PaperlayerView exact revision on arXiv

Source-linked AI summary

Parametric Session States Recover Retrieval Intent without Replaying History

Qingsen Ma, Dianyun Wang, Jiahong Liu, Irwin King

arXiv:2610.33047v1cs.CLcs.AI

TL;DR

Underspecified follow-ups can omit the session identifiers needed to retrieve relevant evidence, while conventional RAG assumes an information-sufficient query. PSS compiles session history and retrieval supervision into a reusable state that restores search intent while keeping facts external. Across controlled sessions, model families, and retrieval backends, correct states improve retrieval, and controls identify signal from ordered history beyond supplied targets.

  • Problem

    Underspecified follow-ups omit session scope needed for retrieval, while conventional RAG assumes the query already contains sufficient information.

    Source: §1 · pp. 1–2

  • Method

    PSS compiles ordered session history and positive and negative retrieval targets into a reusable state that generates a search clue before being unloaded.

    Source: §1 · p. 2

  • Results

    Correct states substantially improve retrieval over ambiguous queries and causal controls across three model families and heterogeneous retrieval backends.

    Source: §1 · p. 3

  • Takeaways & Limitations

    PSS provides a retrieval-scoped memory layer for repeated-use assistants, compiling session intent once for reuse while leaving precise facts external.

    Source: §5 · p. 10

    PSS-320 uses synthetic project names and controlled technical facts, with no human-subject or private user data.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Long-running assistants often interpret follow-up requests whose referents, constraints, and active goals were established many turns earlier. The central challenge is therefore not only context length, but retrieval intent: a current query such as ``use the previous configuration'' may omit the project, entity, and version terms needed to locate relevant evidence. Long-context models can recover those terms by replaying the interaction, but repeatedly pay the cost of processing history. Conventional retrieval-augmented generation (RAG) is cheaper, yet assumes an information-sufficient query; conversational rewriting either rereads the history or depends on an explicitly maintained textual memory. We introduce \emph{Parametric Session State} (\method{}), a reusable session-memory layer that remembers what to search for while keeping factual knowledge external.

1 INTRODUCTION

Underspecified follow-ups omit the session identifiers needed to retrieve relevant evidence, exposing a gap between long-context replay and conventional RAG. PSS compiles session history into a reusable retrieval state, and the paper pairs this architecture with controls designed to identify session-specific retrieval effects.

1 INTRODUCTION

  • Conventional RAG can miss intended evidence when a follow-up omits the project and version identifiers that remain only in session history.A stronger encoder alone cannot recover information absent from the query.Source: §1 · p. 1
  • The systems gap is a reusable layer that restores missing query scope while leaving exact facts external; the evaluation gap is testing that ability with underspecified queries.Ordinary long-context QA often uses self-contained questions that can be answered without recovering conversational scope.Source: §1 · p. 2
  • PSS compiles ordered history and positive and negative retrieval targets into a session-specific state that generates a clue while precise facts remain in external evidence.The adapter is unloaded before the frozen base model answers from retrieved text.Source: §1 · p. 2
  • The evaluation combines public session protocols and PSS-320 with disjoint splits, state controls, null and OOD probes, and session-level inference.The design distinguishes recovered session scope from generic query improvement.Source: §1 · p. 2
  • Correct states substantially improve retrieval over ambiguous queries and causal controls across three model families and heterogeneous retrieval backends.Three-seed confirmation shows that ordered history contributes retrieval signal beyond supplied targets.Source: §1 · p. 3

2 METHOD

PSS keeps the retriever and answer model frozen while a session-specific state reconstructs a search clue from history and write-time supervision. The clue retrieves external evidence; the state is unloaded before the base model answers.

2 METHOD

  • The workflow compiles ordered history and positive/negative clue supervision offline, then uses the resulting adapter to produce a retrieval clue at inference.The answer model uses retrieved evidence after the adapter has been unloaded.Source: Figure 2 · p. 3

2.1 PROBLEM FORMULATION

  • PSS changes query interpretation while leaving the external index, retriever, and answer model unchanged; history and writer supervision remain offline-only.The writer does not receive formal answers or target-document identifiers.Source: §2.1 · p. 3

2.2 LORA AS A BOUNDED SESSION-STATE CARRIER

  • A session-specific, bounded LoRA state supplies the retrieval interface, while the base model parameters remain frozen.Each model uses its native projection layout and a rank-32, single-slot BF16 state.Source: §2.2 · p. 4

