decodable (AUC 0.95–1.00) under a stated 'target-design rule,' with a +0.010-nat tax for 64 balanced targets; a fresh probe scores the verbalizer's true claims above false ones at AUC 0.965 versus 0.823 for the control, and this survives a report-space adversary. The paper is notably candid about limitations: decodability is not verbalizability, the method depends on target choice, a

Significance. If the main claims hold, the paper makes a substantial contribution to interpretability and AI-safety methodology. The audit protocols—claim-level flips with validity controls, the grounded-vs-true cross, the evaluator swap—are well designed, and the exact-ground-truth sandbox gives unusually strong causal evidence that the standard reconstruction recipe can fail in a non-obvious way. The use of fresh independent probes rather than co-trained heads for all decodability measurements is a methodological strength. The adversarial lie-ladder, held-out band null, frozen-threshold anti-circularity check, and class-balanced loss analysis are also careful. The central theoretical point—that reconstruction rewards sufficiency, not claim truth—is sound and important. The main limitation is that the method's transfer to real models rests on an underspecified target-design rule and on a single small model (Pythia-160M); the paper is open about this, but the methodological recipe as written is not yet fully reproducible for new applications.

major comments (2)
  1. [§5.3, Eq. (3), Table 1; §8 Limitations] The RECAP scale transfer depends on the target-design rule stated in §5.3: 'use targets whose loss cannot be minimized without genuine discrimination.' This rule is introduced only after the K=512 naive presence head failed, and no prospective, operationalizable criterion is given for identifying such targets before training. The paper itself concedes in §8 that 'it depends on choosing a good target: a poorly chosen one can report success while decoding nothing.' This is load-bearing because all scale results—Table 1 probe AUCs, §5.4 verbalizer truth, §6 monitor AUC and adversarial robustness—were obtained under this retrospectively stated rule. As written, the rule is unfalsifiable in practice: if a chosen target fails, the method can always attribute the failure to poor target selection. The revision should provide a pre-hoc diagnostic or formal condition for when an auxiliary target's
  2. [§5.4, Table 13; §6] The scale verbalizer still shows a substantial grounded-minus-true gap (+0.33 after RECAP, versus +0.42 control) and per-word truth of only 0.44–0.46; the paper honestly says the content is 'stated more truthfully, not code-free.' This means the sandbox 'codes vanish' result does not fully transfer, and the practical guarantee at scale is explicitly probe decodability, not verbalizer truth. I do not count this as a fatal error because the paper frames the probe as the reliable readout and the adversarial monitor is measured on probe AUC. But the title and abstract should not leave the impression that RECAP makes free-form explanations faithful at scale; the distinction between 'decodability' and 'verbalizability' should be emphasized in the framing, perhaps in the title. If the contribution is intended as a safety mechanism, the scope of the guarantee (designated content only, probe-veri
minor comments (5)
  1. [Abstract and §3.1] The phrase 'as few as ∼2%' is slightly ambiguous: the base system gives 4.2% at τ=0.02, 2.1% at τ=0.05, 1.6% at τ=0.10. The paper correctly calls these sensitivity-limited lower bounds, but the abstract's wording could be read as an upper bound. Consider saying 'at most a few percent' or quoting the τ=0.05 number.
  2. [§5.3 and Appendix Table 12] The auxiliary-weight sweep appears to be single-seed per point (Table 12 caption). For the crucial claim that identity targets remain readable at w=0.03 with tax within noise of zero, a paired multi-seed confidence interval would be more persuasive. The paper gives one for K=64 balanced at w=0.1, but not for the low-weight identity case.
  3. [Appendix C, Table 3] The erosion of decodability under head-free fine-tuning within 250 steps is a major deployment constraint and is only in an appendix. I suggest moving the sentence 'RECAP is a running cost, not a one-time purchase' into the main text, since safety users will need it to design monitoring schedules.
  4. [Figures] Figure 3(b) contains a small grammar error in the generated explanation ('a acorn' instead of 'an acorn'). This does not affect the results but should be fixed for publication.
  5. [§5.4] The offset-8 target is reported as unrecovered with a base-rate caveat (true designated token in only 2.2% of taps). This is appropriately transparent, but the reader may wonder whether it weakens the claim that 'designated content' is decodable. Please add one sentence explaining whether the identity target at offset−8 is considered a failed target under the §5.3 design rule, or simply underpowered.

