Across the full Pythia suite (160M-12B, eight checkpoints, four task families), a linear probe can read a target variable from the residual stream as early as step 1,000 at every scale -- yet steering along that same reading direction remains null-equivalent in 43 of 48 model-checkpoint cells. Internal readability systematically outruns causal efficacy, and the lag does not shrink with scale. We call this structure lagged coupling and decompose it into three dissociable tracks: (i) internal readability, saturated (AUROC >= 0.990) from the first checkpoint everywhere; (ii) behavioral readability, which develops gradually and progressively later at larger scales (12B reaches 0.909 only at the final checkpoint); (iii) causal efficacy, almost always null-equivalent, occasionally counterproductive early, with one isolated positive pulse (12B, step 8,000, z = +2.49) our grid cannot resolve. The ordering is dominantly read-before-write (11/11 units, no inversion). Representation headroom along the probe direction grows up to 57x with training and scale while causal write-in stays below 0.11% of headroom -- the variable is increasingly written into the representation and increasingly ignored by the readout. Under a fully pre-registered protocol, both single-onset hypotheses resolve INDETERMINATE (scale slope +0.24, 95% CI [-0.60, +0.87]; time vote 3:3) -- a disciplined negative explained by the three-track decomposition. A pre-registered OLMo-2 replication preserves the direction at attenuated magnitude. Our results caution against inferring steerability from probe accuracy and establish a developmental bottleneck: representation formation reliably outpaces causal readout consolidation.
Shalom Kachko, Raz Lapid, Margarita Vald +2cs.AI cs.CL cs.CV
Vision-language models (VLMs) process image patches and text tokens in a shared residual stream, but the local geometry through which the two modalities interact remains poorly understood. Most interpretability methods identify global linear directions, which may miss representations that are globally high-dimensional but locally low-dimensional. We introduce LENS (Local Explanation of Neighborhood Subspaces), a method that decomposes VLM activations into local low-rank Gaussian neighborhoods using a Mixture of Factor Analyzers. Applied to LLaVA-1.5-7B and Qwen3-VL-8B, LENS reveals distinct depth-dependent fusion trajectories consistent with each model's fusion mechanism: LLaVA progressively mixes modalities at later layers, whereas Qwen3-VL mixes them early, partially re-segregates them, and recombines them near the output. An automated multimodal labeling pipeline assigns concise semantic descriptions to these neighborhoods. Interpolating activations toward neighborhood centroids causally redirects generation within and across modalities and outperforms difference-in-means and VL-SAE in most evaluated conditions; in one LLaVA vision-to-vision setting, MFA achieves 5.7 times the VL-SAE score. Human evaluation finds MFA steering competitive with prompting and substantially stronger than the other intervention baselines. Finally, the MFA coefficient space improves Qwen3-VL image-to-rendered-text retrieval at the deepest evaluated layer from 14.9% to 48.6% R@1. Ablations show that the reported fusion trajectories are stable across component counts, local ranks, and modality-purity thresholds. These results support local geometric neighborhoods as useful interpretable and causal units for analyzing cross-modal representations in the evaluated VLMs.
Wittgenstein's duck-rabbit poses a question for vision-language models: when a model captions an ambiguous image, where in the model is the commitment to one aspect made? We address this with a 3,320-generation behavioral baseline over 83 bistable stimuli that surfaces three regimes (default-dominant, force-dominant, force-balanced) under neutral vs forced-choice prompting, then probe the underlying representations using a TopK sparse autoencoder we train on the CLIP layer that LLaVA-1.6-7B actually consumes (validation EV 0.93). Across 69 bistable stimuli with both per-aspect feature pools available, 72% (50/69) show simultaneous activation of both pools at the vision tower, including 12/12 default-dominant duck/rabbit and 7/8 force-balanced young/old. Causal steering at CLIP layer 22 flips captions on default-dominant stimuli (33% rabbit-flip rate under a fluency guard) but cannot flip captions on force-balanced young/old at any tested coefficient, despite their vision-side superposition. The dominance bottleneck lives downstream of the vision tower; the gap between vision-side representation and language-side commitment is an empirical handle on the seeing/seeing-as distinction. We also flag a methodological note: rank-based statistics on TopK SAE outputs require tie-corrected ranking to avoid silent row-order bias.