2.3 RETRIEVAL-ALIGNED SESSION COMPILATION

  • The compiler first optimizes history language modeling, then clue readout using positive and same-type archived or conflicting clues.Targets are automatically constructed from controlled session fields or history-parsed active requests; they do not use future probes, formal answers, or target-document IDs.Source: §2.3 · p. 4

2.4 INTENT READOUT, RETRIEVAL, AND UNLOADING

  • At inference, the state generates a textual retrieval clue compatible with varied retrieval backends, then unloads before the frozen base model answers from retrieved evidence.Raw-query and PSS rankings can be fused, while the guarded path routes among raw, card, and PSS-RRF retrieval using observable features.Source: §2.4 · pp. 4–5

3 EXPERIMENTS

The experiments test whether PSS restores omitted session intent while keeping the corpus, index, retriever, and reader fixed. They assess retrieval across public and controlled sessions, models, and backends, then examine output alignment and design interventions.

3.1 EVALUATION DESIGN

  • The evaluation holds corpus, index, retriever, and reader fixed while varying query source or session state across public sessionizations and a shared-corpus PSS-320 test.It uses three backbones and lexical, dense, hierarchical, graph, and query-transformation retrievers, with paired session-level inference and Holm correction.Source: §3.1 · p. 5, §3.2 · p. 5

3.3 RQ1: RETRIEVAL-INTENT RECOVERY GENERALIZES

  • On 262 sealed public sessions, Qwen PSS improves All-Recall@3 over raw, wrong-state, and shuffled-state queries by 35.88, 31.30, and 48.09 points; answer F1 rises from 1.03% to 20.00%.Figure 3 reports session-weighted target-document retrieval and isolated-reader answer F1.Source: §3.3 · p. 5, Figure 3 · p. 6
  • On UltraDomain, correct states reach 98.21–99.11% All-Recall@3 across three backbones, improving raw-query recall by 1.79–4.46 points.The transformation covers financial, legal, and mixed-domain documents.Source: §3.3 · p. 6
  • At K = 16, PSS reaches 83.75% and 95.00% All-Recall@4 on BGE-M3 and BM25, replacing record-wise fusion’s 16 evidence queries with one.PSS performs question-conditioned readout from its compiled parameter state without supplying the text bank to the readout model.Source: §3.3 · p. 7
  • Across three backbones, public-session retrieval improves by 33.97–35.88 points; on PSS-320, state clues improve BGE-M3 by 21.48–45.63 points and the graph-compatible retriever by 32.11–53.13 points.The disjoint 112-session confirmation reports gains over ambiguous raw queries across seven backends, with four satisfying the joint Holm-corrected rule against raw, matched-wrong, and shuffled states.Source: §3.3 · p. 7

3.4 RQ2: DO FUNCTIONAL OUTPUTS ALIGN WITH SESSION-SPECIFIC INTENT?

  • Session-level target alignment improves by 0.1136–0.1779 over raw, matched-wrong, and shuffled controls, with confidence intervals bounded away from zero and Holm-adjusted p-values of 0.000300.The analysis finds functional outputs are session-specific, while the eight coarse semantic categories do not explain state proximity.Source: §3.4 · p. 8, Table 3 · p. 8

3.5 RQ3: DESIGN INTERVENTIONS AND ABLATIONS

  • Registered interventions support contributions from full ordered history, session correspondence, and learned parameter structure; PSS clue retrieval gives the largest mechanism gain, while guarded use gives the best non-oracle score.The canonical shared operating point is rank 32, α=64, last-8, single-slot; backbone-specific screens show sensitivity to writer depth, rank, and slot count.Source: §3.5 · pp. 8–9, Table 4 · p. 8
  • In a three-seed confirmation, adding ordered history improves direct All-Recall@4 over target-only PSS by 5.54 points with BGE-M3 and 10.18 points with BM25.At K = 16, one compiled state accesses distinct records; PSS reaches 83.75% and 95.00% on BGE-M3 and BM25, respectively.Source: §3.5 · p. 9

4 RELATED WORK

Related work addresses long-context efficiency, compact or parametric memory, retrieval organization, query rewriting, and long-term memory evaluation. PSS is positioned as a per-session parameter state for restoring retrieval scope while leaving exact facts external.