Circularity Check

0 steps flagged

No significant circularity: the audit findings are grounded in external ground truth and independent fresh probes, and RECAP's decodability is its explicit training objective rather than a disguised prediction.

full rationale

The central claims are not forced by definition or by self-citation. The reconstruction-test critique is a counterfactual audit on a released Qwen-2.5-7B NLA with per-edit validity controls, a flip-noise null, and a transform profile; the conclusion that specific flips rarely change the score is an empirical measurement, not an equation-level identity. The synthetic-code discovery uses exact ground truth and an evaluator swap, so the 'grounded-and-false' designation is operational and independently verified. RECAP's decodability numbers are held-out fresh-probe AUCs on the same external targets the auxiliary heads supervised; that is verification of the training objective, not a prediction derived from it. The paper explicitly acknowledges this and adds anti-circularity controls: a held-out vocabulary band (words ranked 64-128) shows no RECAP advantage, and flag thresholds are calibrated on disjoint splits. The verbalizer-truth results are not constructed: both RECAP and control verbalizers receive identical cold-start SFT and reconstruction training, so the difference isolates the target model's decodability. The scale truth gain (0.44-0.46 vs near zero) is a non-trivial propagation effect rather than a fitted input. The post-hoc 'target-design rule' in Section 5.3 is a limitation about prospective specification, not a circular reduction; the paper reports the failed K=512 naive target and the balancing fix transparently. No load-bearing self-citation chain is present, and no cited prior work by this author is used to forbid alternatives. The strongest claims are therefore self-contained against external or independently fit quantities.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central empirical claims rest on hand-chosen hyperparameters (w, K, τ, probe threshold) and on assumptions about linear decodability and sandbox ground truth. No new physical or non-physical entity is postulated; the only novel conceptual object is the 'private code' phenomenon, which is an observed behavior, not an invented entity. The most fragile addition is the post-hoc target-design rule that governs RECAP's scale transfer.

free parameters (4)
  • aux weight w = 0.1 (balanced K=64); 0.03 for identity target
    Hand-chosen; the +0.010 nat cost claim and decodability survival depend on this weight, and the naive w=1.0 presence head at K=512 fails (Tables 1, 12).
  • target count K = 8, 64, 512 (main experiments: 64)
    Scaling cost and the K=512 naive failure depend on this choice; no a priori rule is given for selecting K.
  • grounding threshold τ = 0.02, 0.05, 0.10
    The ~2% grounded-fraction claim is threshold-dependent (4.2%/2.1%/1.6% at the three τ values); τ=0.05 is used as the default throughout.
  • probe confidence threshold = 0.9
    The sandbox discrepancy detector's coverage and recall are reported at this single threshold; no sweep is shown.
axioms (5)
  • domain assumption Linear probes reliably indicate whether content is present in a hidden activation.
    The whole RECAP proposal and probe-based monitor treat fresh-probe decodability as the operational definition of 'content present' (§5–6); no evidence is given that linear decodability is the right notion for all representational formats.
  • domain assumption The synthetic templated domains provide exact ground truth for faithfulness.
    The code-discovery result is measured in these sandboxes (Appendix D); transfer to real models is only partial and explicitly limited.
  • domain assumption Closed-vocabulary, template-parsed claims are the right unit for claim-level faithfulness.
    Section 8 explicitly limits conclusions to closed-vocabulary claims; open-ended faithfulness is unmeasured.
  • domain assumption Pythia-160M continued pretraining is a reasonable proxy for frontier LLM behavior.
    All scale results use one 160M model, and the authors state frontier scale is untested.
  • ad hoc to paper There exists a principled way to choose auxiliary targets whose loss cannot be minimized by trivial prediction.
    The 'target-design rule' of §5.3 is introduced after the naive K=512 head failed; no prospective criterion is given.

pith-pipeline@v1.3.0-alltime-deepseek · 22814 in / 10548 out tokens · 96433 ms · 2026-08-01T09:58:52.208394+00:00 · methodology

0 comments
read the original abstract

Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).

Figures

Figures reproduced from arXiv: 2607.20379 by Hiskias Dingeto.

