Earlier uncertainty signals
Manthan / SDA
A hidden-state and attention-head geometry project testing whether spectral disagreement can flag uncertainty before the output logits make it obvious.
Core hypothesis
Uncertainty may appear as disagreement in internal geometry before it appears as entropy in final-token probabilities. Manthan/SDA tests whether that signal is measurable, efficient, and useful as a risk trigger.
Steps taken
- Built Manthan as inference-time hidden-state consensus across expert activations using Sinkhorn mixing, SVD/Karcher-style geometry, and tangent-space refinement.
- Ran ablations for Karcher vs Euclidean consensus, full SVD vs power iteration, and activation-space vs logit-space warning time.
- Built SDA variants over attention-head consensus: V1 activation overwrite, V2 z-score gating, V3 Byzantine/quorum gating, plus metrics for model-space head divergence.
- Pivoted from activation overwrite toward risk-trigger framing after analysis showed Sinkhorn/head collapse can destroy useful dispersion signal.
Results
Manthan geometry ablation: lambda2/lambda1 vs angle Pearson r=0.978, with same output token but different hidden reasoning paths.
Three-step power iteration gave 0.03 degree angle error, 100% token match, and 54.9x speedup vs full SVD.
Activation-space spectral disagreement detected uncertainty a median of +3 tokens before logit entropy; 60% of cases led by at least 2 tokens.
SDA triple ablation on Qwen2-0.5B: baseline PPL 6.366; V1 PPL ratio 1.343 with 96.0% intervention; V3 PPL 6.315, ratio 0.992, needle preserved, 3.1% consensus rate.
The current strongest direction is model-space head-divergence as a selective-generation risk feature, not activation overwrite.
Caveat
This is not a universal uncertainty detector. The viable claim is narrower: internal spectral disagreement can be an early risk signal and should be tested as a trigger for abstention or self-consistency.
Status
Research prototype; strongest as a fast internal-risk signal and a negative-result pivot away from naive interventions.