Pith. sign in

REVIEW 3 major objections 5 minor 4 references

SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read No sparse autoencoder recovers true features even in a best-case linear world.

desk verdict A well-engineered synthetic benchmark with a genuinely interesting MP-SAE finding, but the headline 'architecture bottleneck' claim is undercut by capacity and training-budget confounds. read the letter →

arxiv 2602.14687 v2 pith:LQB2CHW5 submitted 2026-02-16 cs.LG cs.AI

David Chanin , Adrià Garriga-Alonso This is my paper · ORCID
classification cs.LGcs.AI
keywords sparseautoencoderssyntheticbenchmarklinearrepresentationhypothesissuperpositionfeaturerecoveryinterpretabilitydictionarylearningmatchingpursuit
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

SynthSAEBench generates large-scale synthetic activations that obey the Linear Representation Hypothesis by construction: every ground-truth feature is a linear direction, with realistic correlation, hierarchy, and superposition. Because the true features are known, the benchmark can measure feature recovery directly. The paper shows that five current SAE architectures all fall short of perfect feature recovery on the standard 16k-feature model, even though the data is generated exactly as the hypothesis assumes. It also reproduces known LLM failures—reconstruction quality decoupled from latent quality, poor probing, and a precision–recall trade-off driven by L0—which supports using SynthSAEBench as a cheap, low-noise diagnostic. The authors argue that failure here means the bottleneck is in SAE architectures, not in the representation hypothesis.

What carries the argument

The load-bearing object is SynthSAEBench's generative model: a dictionary D of N unit-norm ground-truth feature directions, a low-rank Gaussian-copula firing sampler that introduces structured correlation, rectified-Gaussian magnitudes, and a tree hierarchy with mutually exclusive children applied as post-processing. The benchmark model SynthSAEBench-16k (N=16,384 features, hidden dimension 768, Zipfian firing probabilities, L0≈34) gives a standardized testbed. The argument runs on the contrast between three metric families: variance explained (reconstruction), mean correlation coefficient / uniqueness (feature recovery via matched decoder columns), and per-latent probing F1 (classifier qual

What would settle it

Train any standard SAE on SynthSAEBench-16k with the recommended configuration and observe near-perfect MCC, probing F1, and uniqueness at any L0; that would refute the claim that current architectures fail in the linear world. Alternatively, if a current architecture recovers known ground-truth features perfectly on a real LLM while still failing on SynthSAEBench-16k, the lower-bound transfer assumption would be broken.

Watch

Extended reading notes

Core claim

The central claim is that, under the most favorable possible conditions for sparse autoencoders—data in which the Linear Representation Hypothesis holds exactly, with ground-truth features available—current SAE architectures still fail to recover the true features. Standard L1, BatchTopK, JumpReLU, Matryoshka BatchTopK, and Matching Pursuit SAEs trained on SynthSAEBench-16k all achieve imperfect feature recovery (MCC ~0.6–0.8) and probing F1 far below supervised probes (0.97). The paper further identifies a new failure mode: Matching Pursuit SAEs use their more expressive encoder to exploit superposition noise—improving reconstruction while degrading feature recovery as superposition increas

Load-bearing premise

The claim that failure here predicts failure on real LLMs rests on the assumption that features built from unit-norm directions with Gaussian-copula correlations, tree hierarchies, and Zipfian firing probabilities capture the structure of real neural representations well enough; the paper concedes that synthetic data cannot cover all aspects and that unknown unknowns remain.

Editorial extensions

If this is right

  • SAE architectures that fail on SynthSAEBench-16k cannot blame the Linear Representation Hypothesis, since it holds by construction; architectural innovation, not a richer data model, is the needed step.
  • The MP-SAE overfitting result is a caution for evaluating SAEs on reconstruction loss alone: an architecture can score best on reconstruction while learning worse features, so any real-LLM benchmark that relies on reconstruction metrics can be misled.
  • Matching SAEs against supervised probes on this benchmark quantifies the probing gap at ~0.88 vs ~0.97 F1, showing the SAE-probe gap is a real limitation of SAE architectures rather than an artifact of LLM noise.
  • Because all phenomena are controllable, the benchmark allows researchers to ablate superposition, correlation, or hierarchy one at a time and observe their separate effects on failure modes.
  • The benchmark runs in minutes per SAE on a single GPU, enabling many-seed comparisons that would be expensive on LLMs.

