thinker
active
thinker:openalex-A5092206144

Daniel Wurgaft

Authored
9
Introduces
0
Studies
0
Affiliations
2
Cited by
3

Authored papers (9)

  • Llama-3.1-8B solves cyclic arithmetic (e.g., "what month is six months after August?") not by performing modular addition in the period of the cyclic concept (12 for months, 7 for days of the week) as its circularly structured representations might suggest, but by reusing a single generic base-10 addition mechanism across all such tasks. The mechanism operates in two stages: first, the model computes the raw integer sum of its inputs (six + August = 14), then maps that sum back into cyclic concept space (14 → February). The paper introduces a combined causal-abstraction and feature-geometry analysis to demonstrate that Llama-3.1-8B uses task-agnostic Fourier features with periods characteristic of base-10 arithmetic—specifically periods 2, 5, and 10—rather than the concept-native period. Strikingly, a sparse set of only 28 MLP neurons at layer 18 (approximately 0.2% of that layer's MLP) is responsible for this computation across all cyclic tasks and can be partitioned into disjoint clusters, each computing the sum for a Fourier feature of a distinct period. This implies that structured representational geometry does not entail correspondingly structured computation: a model can build domain-specific circular embeddings and yet delegate the actual arithmetic to a domain-general, base-10 calculator reused opportunistically, suggesting that mechanistic interpretability analyses that read off algorithms from representation geometry alone will systematically mislead.

  • Sparse autoencoders trained on Llama3.1-8B (layer 19, 500M Pile tokens, 25 SAEs across Standard, JumpReLU, TopK, BatchTopK and Matryoshka architectures at VE > 0.85) do not capture continuous concept manifolds as compact subspaces; they spread them across support sets well beyond the manifold's ambient dimension, a fragmented regime the paper calls dilution. The theory separates two ways an SAE can capture a manifold: globally, with a small atom group whose span contains it (Theorem 1, requiring μ < 1/(2k−1) incoherence), or locally, by tiling it with localized features. Tiling splits into shattering (mutually exclusive atoms) and dilution (redundant, mixed-selectivity atoms). A synthetic benchmark of 48 manifold instances (8 types, embedded in R^128, dictionary c=512) shows a sparsity sweet spot near k=4 with shattering below it and dilution above. On Llama, restricted R² plateaus past the ambient dimension and features show tuning curves, for example ones-digit and decade features on the years manifold. The instrument is an Ising-coupling pipeline: a pairwise Ising model fit on binarized codes, then Louvain communities and a PCA spectral-gap check. It recovers temperature, colors and political bias and surfaces a new epistemic-uncertainty manifold, while decoder cosine and Pearson correlation fail. Low reconstruction error therefore does not mean coherent geometry, and single-feature interpretation, steering and auto-interp miss it. The paper argues the unit of interpretability should be geometric objects, meaning groups of atoms, and that featurizers should target manifolds directly.

  • Manifold steering — intervening along activation-space paths defined by the intrinsic geometry of a fitted activation manifold M_h rather than along Euclidean linear directions — produces behavioral trajectories that remain on the behavior manifold M_y, whereas standard linear steering cuts through off-manifold regions and generates unnatural outputs. Across language model reasoning tasks with cyclic, sequential, and graph geometries, and in a video world model whose task geometry corresponds to physical dynamics, the bidirectional relationship M_h ↔ M_y holds: steering that respects M_h tracks M_y, and optimizing interventions in activation space to follow M_y recovers activation trajectories that trace the curvature of M_h. The method fits two manifolds — one to intermediate representations, one to output probability distributions — and then uses geodesic-like paths on M_h as the intervention substrate rather than a single steering vector. This work argues that neural representation geometry is not incidental but is the causal structure organizing model behavior, which implies that the core problem of model steering should be reframed from finding the right direction to finding the right geometry.

  • Manifold steering — intervening on model activations along paths constrained to lie on a learned activation manifold M_h rather than along Euclidean linear directions — produces behavioral trajectories that track the corresponding behavior manifold M_y, while linear (Euclidean) steering cuts through off-manifold regions and generates unnatural outputs. The paper fits M_h to internal representations and M_y to output probability distributions, then tests their bidirectional correspondence via controlled interventions across language models and a video world model. In language models, tasks with cyclic geometries, sequential geometries, and complex graph geometries (in-context learning) all show that manifold-constrained interventions keep behavioral outputs on M_y, whereas linear steering deviates measurably. In a video world model, interventions shaped by physical-dynamics geometry similarly respect M_y. Crucially, the relationship is bidirectional: optimizing interventions in activation space to produce paths along M_y recovers activation trajectories that trace the curvature of M_h. This implies that the core problem of mechanistic steering should be recast not as finding the right direction in a flat Euclidean activation space, but as identifying the correct geometric structure — because representational geometry is not incidental to model behavior but is causally constitutive of it.

