Recency Bias Is Architecture, Not Capability: Axis-First Ordering and the STP Structural Minimum as Context Window Invariants

Read full text Summarise with VIRAT
Authors: Deepak Mohan
Cited by: 0
References: 2

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This note establishes that recency bias in transformer attention is an architectural property, not a capability limitation. Tested across Qwen2.5-14B and Claude Sonnet 4 at full STP-derived budget with all context units present, both models fail identically to recall a planted fact when context is recency-ordered. The failure is deterministic across three independent runs per model. Three empirical findings are reported: (1) conservative retrieval system prompts reliably worsen performance (guided 0/3, blind 2/3), (2) axis-first ordering completely resolves recency bias at adequate budget — a single placement rule producing deterministic 0/3 → 3/3 transitions, (3) the STP structural minimum formula W = 500 + (depth × 200) + (switches × 100) predicts budget failure deterministically before testing. A fourth finding introduces the ego-dissolution perspective depth pass: reclassification of SWITCH events as PUSH (perspective descent) or POP (perspective return) based on the Internal Ear Protocol's frame commitment model. At stress budget 600 tokens, depth-stratified governance keeps the axis return unit (POP) while switch-index governance keeps the first descent unit (PUSH) — structurally correct selection from different scoring logic. At stress budget 450 tokens, the lexical axis scorer degrades while both structural position scorers hold.

Fields of Study & Subjects

Topic Modeling Natural Language Processing Techniques Information Retrieval and Search Behavior

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint