Beyond Plain Demos: A Demo-centric Anchoring Paradigm for In-Context Learning in Alzheimer's Disease Detection

TL;DR

Alzheimer's disease detection from narrative transcripts challenges LLM-based ICL due to scarce pre-training exposure and highly homogeneous demos. The authors propose DA4ICL, a demo-centric anchoring paradigm that combines Diverse and Contrastive Retrieval (DCR) for context width with Projected Vector Anchoring (PVA) for context depth, injecting signals into demo anchors across all Transformer layers rather than the test token. Across three AD benchmarks, DA4ICL achieves large, stable gains over standard ICL and Task Vector baselines, demonstrating improved handling of subtle, out-of-distribution cues. The work highlights the importance of demo-level, layer-wise reasoning for fine-grained, low-resource adaptation in LLMs and suggests a general paradigm for other domains with similar challenges.

Abstract

Detecting Alzheimer's disease (AD) from narrative transcripts challenges large language models (LLMs): pre-training rarely covers this out-of-distribution task, and all transcript demos describe the same scene, producing highly homogeneous contexts. These factors cripple both the model's built-in task knowledge (\textbf{task cognition}) and its ability to surface subtle, class-discriminative cues (\textbf{contextual perception}). Because cognition is fixed after pre-training, improving in-context learning (ICL) for AD detection hinges on enriching perception through better demonstration (demo) sets. We demonstrate that standard ICL quickly saturates, its demos lack diversity (context width) and fail to convey fine-grained signals (context depth), and that recent task vector (TV) approaches improve broad task adaptation by injecting TV into the LLMs' hidden states (HSs), they are ill-suited for AD detection due to the mismatch of injection granularity, strength and position. To address these bottlenecks, we introduce \textbf{DA4ICL}, a demo-centric anchoring framework that jointly expands context width via \emph{\textbf{Diverse and Contrastive Retrieval}} (DCR) and deepens each demo's signal via \emph{\textbf{Projected Vector Anchoring}} (PVA) at every Transformer layer. Across three AD benchmarks, DA4ICL achieves large, stable gains over both ICL and TV baselines, charting a new paradigm for fine-grained, OOD and low-resource LLM adaptation.

Figures (10)

• Figure 1: Schematic comparison of information flow and token processing in standard ICL versus TV methods.
• Figure 2: Progressive enrichment of demo sets, from plain to wide and deep, drives more effective and robust in-context reasoning.
• Figure 3: Overview of the DA4ICL framework. Diverse and contrastive demos are selected and enriched via projected vector anchoring across all Transformer layers to provide both wide and deep context for robust AD detection.
• Figure 4: Accuracy comparison after applying DCR to ICL and TV methods (mean $\pm$ std over 10 runs). DCR consistently improves ICL performance across all datasets, demonstrating the value of contrastive demo construction. However, DCR fails to enhance conventional TV methods, with performance remaining stable or degrading, highlighting the limitations of TV’s injection granularity and alignment. Specific results can be found in Appendix: B.
• Figure 5: Ablation over two-stage retrieval strategies. Diverse main-demo selection (context width) provides the largest gains, while sub-demo enrichment (context depth) offers further complementary improvements.