Reading between the lines

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

  • If the lower-bound logic transfers, then failure on real LLMs should be at least as severe as failure here; a SAE that achieves near-perfect recovery on SynthSAEBench-16k-like models becomes a stronger candidate for LLM deployment, and the benchmark could serve as a filter before expensive LLM runs.
  • The Gaussian-copula correlation model could be extended to non-linear feature manifolds or Minkowski-structured representations; the paper itself flags these as future work, and a SAE that recovers features under those hypotheses would distinguish representation failures from architecture failures.
  • The MP-SAE superposition overfitting suggests a concrete diagnostic for real LLMs: as hidden dimension shrinks (more superposition) and compute is held, a reconstruction improvement accompanied by degraded downstream or interpretability metrics would indicate the same exploit.
  • A testable extension is to train SAEs on SynthSAEBench-16k with targeted ablations of hierarchy only vs correlation only; the paper's results suggest the hierarchy is the dominant correlation structure, but the interaction could be mapped precisely with the released tools.
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

3 major / 5 minor

Summary. The paper introduces SynthSAEBench, a synthetic-data benchmark for sparse autoencoders (SAEs) in which ground-truth features are known by construction. The generative model builds a feature dictionary with N=16,384 unit-norm directions in D=768 dimensions, with Zipfian firing probabilities, low-rank correlations, tree hierarchy with mutual exclusion, and controlled superposition. The authors define SynthSAEBench-16k, train five SAE architectures (standard L1, JumpReLU, BatchTopK, Matryoshka BatchTopK, Matching Pursuit) at width 4096 on 200M samples with L0 targets from 15 to 45 and 5 seeds, and report reconstruction, MCC, probing F1/precision/recall, uniqueness, and dead latents. They find that Matryoshka SAEs achieve the best latent quality despite poor reconstruction, that Matching Pursuit SAEs achieve the best reconstruction but poor latent quality, and that no SAE reaches perfect feature recovery. They also report that MP-SAEs improve reconstruction as superposition increases while latent quality degrades, interpreting this as overfitting superposition noise. The paper positions SynthSAEBench as a lower-bound test for LLM SAEs and concludes that failures indicate a bottleneck in SAE architectures and training dynamics, not in the Linear Representation Hypothesis.

Significance. If the benchmark is valid, it addresses a real gap: LLM SAE benchmarks are noisy and lack ground truth, while existing toy models are small-scale and bespoke. The paper ships code, a standard benchmark model, and 5-seed comparisons with clear metrics; the ability to reproduce several qualitative LLM SAE phenomena (Matryoshka reconstruction/quality disconnect, MP probing failure, the SAE-vs-probe gap, and the L0-mediated precision-recall trade-off) is a genuine strength. The newly reported MP-SAE overfitting pattern, if confirmed, would be a valuable diagnosis with ground-truth metrics. However, the strongest conclusions — that current architectures fail in a best-case LRH world and that the bottleneck is architectural — currently outrun the experiments because of the width choice and the superposition-ablation confound. The benchmark itself is a useful contribution even if these interpretive claims are substantially softened.