Figure 1
Figure 1. Figure 1: The two architectures. (a) A natural-language autoencoder (prior work) (Fraser-Taliente et al. 2026): AV maps an [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Decoding content from a real model’s hidden state (one held-out Pile (Gao et al. 2020) passage; fresh probes read the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: One representative held-out story, two target mod [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: What it takes to restore a failed reader: accuracy [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Decodability exceeds verbalizability at scale [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 16 linked inside Pith

  1. [1]

    Natural Language Autoencoders Produce Unsupervised Explanations of

    Fraser-Taliente, Kit and Kantamneni, Subhash and Ong, Euan and Mossing, Dan and Lu, Christina and others , year =. Natural Language Autoencoders Produce Unsupervised Explanations of

  2. [2]

    2026 , note =

    Pan, Alexander and Chen, Lijie and Steinhardt, Jacob , booktitle =. 2026 , note =

    2026

  3. [3]

    Activation Oracles: Training and Evaluating

    Karvonen, Adam and Chua, James and Dumas, Cl. Activation Oracles: Training and Evaluating. arXiv preprint arXiv:2512.15674 , year =

    arXiv

  4. [4]

    arXiv preprint arXiv:2606.02609 , year =

    Building Better Activation Oracles , author =. arXiv preprint arXiv:2606.02609 , year =

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2512.15712 , year =

    Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants , author =. arXiv preprint arXiv:2512.15712 , year =

    arXiv

  6. [6]

    ICML , year =

    Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models , author =. ICML , year =

  7. [7]

    arXiv preprint arXiv:2605.25903 , year =

    Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation , author =. arXiv preprint arXiv:2605.25903 , year =

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2509.13316 , year =

    Do Activation Verbalization Methods Convey Privileged Information? , author =. arXiv preprint arXiv:2509.13316 , year =

    Pith/arXiv arXiv

  9. [9]

    2017 , note =

    Chu, Casey and Zhmoginov, Andrey and Sandler, Mark , booktitle =. 2017 , note =

    2017

  10. [10]

    arXiv preprint arXiv:2310.18512 , year =

    Preventing Language Models From Hiding Their Reasoning , author =. arXiv preprint arXiv:2310.18512 , year =

    Pith/arXiv arXiv

  11. [11]

    Hidden in Plain Text: Emergence and Mitigation of Steganographic Collusion in

    Mathew, Yohan and Matthews, Ollie and McCarthy, Robert and Velja, Joan and Schroeder de Witt, Christian and Cope, Dylan and Schoots, Nandi , booktitle =. Hidden in Plain Text: Emergence and Mitigation of Steganographic Collusion in. 2025 , note =

    2025

  12. [12]

    arXiv preprint arXiv:2506.01926 , year =

    Large Language Models Can Learn and Generalize Steganographic Chain-of-Thought Under Process Supervision , author =. arXiv preprint arXiv:2506.01926 , year =

    arXiv

  13. [13]

    and Africa, David Demitri and Schroeder de Witt, Christian , journal =

    Rippin, Jimmy Laurence and Marshall, Simon C. and Africa, David Demitri and Schroeder de Witt, Christian , journal =. Tool Use Enables Undetectable Steganography in Multi-Agent

  14. [14]

    arXiv preprint arXiv:2604.01151 , year =

    Detecting Multi-Agent Collusion Through Multi-Agent Interpretability , author =. arXiv preprint arXiv:2604.01151 , year =

    Pith/arXiv arXiv

  15. [15]

    Obfuscated Activations Bypass

    Bailey, Luke and Serrano, Alex and Sheshadri, Abhay and Seleznyov, Mikhail and others , journal =. Obfuscated Activations Bypass

  16. [16]

    arXiv preprint arXiv:2503.11926 , year =

    Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation , author =. arXiv preprint arXiv:2503.11926 , year =

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2606.15980 , year =

    Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness , author =. arXiv preprint arXiv:2606.15980 , year =

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2510.21531 , year =

    Probe-based Fine-tuning for Reducing Toxicity , author =. arXiv preprint arXiv:2510.21531 , year =

    arXiv

  19. [19]

    arXiv preprint arXiv:2506.06609 , year =

    Transferring Linear Features Across Language Models With Model Stitching , author =. arXiv preprint arXiv:2506.06609 , year =

    arXiv

  20. [20]