  • Llama-3.1-8B-Instruct's in-context belief updates while reading a story trace smooth trajectories on a low-dimensional manifold. That manifold is shared between behavior and residual activations, and its geometry predicts how steering spills across concepts. Behavioral beliefs are elicited as probability-weighted 0-10 ratings after each sentence for six concepts per domain (Emotions, Genres, Arbitrary), using 2,500 SimpleStories training stories and a 500-story test set. UMAP/PCA fits on max-activating sentences give manifolds M_y (behavior) and M_z (activations). Concept-centroid distance matrices correlate at r=.92 for Emotions and r=.89 for Genres across M_y and M_z, against r=.71 for the Arbitrary control. The inferred hierarchies match exactly between the two manifolds, and the Emotions geometry correlates r=.93 with human valence-arousal ratings. Layer-9 linear probes predict behavior at RMSE .09 (Emotions, Genres) and .11 (Arbitrary), even though the prompt never names the concepts. Steering entanglement, for example sadness steering raising anger, tracks manifold distance (r=.65 Emotions, r=.42 Genres). The paper introduces the conceptual belief space, which lifts Gärdenfors' conceptual spaces to distributions over concepts, so that ICL trajectories become paths through it. Single-layer probe steering is undone by later layers, and a 7-layer span fixes this. The implication is that Bayesian accounts of ICL on freeform text can be grounded in structured conceptual geometry. Steering and interpretability can then use that geometry directly, instead of treating concepts as independent directions.

  • Lifting the featurizer's atom from a direction to a block of directions recovers concept manifolds that SAEs shatter, and block sparsity is the MAP-matched prior for a representation that is a sparse Minkowski sum of low-dimensional manifolds. The paper introduces block-sparse featurizers (BSFs) in three variants: Vanilla (top-k blocks by norm, untied encoder), Grassmannian (tied encoder with orthonormal charts), and Group Lasso (block soft-threshold). On a synthetic manifold-superposition toy, all three reach per-block R² of 0.93–0.97 against an oracle of 0.99, while a TopK SAE collapses to 0.53. On DINOv3 ViT-B, a minimum-description-length analysis at the 20% distortion floor shows blocks of dimension 2–4 describe activations in fewer bits than b=1 SAEs at every dictionary width and sparsity, and stable rank saturates near 3 even when b=16 is allowed. In InceptionV1, the curve-detector neurons and SAE atoms are fragments of one curve manifold, whose first three Fourier harmonics carry 89% of variance (59/18/12%) and survive a matched null by factors of 470 and 314. A DINOv3 block tracks sun luminance and shadow across twelve Blender models, and SDXL blocks support manifold steering, as in the pretzel block. The implication is that a featurizer is a hypothesis about the data-generating process, so interpretability should infer what the atoms must be rather than assume them. Blocks are one answer, matched to vision but probably not to language.

  • Larger models learn tasks smaller models never learn, and the cause is data-induced competition for neurons, not expressivity: small models spend capacity on frequent or simple tasks, and their updates overwrite rare-task features before the next rare observation arrives. The argument starts from power-law scaling. With a compute-optimal exponent γ=0.34 and an asymptotic parameter exponent α≈0.46, α>γ implies that a smaller model's infinite-data loss still exceeds a larger model's finite-compute loss. The paper formalizes this as 'learnable via data scaling' versus 'learnable via model scaling' (Definitions 1–2). In a synthetic mixture of K=32 linear-regression tasks, Theorem 3 shows a width-N encoder keeps the N features with the highest utility π_kλ_kj. Theorem 4 and Proposition 6 show that once common-task residual falls, rare-task features stop being overwritten. A matched-frequency injection protocol isolates retention from frequency. The same pattern appears in OLMo models (4M–4B parameters) pretrained on Dolma v1.7 with injected comparison and modular-addition tasks at frequencies from 7.8×10⁻³ to 2.4×10⁻⁸. Only larger models learn the rare tasks. They carry more task features, such as a 1-D global-order subspace with 96% DAS interchange success. They also show near-orthogonal non-task gradients (7.58×10⁻⁵ at 1B versus 0.10 at 20M). The implication is that scaling is partly a data-mixture problem: raising a target task's frequency may substitute for adding parameters.

  • In-context learning and activation steering are the same operation on a latent-concept belief: context adds log-likelihood evidence while a steering vector adds a constant shift to the log prior odds, so the two sum in log-odds space. A four-parameter closed-form model (α, γ, a, b), log o(c|x) = b + γN^(1−α) + a·m, predicts Llama-3.1-8B-Instruct behavior on held-out steering magnitudes at r = 0.98 averaged over five persona datasets (Machiavellianism, Psychopathy, Narcissism, and two moral-nihilism sets, from Perez et al. 2022). It also predicts the crossover shot count N* at r = 0.97. Fitting is by L-BFGS with 10-fold cross-validation over steering magnitude, using Contrastive Activation Addition vectors applied at layer 12. The same model holds for Qwen-2.5-7B and Gemma-2-9B (N* r = 0.91 and 0.97), Llama-3.1-70B (r = 0.98), and a flipped-label financial sentiment task (r = 0.96). Posterior belief is sigmoidal in N^(1−α), which explains the sudden many-shot jailbreaking curves of Anil et al. (2024). Steering moves the curve left or right, and the joint (shots × magnitude) space splits into sharp phases whose boundary is computable. Steering is linear in log-odds only within a bounded range of m, beyond which behavior collapses to chance. The result implies that prompting and steering are interchangeable levers on one belief state. Small changes in either control can therefore flip behavior abruptly, and N* can be forecast for safety purposes.

More papers — OpenAlex / S2

Affiliations (2)

Co-authors (12)

Recent mentions (9)