PaperlayerView exact revision on arXiv

Source-linked AI summary

MiA-Signature: Approximating Global Activation for Long-Context Understanding

Yuqing Li, Jiangnan Li, Mo Yu, Zheng Lin, Weiping Wang, Jie Zhou

arXiv:2605.06416v1cs.CL

TL;DR

Cognitive evidence suggests that global activation supports reportable access while only a compact approximation is accessible. The paper introduces MiA-Signatures to represent query-induced activation and reports consistent gains across long-context understanding tasks in RAG and agentic systems.

  • Problem

    Cognitive evidence links reportable conscious access to distributed global activation, while people cannot directly access or enumerate all activated contents.

  • Method

    MiA-Signatures compress query-induced activation by selecting high-level concepts covering the activated context space, optionally refining them with lightweight working-memory updates.

  • Results

    10.9% average R@10 and 3.8% average task performance gains accompanied consistent improvements across multiple long-context understanding tasks.

  • Takeaways & Limitations

    MiA-Signatures provide a tractable interface between broad memory activation and downstream LLM computation without replacing local evidence or requiring full-state access.

  • Takeaways & Limitations

    Experiments center on literary and narrative domains, and transfer to code, scientific literature, or multimodal interaction remains untested.

Abstract

from arXiv · show

A growing body of work in cognitive science suggests that reportable conscious access is associated with \emph{global ignition} over distributed memory systems, while such activation is only partially accessible as individuals cannot directly access or enumerate all activated contents. This tension suggests a plausible mechanism that cognition may rely on a compact representation that approximates the global influence of activation on downstream processing. Inspired by this idea, we introduce the concept of \textbf{Mindscape Activation Signature (MiA-Signature)}, a compressed representation of the global activation pattern induced by a query. In LLM systems, this is instantiated via submodular-based selection of high-level concepts that cover the activated context space, optionally refined through lightweight iterative updates using working memory. The resulting MiA-Signature serves as a conditioning signal that approximates the effect of the full activation state while remaining computationally tractable. Integrating MiA-Signatures into both RAG and agentic systems yields consistent performance gains across multiple long-context understanding tasks.

1 Introduction

The paper models memory access as global activation over a semantic mindscape followed by a compact, usable representation. It introduces MiA-Signatures, constructed by submodular concept selection with optional iterative refinement, and integrates them into RAG and agentic systems for long-context understanding.

  • Motivation: Cognitive evidence motivates modeling memory access as global ignition followed by a compact representation of partially accessible activated contents.Global ignition is described as large-scale activation over distributed memory systems, while people cannot directly enumerate all activated contents.
  • Method: MiA-Signature is a compressed, query-induced activation pattern over a global semantic memory space called the mindscape.It provides a compact query-conditioned global state for retrieval, generation, and agentic memory.
  • Method: The construction selects high-level concepts submodularly to cover activated context, with optional lightweight iterative refinement using working memory.The selected concepts form a tractable representation that guides downstream computation.
  • Evaluation: MiA-Signatures yield consistent performance gains across multiple long-context understanding tasks when integrated into RAG pipelines and agentic systems.The results suggest that approximating global activation is a more effective memory interface than relying solely on local retrieval.

2 Related Work

Existing theories associate conscious access with globally distributed activation but leave open how this activation becomes a usable representation for downstream computation. MiA-Signature addresses this gap by proposing a compact, query-conditioned representation that approximates global activation and guides retrieval and reasoning.

  • Global workspace and global ignition: Global Workspace Theory links conscious accessibility to broadcasting information across distributed cognitive modules, while Global Neuronal Workspace theory associates conscious access with nonlinear global ignition.These frameworks motivate modeling cognition through distributed global activation.
  • Limits of access and partial awareness: Recurrent Processing Theory and empirical studies of partial awareness suggest that globally activated representations are not all accessible or directly enumerable as reportable contents.This establishes a gap between global activation and its cognitively available form.
  • Integration and compression of global states: Integrated Information Theory characterizes conscious states as integrated and structured, supporting the view that global activity can be represented in compressed form.IIT differs theoretically from GNW but likewise emphasizes representations reflecting global structure.
  • From global activation to usable representations: Existing theories do not specify how distributed activation guides downstream computation, while typical RAG systems rely on local retrieval from a small set of documents.The paper identifies this as a limitation of direct retrieval-based memory access for long-context reasoning.
  • Our perspective: MiA-Signature is a compressed representation of a query-induced global activation pattern over semantic memory, used as a compact signal for retrieval and reasoning.The framework treats memory access as a two-stage process beginning with global activation rather than direct retrieval.