major comments (3)
  1. [§5.1, §8; Fig. 5] All evaluated SAEs use L=4096 latents while the generating dictionary has N=16,384 features. The abstract and §8 state that 'no SAE architecture achieves perfect feature recovery' and that 'the bottleneck is in the SAE architectures and training dynamics themselves.' With a 4× undercomplete dictionary, full recovery of all 16k ground-truth features is impossible for any architecture, so the observed sub-perfect MCC and F1 do not by themselves distinguish an architectural/training failure from a capacity limitation. A width-matched run (L=N) or, at minimum, a report of how many of the 16,384 ground-truth features are actually recovered (rather than only the matched-subset MCC) is needed to support the bottleneck claim. Without this, the headline conclusion is underdetermined.
  2. [§6.2, Fig. 7] The superposition ablation varies hidden dimension D from 256 to 1536 and simultaneously scales training samples by (D/768)^0.6. D controls not only superposition level ρmm but also activation statistics, SAE parameter count, and the intrinsic difficulty of reconstruction. The conclusion that MP-SAEs 'exploit superposition noise' rests on comparing across this joint change. Please hold the training budget fixed (or matched by compute) and vary only the feature dictionary geometry, e.g. by changing N at fixed D, or otherwise isolate superposition from the D/training confound. As written, the 'overfitting superposition noise' diagnosis is not uniquely identified.
  3. [Abstract and §8; Appendix A] The lower-bound claim — 'SAE architectures that fail when the LRH holds by construction have little hope on real LLMs' — is an external-validity conjecture. The paper demonstrates qualitative similarity to four previously reported LLM phenomena, but does not show that the relative ordering of architectures on SynthSAEBench-16k predicts their relative ordering on LLM benchmarks. Appendix A itself concedes there may be 'unknown unknowns'. At minimum, the claim should be softened to a hypothesis, or validated by a rank-correlation comparison between SynthSAEBench-16k and SAEBench (or another LLM benchmark) across the same set of architectures.
minor comments (5)
  1. [§4.2, Eq. (18)] Uniqueness should use min(L,N) in the denominator to remain valid when L>N; currently a perfect dictionary with more latents than features would receive a uniqueness score below 1.
  2. [§6.2 and Fig. 7] The x-axis is labeled 'Superposition level (ρmm)' but the manipulation is hidden dimension D. Please state explicitly that D was varied, how ρmm was computed for each configuration, and report D values or a secondary axis.
  3. [Figs. 5 and 6] The shaded standard-deviation bands are often invisible at the plotted scale. Consider adding numeric error bars or a supplementary table so the 5-seed variability is actually readable.
  4. [§5.1] The rationale for width 4096 is stated as 'realistic', but this choice is central to the undercompleteness issue. Please justify it with a reference or a sensitivity analysis, or explicitly acknowledge that the benchmark targets a fixed undercomplete regime.
  5. [Appendix I] The SAEBench noise figure would be more informative with repeated-seed statistics or error bars; currently it shows single trajectories, which makes the 'noise' claim hard to quantify.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional or fitted-input circularity: the benchmark is self-contained against independent ground-truth metrics. The main caveats are a capacity confound (4x undercomplete dictionary) and mildly self-referential validation, neither of which is a circular derivation.

full rationale

SynthSAEBench is an empirical benchmark rather than a derivation chain, and its headline quantities (MCC, probing F1, R^2, L0, dead latents) are computed from ground-truth feature directions and firings that are independently generated by the model in Section 3. No parameter is fitted to reproduce the target SAE phenomena; the generator's correlation, hierarchy, superposition, Zipfian firing, and magnitude settings are chosen from independent observations (Sections 3.1-3.5, 5), and the resulting SAE failure metrics are measured, not imposed. The 'reproduces known LLM SAE phenomena' claims are validated against metrics measured on SAEs trained on the synthetic data, not by construction, so the central result (no tested SAE achieves perfect recovery) does not reduce to an input fit. Two concerns keep this from being a clean score 0-1. First, the recommended width of 4096 latents against N=16,384 ground-truth features (Section 5.1) makes full feature recovery impossible for any architecture, so the Section 8 inference that 'the bottleneck is in the SAE architectures and training dynamics themselves' is partly confounded by capacity, though this is a design/validity issue rather than a circular reduction. Second, the paper leans on several self-authored prior results (Chanin 2025; Chanin et al. 2025a,b; Chanin & Garriga-Alonso 2025) when asserting that observed synthetic failures reproduce LLM phenomena, and the generator explicitly bakes in the correlation/hierarchy/superposition structures from those works; this makes the external-transfer claim partially scaffolded by the authors' own prior assumptions. Appendix A honestly flags this: 'Synthetic data cannot capture all aspects of real neural network representations' and warns of 'unknown unknowns'. These are stated limitations, not hidden circular steps. Overall, the benchmark's ground-truth evaluation is self-contained, so no formal circularity is present; the score reflects the mildly self-referential validation rather than a definitional or fitted-input equivalence.