Compact and parametric memory

  • Long-context and compression methods retain or reduce the cost of historical tokens, while compact-memory methods learn token or activation interfaces; PSS instead uses a bounded per-session LoRA state to recover retrieval scope and unloads it before answering.LoRA provides a parameter-efficient carrier [Hu et al., 2021], while cited document or corpus modules internalize content [Charakorn et al., 2026].Source: §4 · p. 9

RAG organization and query recovery

  • Hierarchical and graph retrieval organize evidence, while query rewriters bridge queries and evidence; PSS stores a per-session parameter delta and tests ambiguous follow-ups with wrong-session interventions.MemoRAG is noted as closest in functional decomposition because memory generates clues before an expressive reader.Source: §4 · p. 9

Long-term memory evaluation

  • Prior long-term-memory studies examine contextual query interpretation, indirectly expressed information, and cross-session reasoning; these protocols provide explicit session associations and information-matched text controls to study compilation into parameters and retrieval use.The public extension varies whether the query reveals its active request, while PSS-320 supports causal state interventions.Source: §4 · p. 9

5 CONCLUSION

The conclusion presents PSS as a reusable retrieval-intent layer that compiles ordered history and retrieval supervision into a session-specific state while keeping exact knowledge external. The accompanying task contract, evaluation rationale, and benchmarks define how this capability is isolated and tested.

5 CONCLUSION

  • PSS addresses underspecified follow-ups by compiling retrieval supervision and ordered history into a reusable state that reconstructs search intent without replaying historical tokens.The conclusion reports controlled retrieval gains over a compact card and three-seed evidence that ordered history contributes beyond supplied targets.Source: §5 · p. 10

A FORMAL TASK CONTRACT AND ALGORITHMS

  • The stage contract changes only a session-local adapter during writing and clue generation; retrieval uses an unchanged external index, and answering uses the unloaded frozen base model.The paired comparison holds the corpus, retriever, cutoff, and reader fixed while replacing the query source.Source: A FORMAL TASK CONTRACT AND ALGORITHMS · p. 14, Table 5 · p. 14

B.1 EVALUATION RATIONALE AND METHOD VOCABULARY

  • The evaluation defines PSS clue retrieval as using the correct session state to emit a clue for retrieval alone; Raw + PSS RRF fuses rankings, while Guarded PSS routes among raw, card, or fused rankings.The evaluation rationale also uses public sessionizations, PSS-320 state controls, frozen retrievers, cross-backbone tests, and same-evidence comparisons.Source: §B.1 · p. 14

B.2 WHY SESSION-INTENT MEMORY BENEFITS FROM A DIFFERENT EVALUATION

  • The causal test asks whether a state from session s restores that same session’s omitted retrieval intent, using underspecified queries, shared candidate corpora, and matched-wrong, shuffled, and zero-state controls.Null, out-of-distribution, revision, and conflict probes test false activation and stale-scope failures.Source: §B.2 · pp. 15–16, Table 6 · p. 16
  • Public sessionizations test transfer to released data, while PSS-320 supplies a shared cross-session retrieval pool and controlled state interventions to identify session-intent effects.The two tracks answer different questions: transfer to released data and whether gains arise from recovering the corresponding session’s omitted intent.Source: §B.2 · p. 16

B.3 BENCHMARKS AND SESSIONIZATION

  • The public benchmark derives 326 sessions from seven QA tasks and evaluates 262 sealed test sessions; its parsed active-request reference uses explicit activity markers and therefore tests retrieval access under shared request supervision.The separate 112-session UltraDomain transformation tests finance, legal, and miscellaneous documents; the original task has no session history, so PSS is inapplicable there.Source: §B.3 · p. 16
  • PSS-320 contains 320 AI-controlled synthetic technical-project sessions and 1,920 probes, with a source-group-disjoint test of 256 sessions and audits of evidence alignment and scope.Because the sessions are synthetic, PSS-320 is a controlled mechanism benchmark.Source: ETHICS STATEMENT · p. 10, §B.3 · pp. 16–17

B.4 MODELS, CONTROLS, AND EVALUATION

The evaluation freezes model-specific states, baselines, controls, and inference procedures before testing portability. State fingerprints show no category clustering, while generated clues align with session-specific retrieval targets in the original embedding space.

