Daily Paper Briefing

Abstract Overview

This paper introduces Mindscape Activation Signature (MiA-Signature), a compact representation that approximates the global memory activation pattern induced by a query over a structured semantic memory space (the "mindscape"). The method models memory access as a two-stage process: broad activation over the memory space followed by construction of a tractable signature via submodular selection of high-level memory units that balances query relevance, coverage, and diversity. The signature can be used as a fixed conditioning signal in static RAG or maintained as an evolving global state in an iterative agent loop. The authors evaluate on four long-context benchmarks (DetectiveQA, NarrativeQA, NovelHopQA, NoCha) spanning multiple-choice QA, open-ended QA, multi-hop QA, and claim verification. Results indicate that MiA-Signature primarily improves evidence selection at retrieval time, while its benefit at answer generation time is more task- and model-dependent.

Novelty

The distinctive contribution is modeling LLM memory access through query-induced global activation over a structured memory space and then approximating that activation with a compact signature built via submodular selection, rather than relying on local retrieval alone. The work also presents a unified signature interface that functions both as a one-shot conditioning signal in static RAG and as an evolving memory state refined iteratively in agent loops.

Results

In static RAG, conditioning retrieval on MiA-Signature improves average Recall@10 by 10.9% and average task performance by 3.8% under the same retriever and generator backbone (MiA-Emb / DS-V3.2). In the iterative agent setting, MiA-Agent improves retrieval recall on every benchmark with retrieval annotations relative to the no-signature agent, with especially clear gains on DetectiveQA-ZH and NovelHopQA. Answer-time ablations show the signature is a more reliable aid for retrieval than for generation, while combining signature and accumulated evidence is particularly helpful on NoCha (71.4 PairAcc vs. 57.1 without any memory state).

Key Points

1. MiA-Signature is defined as a compact surrogate of a query-induced global activation pattern over a structured memory space, constructed via submodular selection of high-level memory units for relevance, coverage, and diversity, rather than serving as a simple summary replacement for retrieved evidence.
2. The method provides a unified memory interface used both as a one-shot conditioning signal in static RAG and as an evolving global state refined iteratively alongside query rewriting and evidence accumulation in an agent loop.
3. Empirically, the approach yields consistent retrieval improvements across four long-context benchmarks in both static RAG and agent settings, while generation-side gains depend on the task and the generator's capacity to exploit global context.

References

• arXiv: https://arxiv.org/abs/2605.06416v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2605.06416v1
• Hugging Face Papers: https://huggingface.co/papers/2605.06416