Assumptions & free parameters 10 free parameters · 7 assumptions · 0 invented entities

No new physical or mathematical entities are postulated. The free parameters are the benchmark model hyperparameters, which are hand-chosen rather than fitted to LLM data; the axioms are the LRH and simplified dependence structures. The central claim's external validity depends on these choices, many of which are acknowledged in Appendix A as approximations.

free parameters (10)
  • Benchmark dimensions: N=16,384, D=768, ρmm≈0.15 = N=16,384; D=768
    Scale chosen for tractability on one GPU and to roughly match Pythia-160m layer 10 activation norms; not fitted to target phenomena.
  • Zipfian firing distribution = pmax=0.4, pmin=5e-4, exponent 0.5
    Sets feature frequencies to be power-law-like; these constants determine L0≈34 and which features are learnable. Chosen by hand, not inferred from LLM data.
  • Firing magnitude parameters = mean linear 5.0→4.0; σ~|N(0.5,0.5²)|
    Heterogeneous magnitudes chosen for realism; there is no LLM ground truth to calibrate them.
  • Hierarchy structure = 128 roots, branching 4, depth 3, all children mutually exclusive, parent-scaled magnitudes
    Defines the hierarchy-induced phenomena the benchmark is meant to test; no evidence this topology matches any language model hierarchy.
  • Low-rank correlation structure = rank r=25, scale s=0.1
    Controls strength and dimensionality of feature correlations; arbitrary but not fitted to target results.
  • Orthogonalization of dictionary = 100 steps, lr 3e-4
    Reduces spurious pairwise cosine similarity, changing effective superposition geometry.
  • L0 targets for architecture comparison = L0 15..45; L0=25 for superposition ablations
    Architectures are compared at matched L0; choice of target affects all rankings and conclusions.
  • JumpReLU init threshold and latent norm = threshold=0.5, latent norm=0.5
    Ad hoc tuning to reduce dead latents on SynthSAEBench; deviates from Anthropic defaults and may not transfer to LLMs.
  • L0 autotuner constants = α=0.99, αr=0.95, Ki=3e-4, s=10, γc=0.01, clamp [0.01,100]
    Controller hyperparameters used to match target L0; could affect comparisons if unstable.
  • Superposition ablation training-sample scaling = (d/768)^0.6
    Training budgets scaled with dimension following Gao et al.; exponent imported from a prior scaling law and may not be appropriate here.
assumptions (7)
  • domain assumption Linear Representation Hypothesis: each concept is a linear direction and activations are a sparse combination of these directions.
    The generative model encodes features as unit vectors and activations as D^T c + b (Eq. 10). The benchmark's lower-bound logic relies on this hypothesis being a useful proxy for LLM features.
  • standard math Gaussian copula thresholding produces the desired marginal firing probabilities and correlation structure.
    Uses Φ^{-1}(1-p_i) thresholds on correlated Gaussian samples (Eq. 8); correctness relies on standard copula theory, not verified in the paper.
  • domain assumption Hierarchical constraint c_child ← c_child * 1[c_parent>0] faithfully models real feature hierarchy.
    This multiplicative gating (Eq. 15) is a simplified model of hierarchical dependencies; real LLM features may not follow a tree structure.
  • standard math Low-rank factor model Σ=FF^T+diag(δ) preserves the intended pairwise correlation structure.
    Eqs. 11-12 are valid for sampling but impose a factor structure on correlations.
  • domain assumption The cited LLM SAE phenomena are correctly characterized and are the right target phenomena for benchmark validation.
    The benchmark's validity rests on qualitative agreement with prior empirical results (Matryoshka disconnect, MP-SAE gap, probe underperformance, precision-recall tradeoff), several from the authors' own group.
  • domain assumption Training procedures inherited from prior work are appropriate on synthetic data.
    Appendix E prescribes training details from SAELens, Anthropic, and Gao et al.; there is no proof these are optimal for the benchmark, and some are modified ad hoc.
  • domain assumption Pythia-160m layer 10 activation norm is a reasonable reference for 'realistic' synthetic activations.
    Used to set benchmark scaling; a single LLM layer may not generalize to other models or layers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data." pith.science (2026). https://pith.science/paper/LQB2CHW5

@misc{pith2026260214687,
  author       = {Pith},
  title        = {Pith review of: SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQB2CHW5}},
  note         = {Machine review of arXiv:2602.14687}
}
read the original abstract

