AuthorsMehrdad Farahani, Richard Johansson
TL;DR
Deciphering the Interplay of Parametric and Non-parametric Memory uses causal mediation on ATLAS to show non-parametric copying dominates parametric recall when contexts conflict.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Deciphering the Interplay of Parametric and Non-parametric Memory shows statistically different behavior between parametric and non-parametric subsets (p-value=1.60e-4, Cohen’s d=-0.9851), with non-parametric answers highly variable.
When ATLAS receives conflicting contexts like “Milan became the official capital of Sweden,” Deciphering the Interplay of Parametric and Non-parametric Memory finds ATLAS often copies the counterfactual with high probability, degrading factual reliability.
HOW IT WORKS
Deciphering the Interplay of Parametric and Non-parametric Memory combines Experiment 1, Experiment 2, causal mediation analysis, and Path Specific Effects (PSE) to isolate copying versus recall inside ATLAS.
Conceptually, Deciphering the Interplay of Parametric and Non-parametric Memory treats ATLAS like a computer where parametric weights are long-term storage and retrieved passages are a volatile clipboard that can overwrite answers.
This KEY_MECHANISM of token-level mediation tracing lets Deciphering the Interplay of Parametric and Non-parametric Memory expose when ATLAS chooses context copying over parametric recall, beyond what a plain context window reveals.
DIAGRAM
This diagram shows how Deciphering the Interplay of Parametric and Non-parametric Memory runs corrupted and restoration passes through ATLAS layers to compute indirect effects for specific tokens.
DIAGRAM
This diagram shows how Deciphering the Interplay of Parametric and Non-parametric Memory constructs synthetic contexts, filters questions, and groups behaviors into parametric versus non-parametric subsets.
PROCESS
Data Preparation
Deciphering the Interplay of Parametric and Non-parametric Memory uses PopQA and PrincetonEntityQuestion to build synthetic contexts from factual triples and query templates.
Experiment 1
Deciphering the Interplay of Parametric and Non-parametric Memory replaces object tokens with counterfactuals to measure copying via total and indirect effects.
Experiment 2
Deciphering the Interplay of Parametric and Non-parametric Memory injects noise into subject and relation embeddings to study context relevance decisions.
Path Specific Effects
Deciphering the Interplay of Parametric and Non-parametric Memory severs MLP and Attention paths to attribute copying and relevance to specific transformer modules.
KEY CONTRIBUTIONS
Examine how ATLAS uses different types of memory
Deciphering the Interplay of Parametric and Non-parametric Memory applies causal mediation analysis to ATLAS to separate parametric recall from non-parametric copying at token and layer level.
Show when ATLAS prefers one memory type
Deciphering the Interplay of Parametric and Non-parametric Memory demonstrates that ATLAS behavior resembles the non-parametric subset, with TE distributions shifting strongly toward counterfactuals when contexts change.
Identify parts crucial for copying and relevance
Deciphering the Interplay of Parametric and Non-parametric Memory finds mid layer MLP blocks and object tokens dominate copying, while subject and relation tokens drive context relevance.
RESULTS
TE parametric vs non parametric
p-value=1.60e-4
Cohen’s d=-0.9851 between parametric and non_parametric subsets
Effect size
-0.9851
vs parametric subset in TE distribution
Subject vs relation TE
p-value=3.57e-3
Cohen’s d=-6.87e-2 between subject and relation tokens
Subject relation effect size
-6.87e-2
shows similar contribution of subjects and relations to relevance
Deciphering the Interplay of Parametric and Non-parametric Memory evaluates ATLAS on PopQA and PrincetonEntityQuestion synthetic contexts, focusing on TE, AIE, and PSE rather than accuracy. These statistics show that non-parametric copying dominates behavior and that subject and relation tokens contribute comparably to relevance decisions.
BENCHMARK
Deciphering the Interplay of Parametric and Non-parametric Memory evaluates ATLAS on PopQA and PrincetonEntityQuestion synthetic contexts, focusing on TE, AIE, and PSE rather than accuracy. These statistics show that non-parametric copying dominates behavior and that subject and relation tokens contribute comparably to relevance decisions.
BENCHMARK
Relative strength of total effects (TE) and effect sizes for parametric vs non_parametric behavior and subject vs relation relevance.
KEY INSIGHT
Deciphering the Interplay of Parametric and Non-parametric Memory shows non-parametric behavior has Cohen’s d=-0.9851 versus parametric behavior, indicating strong copying from altered contexts.
This is surprising because ATLAS is a knowledge intensive RAG system, yet Deciphering the Interplay of Parametric and Non-parametric Memory finds external context can override memorized facts even with obvious counterfactuals.
WHY IT MATTERS
Deciphering the Interplay of Parametric and Non-parametric Memory gives a concrete recipe to localize copying and relevance mechanisms in RAG models using mediation and PSE.
Armed with these insights, builders can now design ATLAS like systems that explicitly modulate copying versus recall by targeting mid layer MLPs, subject tokens, relation tokens, and object token pathways.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Yijie Zhong, Yunfan Gao, Haofen Wang
· 2026
HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.
Answers use this explainer on Memory Papers.
Checking…