finding

active

finding:two-exemplars-2-3-5-7-4-11-induce-reinterpretation-of-as-addition-on-held-out-queries-across-mainstream-llms

Two exemplars (2−3=5, 7−4=11) induce reinterpretation of '−' as addition on held-out queries across mainstream LLMs

E1 qualitative finding demonstrating anchor rebinding of strong arithmetic prior

Source paper

extracted_from

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

(2025) · Edward Yi Chang · Kaya, Zeyneb N. · Ethan Chang

Neighborhood — ranked by edge-count

Claims (1)

claim

Anchors recruit and bind latent structure; they do not create new knowledge in the model
supports
Scope-limiting claim clarifying UCCT's interpretation of what anchoring does

Related by similarity (8)

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.

2-shot reinterpretation of '-' yields 23 for 15-8 on held-out queryfinding0.760
E1 qualitative: two exemplars (2-3=5, 7-4=11) cause LLMs to output 23 for 15-8.
Two-shot redefinition of "−" operator flips model output from -1 to 23 on 15-8=?finding0.745
Demonstration of strong prior rebinding via small coherent anchors
Adding a single disambiguating example (12−9=21) aligns divergent M1-M4 interpretations under tested seedsfinding0.742
E1 finding consistent with threshold-crossing: near-threshold state resolved by one additional anchor
We hypothesize that LLMs represent correctness of arithmetic expressions differently from factual statements.hypothesis0.734
Core working hypothesis motivating the factual vs. arithmetic task split in the experimental design.
Ali et al. 2025 found contrastive activation addition less effective at larger model scale, consistent with ESR in 70Bfinding0.734
Prior finding from related work that aligns with ESR being strongest in the largest model tested
Llama-3.1-8B implements a two-stage algorithm: (1) compute integer sum via base-10 addition (e.g., six + August = 14), then (2) map sum to cyclic concept space (14 → February)finding0.733
The complete mechanistic algorithm discovered for cyclic concept reasoning
Subcomponent L2.MLP.down:3382 (density 0.00%) predicts emoticon continuations after colon, semicolon, or equalsfinding0.732
Specific discovered subcomponent that activates on punctuation like ' :', ' ;', ' =', ':-' and predicts the rest of emoticons/emojis.
All cases satisfying Criteria 1 and 2 (two out of three) originate from deeper transformer layers and/or the 2/3 layer of LLMs.finding0.730
Consistent with literature that deeper layers encode semantic information and align with human brain activity.