Improving Sparse Autoencoders (SAEs) requires benchmarks that can precisely validate architectural innovations. Current LLM-based SAE benchmarks are too noisy to differentiate architectural improvements, while commonly used synthetic-data experiments are too small-scale, unstandardized, and unrealistic to be meaningful. We introduce SynthSAEBench, a benchmark and toolkit for evaluating SAEs against large-scale synthetic data with realistic feature characteristics including correlation, hierarchy, and superposition, while providing ground-truth features and firings. SynthSAEBench acts as a controlled lower-bound test: SAE architectures that fail when the Linear Representation Hypothesis holds by construction have little hope on real LLMs. The benchmark reproduces known LLM SAE phenomena including the disconnect between reconstruction and latent quality, poor SAE probing, and a precision-recall trade-off mediated by L0, demonstrating that SynthSAEBench findings reproduce results on LLM SAEs. We further identify a novel failure mode: Matching Pursuit SAEs exploit superposition noise to improve reconstruction without learning ground-truth features, suggesting more expressive encoding procedures can easily overfit. SynthSAEBench complements LLM benchmarks with ground-truth features and controlled ablations for diagnosing SAE failure modes, while providing a clear target for SAE architecture work.

Figures

Figures reproduced from arXiv: 2602.14687 by the authors.

Figure 1
Figure 1. SynthSAEBench provides a large-scale synthetic data model with realistic feature characteristics including correlation, hierarchy, superposition and zipfian firing distributions, scalable to hundreds of thousands of features and realistic hidden dimension sizes. non-orthogonal directions, a phenomenon known as super￾position (Elhage et al., 2022). Superposition is efficient but makes interpreting activations difficu… view at source ↗
Figure 2
Figure 2. Overview of process to generate a single training activation, a. lt = argmaxiWdec,i · aˆt (5) aˆt+1 = ˆat − Wdec,lt · aˆt (6) Where aˆ0 = a. The reconstruction loss for the MP-SAE is then calculated as follows: L = ∥aˆk∥ 2 2 . (7) The variant of MP-SAEs we use in this paper does not do any early stopping based on ∥aˆ∥2 or based on selecting the same latent multiple times, as early stopping adds more complication to … view at source ↗
Figure 3
Figure 3. SynthSAEBench-16k feature firing probabilities. 0 1 2 3 Depth 0 2000 4000 6000 8000 Feature count Features by depth level [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: SynthSAEBench-16k hierarchy distribution. 4.3. Classification Metrics We evaluate each SAE latent as a binary classifier for its best-matching ground-truth feature. For latent j matched to feature i ∗ (j), we compute standard classification metrics over evaluation samp…
Figure 5
Figure 5. Figure 5: Variance explained (left), MCC (middle), and F1-score (right) for SAEs trained on SynthSAEBench-16k across varying L0 values. Shaded area is stdev with 5 seeds (too small to be visible for most SAEs). 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Precision 15 20 25 30 35 40 45 SAE L0 0.…
Figure 6
Figure 6. Figure 6: Probing precision and recall for SAEs trained on SynthSAEBench-16k across varying L0 values. Higher L0 in￾creases recall at the cost of precision. Shaded area is stdev. 6.2. MP-SAEs overfit superposition noise We next investigate the effect of superposition noise on SA…
Figure 7
Figure 7. Figure 7: Variance explained (left), MCC (middle), and F1-score (right) for SAEs trained on variants of SynthSAEBench-16k with different levels of superposition. Interestingly, MP-SAEs increase their variance explained at high superposition, implying they are able to effectively…
Figure 10
Figure 10. Figure 10: Dead latents vs L0 for SynthSAEBench-16k SAEs. Shaded area is 1 stdev with 5 random seeds. With default settings suggested by Anthropic, we see sim￾ilar numbers of dead latents in both SynthSAEBench-16k and GPT2-small SAEs. We find that changing the initial JumpReLU t…
Figure 8
Figure 8. Figure 8: Dead latent counts during JumpReLU SAE training on GPT-2 small with default settings from Anthropic [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: shows the same for SynthSAEBench-16k, both width=4096 SAEs trained on 200M tokens [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 11
Figure 11. Figure 11: Comparing JumpReLU SAEs trained on SynthSAEBench-16k using the standard auxiliary loss suggested by Anthropic (labeled “JumpReLU”) with TopK aux loss (labeled “JumpReLU (TK-aux)”). Using TopK aux loss results in a more consistently better SAE than the standard JumpReL…
Figure 12
Figure 12. Figure 12: Comparing Matryoshka SAEs trained with standard TopK auxiliary loss (TK-aux) and a Matryoshka-optimized TopK auxiliary loss (mat-aux). The matryoshka-optimized loss results in better SAEs, but the difference is especially pronounced at low L0. at low L0. We show SAE q…
Figure 13
Figure 13. Figure 13: Dead latents for Matryoshka SAEs trained with standard TopK auxiliary loss (TK-aux) and a Matryoshka-optimized TopK auxiliary loss (mat-aux). The matryoshka-optimized loss results in fewer dead latents, especially pronounced at low L0. G.3. Dead latents in Matching Pu…
Figure 14
Figure 14. Figure 14: Synthetic model throughput by number of features. I. Noise in SAEBench metrics One motivation for this work is that existing SAE bench￾marks like SAEBench, while extremely important and indis￾pensable for SAE architecture development, tend to be noisy. For instance, w…
Figure 15
Figure 15. Figure 15: SAEBench SCR, TPP, and Sparse Probing metrics for Gemma-2-2b layer 12 width 16k SAEs. 500 1000 1500 2000 2500 3000 3500 4000 Hidden Dimension 0.05 0.10 0.15 0.20 0.25 0.30 Superposition level (ρmm) Without Orthogonalization 500 1000 1500 2000 2500 3000 3500 4000 Hidde…
Figure 16
Figure 16. Figure 16: Amount of superposition (ρmm) as a function of number of features and hidden dimension with no orthogonalization (left) and with 100 iterations of orthogonalization (right) [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: Full results for SynthSAEBench-16k. Shaded area represents 1 stdev with 5 seeds per SAE. 0.00 0.05 0.10 0.15 0.20 RMS Correlation 0.775 0.800 0.825 Variance Explained 0.00 0.05 0.10 0.15 0.20 RMS Correlation 0.6 0.7 0.8 MCC 0.00 0.05 0.10 0.15 0.20 RMS Correlation 0.4…
Figure 18
Figure 18. Figure 18: Results varying correlation strength, while keeping the remaining model hyperparameters set at default values for SyntheSAEBench-16k. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 linked inside Pith

