thinker
active
thinker:ethan-y-chang

Ethan Y. Chang

Authored
1
Introduces
0
Studies
0
Affiliations
2
Cited by
0

Authored papers (1)

  • The Guanyin Protocol: A Framework for Immediately Establishing an Understanding of Both Causality and Compassion in LLM Systems Using Semantic Anchoring2025

    Semantic anchoring — the binding of a pretrained model's latent patterns to task-specific targets via external structure — predicts threshold-like performance flips with a single calibrated score S = ρd − dr − log k, where ρd measures within-cluster cohesion, dr measures prior-target mismatch, and k is the anchor budget. This formalization, called Unified Contextual Control Theory (UCCT), strictly generalizes in-context learning and recasts retrieval-augmented generation and fine-tuning as variants of the same anchoring process acting on one measurable quantity. Three controlled experiments supply evidence. Across numeral bases (base-10, base-8, base-9) at fixed computational complexity, few-shot shot midpoints follow the ordering k50(B10) = 0.28 ± 0.05 < k50(B8) = 1.83 ± 0.12 < k50(B9) = 2.91 ± 0.18, with phase widths and final accuracies (94.8%, 92.4%, 89.7%) tracking the heuristic k50 ∝ dr/ρd. On Meta-Llama-3.1-8B-Instruct, layer-wise anchoring peaks at layer 9 (S ≈ −1.90), with math/code tasks achieving S ≈ −1.65 at layers 8–12 versus commonsense at S ≈ −2.15, and the correlation between layer-wise scores and task accuracy reaches ρ = −0.73 (p < 0.001). The geometry summaries Sbmax and AUSN — the peak and normalized area of the per-layer S(ℓ) trajectory — correlate with internal few-shot midpoints θ50 across backbones (Meta-LLaMA-3.1-8B, Phi-4, Gemma-3-4B-it). UCCT implies that prompt design, retrieval filtering, and light fine-tuning are unified under a single diagnostic: compute S relative to the task-dependent critical threshold Sc to predict whether anchoring will succeed, and prescribe exactly how many additional examples or how much retrieval boost is needed to cross it.

More papers — OpenAlex / S2

Affiliations (2)

Co-authors (2)

Recent mentions (4)