Keiran Thompson

Valeria Ruscio, Keiran Thompson

We show that language models hallucinate not because they fail to detect uncertainty, but because of a failure to integrate it into output generation. Across architectures, uncertain inputs are reliably identified, occupying high-dimensional regions with 2-3$\times$ the intrinsic dimensionality of factual inputs. However, this internal signal is weakly coupled to the output layer: uncertainty migrates into low-sensitivity subspaces, becoming geometrically amplified yet functionally silent. Topological analysis shows that uncertainty representations fragment rather than converging to a unified abstention state, while gradient and Fisher probes reveal collapsing sensitivity along the uncertainty direction. Because cross-entropy training provides no attractor for abstention and uniformly rewards confident prediction, associative mechanisms amplify these fractured activations until residual coupling forces a committed output despite internal detection. Causal interventions confirm this account by restoring refusal when uncertainty is directly connected to logits.

Valeria Ruscio, Eli-Shaoul Khedouri, Keiran Thompson

Cross-entropy pretraining and preference alignment update the same transformer weights, but leave geometrically distinct traces. We characterise this asymmetry with a relative-subspace-fraction probe that tracks how weight deltas align with residual-stream activation subspaces and with the prediction subspace defined by the unembedding. Alignment deltas concentrate in the read pathway ($W_Q$, $W_K$), along principal directions of attention-input activations, while remaining near-isotropic in the write pathway ($W_O$, $W_2$) relative to the prediction subspace. We explain this pattern through anisotropic gradient accumulation: updates to a matrix $W$ are sums of outer products $δ_t a_t^\top$, and inherit directional structure from whichever side has concentrated covariance. For read-pathway matrices, this side is the input activation $a_t$, whose covariance is spiked in trained transformers and therefore produces objective-agnostic concentration. For write-pathway matrices, the relevant side is the upstream gradient $δ_t$, whose anisotropy depends on the loss. Cross-entropy supplies the canonical sharp per-sample signal, inducing write-pathway prediction geometry during pretraining; alignment objectives typically add little further write-side concentration. We support this explanation with a within-checkpoint trajectory, a graded contrastive-objective control, and a closed-form rank-1 intervention with matched direction controls, providing causal evidence for the proposed weight-space geometry.