3 Method

MiA-Signature is a compact, query-conditioned global state that approximates the activated region of a long-context memory pool. The method constructs this signature through coverage-aware high-level memory selection and uses it as a conditioning signal in both static RAG and iterative agent retrieval.

  • 3.1 Mindscape and MiA-Signature: MiA-Signature selects high-level memory units that jointly approximate query-induced activation through relevance, coverage, and redundancy avoidance.It represents the activated global state rather than a shortened source summary and coexists with locally retrieved evidence.
  • 3.1 Mindscape and MiA-Signature: The mindscape organizes fine-grained source evidence into redundant, overlapping memory units at multiple abstraction levels.This structure motivates representing the globally relevant region compactly rather than relying on the full memory pool.
  • 3.2 Static RAG: In static RAG, E2 retrieves evidence conditioned jointly on the query and fixed signature, scoring candidates as (1 − α) query relevance plus α signature consistency.The signature changes the retrieval interface but does not replace retrieved evidence, and may also condition generation.
  • 3.2 Agent Setting: In the agent setting, E2 repeatedly retrieves with the evolving pair (qt, σt), while updates jointly refine the query, local evidence memory, and global signature.The agent answers when its stopping decision triggers generation or when the refinement budget is exhausted.
  • 3.2 Signature Construction: The initial signature retrieves K0 = 50 fine-grained candidates with E1, maps them to high-level memory units, and applies greedy coverage-aware selection instead of First-K truncation.The objective balances query relevance, activated-region coverage, and diversity among selected units.

4 Experiments

Experiments evaluate MiA-Signatures in static RAG and iterative-agent settings across four long-context benchmarks, measuring task performance and retrieval recall. Results show consistent retrieval benefits, while answer-time gains depend on the task and generator.

  • Evaluation: Four benchmarks cover multiple-choice QA, open-ended QA, multi-hop QA, and claim verification over long-context narrative documents.DetectiveQA, NarrativeQA, NovelHopQA, and NoCha evaluate complementary long-context capabilities.
  • Evaluation: Evaluation uses accuracy for multiple-choice QA, F1 for open-ended QA, accuracy and pair accuracy for NoCha, plus Recall@10 where gold evidence is available.These metrics distinguish answer quality from evidence retrieval quality.
  • RQ1: Static RAG: 10.9% average R@10 and 3.8% average task-performance gains result from conditioning retrieval on the MiA-Signature under the same retriever and generator backbone.The generator still receives retrieved chunks only, so the improvement comes from evidence selection before generation.
  • RQ2: Iterative Agents: MiA-Agent improves retrieval recall over Agent w/o Sig. on every benchmark with retrieval annotations, extending the signature’s benefit from one-shot retrieval to iterative search.The signature functions as an evolving global state that keeps iterative retrieval aligned with the query-induced activated region.
  • RQ3: Answer-Time Memory: Answer-time use of the signature is selective: MiA-RAG improves over MiA-Emb, while the Gen-14B variant achieves the best NarrativeQA F1 without dominating across benchmarks.The signature summarizes broader activated context, whereas evidence memory preserves grounded facts accumulated during agentic retrieval.

5 Conclusion … A.2 Objective Formulation

MiA-Signature is introduced as a compact, tractable representation of query-induced global activation for conditioning downstream LLM computation. Its construction selects cached session summaries with submodular objectives combining relevance, coverage, and diversity for static RAG or a relevance-only initialization for online agent refinement.