    ICLR , year =

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

  21. [21]

    arXiv preprint arXiv:2503.09543 , year =

    van der Wal, Oskar and Lesci, Pietro and M. arXiv preprint arXiv:2503.09543 , year =

    Pith/arXiv arXiv

  22. [22]

    CVPR , year =

    Understanding Image Representations by Measuring Their Equivariance and Equivalence , author =. CVPR , year =

  23. [23]

    ICML , year =

    Similarity of Neural Network Representations Revisited , author =. ICML , year =

  24. [24]

    arXiv preprint arXiv:2303.08112 , year =

    Eliciting Latent Predictions from Transformers with the Tuned Lens , author =. arXiv preprint arXiv:2303.08112 , year =

    Pith/arXiv arXiv

  25. [25]

    AISTATS , year =

    Deeply-Supervised Nets , author =. AISTATS , year =

  26. [26]

    CVPR , year =

    Going Deeper with Convolutions , author =. CVPR , year =

  27. [27]

    ICML , year =

    Concept Bottleneck Models , author =. ICML , year =

  28. [28]

    arXiv preprint arXiv:2106.13314 , year =

    Promises and Pitfalls of Black-Box Concept Learning Models , author =. arXiv preprint arXiv:2106.13314 , year =

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2105.04289 , year =

    Do Concept Bottleneck Models Learn as Intended? , author =. arXiv preprint arXiv:2105.04289 , year =

    Pith/arXiv arXiv

  30. [30]

    Nature Machine Intelligence , volume =

    Concept Whitening for Interpretable Image Recognition , author =. Nature Machine Intelligence , volume =

  31. [31]

    arXiv preprint arXiv:2310.17230 , year =

    Codebook Features: Sparse and Discrete Interpretability for Neural Networks , author =. arXiv preprint arXiv:2310.17230 , year =

    Pith/arXiv arXiv

  32. [32]

    ICLR , year =

    Concept Bottleneck Large Language Models , author =. ICLR , year =

  33. [33]

    arXiv preprint arXiv:2410.04332 , year =

    Gradient Routing: Masking Gradients to Localize Computation in Neural Networks , author =. arXiv preprint arXiv:2410.04332 , year =

    Pith/arXiv arXiv

  34. [34]

    CVPR , year =

    Exploring Simple Siamese Representation Learning , author =. CVPR , year =

  35. [35]

    ICML , year =

    Understanding Self-Supervised Learning Dynamics without Contrastive Pairs , author =. ICML , year =

  36. [36]

    arXiv preprint arXiv:2606.32038 , year =

    Introspective Coupling: Self-Explanation Training Tracks Behavioral Change Despite Fixed Supervision , author =. arXiv preprint arXiv:2606.32038 , year =

    Pith/arXiv arXiv

  37. [37]

    NeurIPS , year =

    Fine-Grained Human Feedback Gives Better Rewards for Language Model Training , author =. NeurIPS , year =

  38. [38]

    arXiv preprint arXiv:2407.03572 , year =

    Core: Robust Factual Precision with Informative Sub-Claim Identification , author =. arXiv preprint arXiv:2407.03572 , year =

    Pith/arXiv arXiv

  39. [39]

    ICML , year =

    Language Models with Conformal Factuality Guarantees , author =. ICML , year =

  40. [40]

    Proceedings of the 6th BlackboxNLP Workshop , year =

    Rigorously Assessing Natural Language Explanations of Neurons , author =. Proceedings of the 6th BlackboxNLP Workshop , year =

  41. [41]

    ACL , year =

    Faithfulness Tests for Natural Language Explanations , author =. ACL , year =

  42. [42]

    NeurIPS , year =

    Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting , author =. NeurIPS , year =

  43. [43]

    arXiv preprint arXiv:2307.13702 , year =

    Measuring Faithfulness in Chain-of-Thought Reasoning , author =. arXiv preprint arXiv:2307.13702 , year =

    Pith/arXiv arXiv

  44. [44]

    ICLR , year =

    Sparse Autoencoders Find Highly Interpretable Features in Language Models , author =. ICLR , year =

  45. [45]

    Gao, Leo and Biderman, Stella and Black, Sid and others , journal=. The

  46. [46]

    International Conference on Machine Learning , year=

    Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling , author=. International Conference on Machine Learning , year=

  47. [47]

    Penedo, Guilherme and others , journal=. The