  1. [2000]

    JumpReLU

    adapts controller parameters based on operating con- ditions. The key insight is that when the system is converg- ing toward the target (error decreasing), we should reduce the controller gain to prevent overshoot. We detect conver- gence when the error and rate have opposite signs: converging= ( ¯ℓt −ℓ ∗)· ˙ℓt <0(31) where ℓ∗ is the target L0. When conve...

    2025

  2. [2024]

    Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., et al

    URL https://openreview.net/forum? id=F76bwRSLeK. Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., et al. Toy models of superposition.arXiv preprint arXiv:2209.10652, 2022. Engels, J., Michaud, E. J., Liao, I., Gurnee, W., and Tegmark, M. Not all language model features are one-...

    arXiv 2022

  3. [2025]

    Fel, T., Wang, B., Lepori, M

    URL https://openreview.net/forum? id=d63a4AM4hb. Fel, T., Wang, B., Lepori, M. A., Kowal, M., Lee, A., Balestriero, R., Joseph, S., Lubana, E. S., Konkle, T., Ba, D., et al. Into the rabbit hull: From task-relevant concepts in dino to minkowski geometry.arXiv preprint arXiv:2510.08638, 2025. Gao, L., la Tour, T. D., Tillman, H., Goh, G., Troll, R., Radfor...

    arXiv 2025

  4. [8856]

    unknown unknowns

    URL https://openreview.net/forum? id=JYs1R9IMJr. Kantamneni, S., Engels, J., Rajamanoharan, S., Tegmark, M., and Nanda, N. Are sparse autoencoders useful? a case study in sparse probing.arXiv preprint arXiv:2502.16681, 2025. Karvonen, A., Rager, C., Marks, S., and Nanda, N. Eval- uating sparse autoencoders on targeted concept erasure tasks.arXiv preprint ...

    arXiv 2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.