  • 5 Conclusion: MiA-Signature compresses query-induced global activation into a tractable interface for downstream LLM computation across static RAG and agentic systems.The paper reports improved access to external memory across inference settings and long-context benchmarks, but the supplied passage truncates the detailed results.
  • A Submodular Initialization: Coverage-aware vs. First-K: The initialization framework offers Coverage-aware submodular and First-K submodular variants, assigning the former to static RAG and the latter to agents with online refinement.Coverage-aware selection uses relevance, chunk-level coverage, and diversity; First-K is a modular relevance-only variant.
  • A.1 Problem Setup and Motivation: Given ranked candidate chunks and their session summaries, initialization selects A ⊆ S with |A| ≤ K whose concatenation forms the initial signature.Distinct chunks may share a session summary, so the number of summaries is typically much smaller than the number of chunks.
  • A.1 Problem Setup and Motivation: Session summaries are created before querying by summarizing non-overlapping source-order windows of W = 20 chunks with one fixed GPT-4o call per window.The summaries are cached and reused across queries, so construction adds no query-time LLM calls.
  • A.1 Problem Setup and Motivation: The deterministic source-order mapping assigns each retrieved chunk to its cached session summary, with the same construction settings used across datasets.The high-level memory set is instantiated as H(D) ≡ S, and the mapping is fixed at indexing time.
  • A.1 Problem Setup and Motivation: Coverage-aware selection improves on taking summaries associated with the top-K chunks by reducing redundancy and addressing missed near-cutoff summaries and summary-level misalignment.The method adds chunk-level coverage and diversity to the relevance-only objective.
  • A.2 Objective Formulation: The combined marginal gain weights max-normalized query relevance and chunk coverage together with diversity: Δ(s | A) = λQ Δf̃Q(s | A) + λC Δf̃C(s | A) + λD fD(s | A).Coverage uses rank-decaying wi = 1/(ri + 1) and semantic match m(s, ci), which checks that selected summaries genuinely reflect candidate chunks.

A.3 Submodularity Analysis · A.4 Algorithmic Procedure

The query-relevance and coverage components support monotone submodular greedy selection, while the diversity term intentionally breaks global monotonicity and submodularity. Algorithm 2 implements coverage-aware greedy selection with cached BGE-M3 embeddings and complexity O(K · N · M) after embedding computation.

  • A.3 Submodularity Analysis: fQ is modular because it sums element-wise query-relevance quantities, and therefore is trivially submodular and monotone.
  • A.3 Submodularity Analysis: fC is monotone submodular because it forms a weighted MAX-COVERAGE objective whose non-negative maximum terms preserve both properties.
  • A.3 Submodularity Analysis: The diversity term fD is not monotone, so the full objective is not globally submodular; it serves mainly as a tie-breaker among candidates with similar fQ/fC profiles.The (1 −1/e) approximation guarantee remains valid for the fQ + fC sub-objective when λD = 0.
  • A.3 Submodularity Analysis: (1 −1/e) ≈0.632 is the greedy approximation bound for the dominant monotone component fQC(A) = λQ ˜fQ(A)+λC ˜fC(A).The guarantee applies when the diversity term is ignored and A∗ is the optimal size-K subset.
  • A.4 Algorithmic Procedure: Embeddings are computed once per call in a single batched forward pass through BGE-M3 and then cached.BGE-M3 uses batched CLS embeddings with ℓ2-norm, while query-summary similarities are pre-computed before selection.
  • A.4 Algorithmic Procedure: O(K · N · M) set operations are required after the embedding pass, which is negligible relative to the retriever call itself.The BGE-M3 model is loaded once per process and cached, so disabling submodular selection adds no overhead.
  • A.4 Algorithmic Procedure: Algorithm 2 greedily maximizes Eq. 10 by selecting up to K summaries using query relevance, coverage, and diversity marginal gains.The procedure computes candidate and summary embeddings, normalizes relevance and coverage scores, tracks covered chunks, and returns the selected set.

A.5 Coverage vs. First-K Initialization · B Retriever Mechanism

The paper compares coverage-aware submodular initialization with First-K initialization under fixed system components, finding modest but consistent gains from coverage-aware selection. It also distinguishes query-only retrievers from the mindscape-aware MiA-Embedding retriever, which uses both the query and signature.

