LoRA+CoT

Fine-tuning with chain-of-thought rationales aiming to reduce dr via procedural alignment.

Neighborhood — ranked by edge-count

concept

Fine-tuning
implements
Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
Prior-Target Mismatch (dr)
associated_with
Measures how far the target PT is from the prior P_prior; increases anchoring difficulty

method

E2: Numeral-Base Arithmetic Controlled Study
uses
Quantitative study varying representational familiarity via numeral bases B10/B8/B9 at fixed computational complexity

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Short rationales (LoRA+CoT) sometimes improve in-distribution performance but do not reliably reduce cross-base harmfinding0.752
E2 finding showing CoT's limited benefit for OOD transfer, consistent with larger dr out of scope
LoRA Adaptersmethod0.741
Parameter-efficient fine-tuning method used in both distillation and introspection stages
LoRA adapterconcept0.739
Fine-tuning technique that can effectively add a constant vector to the residual stream, equivalent to steering along a persona direction
LoRA (Low-Rank Adaptation)method0.738
Parameter-efficient fine-tuning method used for both SDF and expert iteration stages.
LoRA Fine-Tuningmethod0.731
Adaptation method used via Tinker API for DeepSeek-V3.1 and Qwen3-235B fine-tuning with rank 32
LoRA Fine-Tuning with Axolotlmethod0.717
Specific fine-tuning implementation using LoRA rank 32, learning rate 2e-4, AdamW 8-bit optimizer
Low-Rank Adaptation (LoRA)method0.711
Parameter-efficient fine-tuning method used to implement SOO fine-tuning on LLMs
LoRA SFTmethod0.704
Light fine-tuning method used in E2 to reduce mismatch dr.