Query

I'm looking for prior work on a specific research problem. Please search exhaustively and report only papers that directly address it. Do not include tangentially related work, architecture proposals that merely use neuroscience vocabulary, or memory-augmented LLM systems. THE PROBLEM: Whether LLMs process in-context information through mechanisms that correspond to neuroscientifically-defined explicit memory (hippocampal/declarative: sparse coding, pattern separation, pattern completion, one-shot binding, reconstructive retrieval) versus implicit memory (basal ganglia/procedural: distributed coding, stimulus-response automaticity, gradual interference-prone associations). The target work must: 1. Take the explicit/implicit distinction from neuroscience (not abstract dual-process / System 1-2 framings). 2. Apply those criteria to diagnose or probe how LLMs internally process information given in the context window — not parametric knowledge, not external memory modules. 3. Use mechanistic analysis (probing, causal intervention, representational analysis) on LLM internals, not just behavioral benchmarks. 4. Optionally: map findings back to human cognition. EXPLICITLY OUT OF SCOPE: - Memory-augmented architectures (HippoRAG, Larimar, Memoria, Memory^3, MemoryLLM, Titans, etc.) — these add modules, they do not diagnose existing LLM processing. - Dual-process / System 1-2 framings that don't use neuroscientific memory taxonomy. - Behavioral benchmarks of implicit/explicit memory without internal mechanistic analysis (e.g., ImplicitMemBench). - Brain-LLM alignment studies (Schrimpf, Goldstein, Tuckute) — these compare representations, they don't diagnose LLM processing through the memory-systems lens. - General mechanistic interpretability work without the explicit/implicit memory framing. SEARCH PRIORITIES: - arXiv, ACL Anthology, OpenReview (NeurIPS/ICML/ICLR), bioRxiv - CCN (Cognitive Computational Neuroscience) and RLDM proceedings - NeurIPS workshops: CogInterp, NeuroAI, MemARI - Recent work (2023–2026) from groups doing neuroscience-inspired LLM analysis (Momennejad, Hasson lab, Andreas, Bau, etc.) OUTPUT: For each candidate paper, give: - Citation - One-paragraph summary of what they actually do - Explicit verdict: does it match the problem above, or is it adjacent? If adjacent, state precisely which of the four criteria above it fails. If you find nothing that fully matches, say so clearly. Do not pad the list.

Run e700d8e7 · deepest · started 5/27/2026, 11:06:27 PM