  • A.5 Coverage vs. First-K Initialization: A.5 Coverage vs. First-K Initialization: All main results use submodular initialization, while the ablation compares it with First-K under the same retriever, generator, and refinement budget.First-K selects session summaries associated with the top-K chunks from the step-0 query-only ranking.
  • A.5 Coverage vs. First-K Initialization: A.5 Coverage vs. First-K Initialization: First-K and coverage-aware submodular share the same step-0 query-only candidate pool, leaving the initial evidence frontier unchanged.Their difference lies in which high-level summaries are selected.
  • A.5 Coverage vs. First-K Initialization: A.5 Coverage vs. First-K Initialization: Coverage-aware submodular yields a small but consistent improvement in average R@10 across all three signature-based variants.It also improves average task performance in each case.
  • A.5 Coverage vs. First-K Initialization: A.5 Coverage vs. First-K Initialization: The clearest coverage-aware gain appears on NarrativeQA, where activated context is broad and redundant.Selecting summaries for chunk coverage is more useful there than simply taking the first K ranked summaries.
  • A.5 Coverage vs. First-K Initialization: A.5 Coverage vs. First-K Initialization: Effects are smaller and sometimes mixed on DetectiveQA, NovelHopQA, and NoCha, where narrower activation or precise local distinctions can favor First-K.The ablation characterizes coverage-aware terms as a modest but reliable improvement for static RAG.
  • B Retriever Mechanism: B Retriever Mechanism: The system uses query-only retrievers that encode the query with conventional last-token pooling, including SFT-Embedding and Qwen3-Embedding.These retrievers search chunks using the query alone.
  • B Retriever Mechanism: B Retriever Mechanism: The mindscape-aware MiA-Embedding retriever takes both the query and the signature as input following Eq. 4.It captures two sources of information via interpolation.

C Dataset Construction and Statistics · C.1 Series Aggregation Details · C.2 Single-Book vs. Series-Book Control Details

The paper constructs coherent multi-volume series from existing DetectiveQA and NarrativeQA books, then controls for indexing granularity by comparing single-book and merged series-book retrieval under identical pipeline conditions.

  • C.1 Series Aggregation Details: DetectiveQA aggregates 13 novels into the Miss Marple and Hercule Poirot series.This aggregation forms coherent multi-volume series from the original benchmark books.
  • C.1 Series Aggregation Details: NarrativeQA aggregates 37 books into 11 series based on sequential arcs or shared protagonists.The resulting series are designed around narrative continuity or recurring protagonists.
  • C.1 Series Aggregation Details: Table 4 summarizes the aggregated-series statistics for DetectiveQA and NarrativeQA.The table is presented as the summary of the series aggregation details.
  • C.2 Single-Book vs. Series-Book Control Details: The control tests whether series-book indexing is harder than single-book indexing while holding all other pipeline components fixed.This is framed as a retrieval-side sanity check.
  • C.2 Single-Book vs. Series-Book Control Details: Both indexing conditions use the same query-only retriever, SFT-Emb-8B.The retriever is held constant across the comparison.
  • C.2 Single-Book vs. Series-Book Control Details: Single-Book indexes only the gold book, whereas Series-Book merges all books in a series into one document and retrieves from the merged index.Questions are answered against their own book in the Single-Book condition.

C.3 On Full Global-Summary Baselines

The paper distinguishes prior MiA-RAG’s document-level global-summary conditioning from its query-induced MiA-Signature. It excludes the full global-summary baseline in series-book retrieval because merged summaries create semantic interference, while gold-book summaries leak the target.

  • Baseline distinction: Prior MiA-RAG conditions retrieval and generation on a pre-existing document-level global summary, whereas this paper uses a query-induced MiA-Signature.The conditioning mechanism is the key difference between the prior baseline and the system evaluated here.
  • Baseline exclusion: The full global-summary baseline is omitted for series-book retrieval because merged summaries mix plots, characters, and events, while gold-book summaries reveal the target book.The series-book setting instead tests identification of the query-relevant region within an overcomplete memory space.
  • Series-book control: Merging related books lowers query-only retrieval recall, indicating that cross-book context introduces semantic interference rather than useful side information.This control motivates evaluating the main experiments under the harder series-book setting, before chunk-level matching.

D Query-Rewrite Ablation

The ablation tests whether the query should be rewritten at each retrieval step alongside the evolving activation state. Rewriting generally helps narrow refinement searches, but NovelHopQA benefits from keeping the query fixed to preserve parallel evidence paths.

  • Ablation setup: The ablation compares final performance with query rewriting enabled versus disabled across four answer-time interfaces.The retrieved evidence representation σt continues evolving at each step in both conditions.
  • When rewriting helps: Rewriting most clearly improves NarrativeQA and NoCha by converting partial evidence into more specific follow-up queries.Its benefit is tied to refinement settings where narrowing the search is desirable.
  • When rewriting hurts: NovelHopQA achieves higher F1 with a fixed query because multi-hop questions benefit from preserving parallel evidence paths.Accordingly, the main experiments keep qt fixed on NovelHopQA and rewrite it elsewhere.

E Case Study · F Prompt Templates

The DetectiveQA case study shows that MiA-Signature resolves a cross-context identity binding that defeats local retrieval, while the prompt templates operationalize signature-guided refinement, evidence accumulation, and task-specific answering.

  • E Case Study: MiA-Agent answers option D because signature refinement binds the hostess to Charlotte-as-Letitia, the agent of the pill swap.The binding is encoded after refinement, enabling later retrieval to promote the identity-bound reveal; confidence changes from MEDIUM to HIGH over two refinement steps.
  • E Case Study: MiA-Emb, MiA-RAG, and the agent without a signature all predict B because they fail to preserve the required global identity binding.MiA-Emb lacks a signature, MiA-RAG has not bound the hostess role to Charlotte-as-Letitia, and the unsigned agent carries no global signature across steps.
  • E Case Study: The initial signature activates both the identity and poisoning subplots but does not yet bind the hostess identity to the pill substitution.Query-only retrieval produces |H0(q)| = 10 candidate summaries, and First-K submodular selection retains K = 5.
  • E Case Study: At Step 1, the agent refines rather than answers by rewriting the query to identify who substituted the pills and which poison caused Dora’s death.The refined query is: “Who substituted the aspirin pills that Dora Bunner took, and what specific poison caused her death?”
  • E Case Study: At Step 2, signature-guided retrieval identifies aconite as the specific poison and supports the causal explanation that Charlotte substituted the poisoned pills.Option B remains the local surface mechanism, whereas option D provides the causal explanation.
  • F Prompt Templates: The prompt suite uses SESSION-SUMMARY during offline preprocessing, UPDATE during refinement, and an answer prompt once the agent decides to answer.These prompts construct H(D), update the signature through Mupd, and send the final context to Mgen.
  • F Prompt Templates: The retrieval-planning prompt treats the signature as a narrative compass and the rewritten query as a target for missing evidence, combining both signals during retrieval.The prompt also distinguishes session summaries, retrieved passages, and running evidence memory, and directs the agent to REFINE when key evidence is missing or confidence is low.
  • F Prompt Templates: Answer-time contexts prepend the final chunks with no memory, the signature, evidence memory, or both, followed by formats tailored to multiple-choice, open-QA, and claim verification.Multiple-choice outputs require JSON with an A/B/C/D answer, open QA requests a concise phrase without explanation, and NoCha requires TRUE/FALSE with a bounded explanation.

G Limitations

MiA-Signatures are effective for long-context narrative understanding, but the reported experiments do not establish whether the formulation transfers beyond literary and narrative domains.

  • MiA-Signatures provide an effective memory interface for long-context narrative understanding, particularly when evidence is dispersed across a large source.
  • The experiments focus on literary and narrative domains where memory naturally forms chapter- or session-level units.
  • Transfer to code repositories, scientific literature, or multimodal interaction remains untested.
Loading 2605.06416v1…