B.4 MODELS, CONTROLS, AND EVALUATION

  • The frozen evaluation uses Qwen2.5-7B-Instruct with BF16, greedy decoding, BM25, and registered correct, wrong, shuffled, and zero-state controls.The same-query, same-evidence answer ablation varies only decoding state; inference treats sessions as independent units and applies paired intervals and Holm correction.Source: §B.4 · p. 17

C STATE FINGERPRINTS AND FUNCTIONAL-OUTPUT ALIGNMENT

  • The category cosine silhouette is -0.0090, and same-category nearest neighbors occur for 1.17% of sessions versus a 12.16% chance rate.These statistics precede UMAP projection; Figure 6 visualizes factorization-invariant fingerprints and median effective-update energy.Source: §B.4 · p. 17, Figure 6 · p. 17
  • Fingerprints use effective updates and deterministic projections because raw LoRA factors are rescaling-nonidentifiable; functional tests compare state-generated clues with raw, control, and gold rewrites.UMAP is descriptive; functional effects are calculated in the original embedding space, and the gold rewrite is only an evaluation target.Source: C STATE FINGERPRINTS AND… · p. 18, Figure 7 · p. 18
  • PSS-clue progress over raw was +0.1136 (95% CI [+0.1008, +0.1266]; Holm p = 0.00029997), and exceeded matched-wrong and shuffled controls.The follow-up covers 256 sessions and uses original 1,024-dimensional BGE-M3 embeddings; it validates session-specific output alignment, not parameter geometry.Source: C STATE FINGERPRINTS AND… · p. 18, Figure 7 · p. 18

D.1 CROSS-BACKBONE HYPERPARAMETER SENSITIVITY

Hyperparameter effects differ by backbone, while the locked single-slot configuration remains a shared robust choice. A disjoint carrier confirmation finds retrieval signal from ordered history beyond supplied targets, but the frozen guard misses its transfer gate.

D.1 CROSS-BACKBONE HYPERPARAMETER SENSITIVITY

  • Qwen’s last-16 screen winner gains 17.50 points but confirms at only +7.50; Phi’s last-16 winner gains and confirms at +7.50 points.Qwen confirmation: 95% CI [−6.25, 21.25], Holm p=0.410; Phi confirmation: 95% CI [0.00, 17.50], Holm p=0.258. The screen and disjoint confirmation are shown in Figure 8.Source: §D.1 · pp. 18–19, Figure 8 · p. 19
  • The canonical single-slot topology remains shared, but parameter effects are non-monotone: doubling rank helps Qwen, slightly helps Phi, and harms Mistral.A second slot loses 20.00, 2.50, and 8.75 points for Qwen, Mistral, and Phi, respectively; canonical states occupy 44, 40, and 24 MiB.Source: §D.1 · p. 19, Table 7 · p. 19

E SUPERVISION AND CARRIER CONFIRMATION

  • History increment is positive after correction for both retrievers and every seed, beyond the target-only state.The carrier effect depends on representation: PSS beats matched text with BGE-M3, whereas BM25 favors lexically richer text; Table 8 reports session-weighted retrieval and guard results.Source: E SUPERVISION AND CARRIER CONFIRMATION · p. 20, Table 8 · p. 20
  • The frozen guard exceeds its 5% null/OOD threshold, with 17.86% activation for correct PSS, so its transfer gate fails.Direct retrieval remains the registered primary analysis; the confirmation is not tuned on its test set. Table 9 reports preregistered effects and seed-direction criteria.Source: E SUPERVISION AND CARRIER CONFIRMATION · p. 20, Table 9 · p. 20

F LIMITATIONS AND THREATS TO VALIDITY

  • The reported retrieval improvements are scoped by qualifications on router transfer, answer decoding, costs, backend status, benchmark construction, privacy, and continual learning.These limits define current generalization settings rather than changing the registered retrieval improvements.Source: F LIMITATIONS AND THREATS TO VALIDITY · p. 20

F.1 EXPERIMENTAL STATUS

  • The carrier confirmation uses a newly generated, disjoint 320-session cohort and holds the writer design and seeds fixed across reported runs.The confirmation tests history increment and a matched text baseline under a separate preregistration.Source: §F.1 · pp. 20–21

F.2 SYNTHETIC BENCHMARK AND SAMPLED HUMAN AUDIT

The discussion bounds the retrieval findings by benchmark construction, routing, state capacity, costs, and evaluation scope. Answer-generation results are auxiliary, and portability across tested settings does not establish broader guarantees.

F.2 SYNTHETIC BENCHMARK AND SAMPLED HUMAN AUDIT

  • PSS-320 uses controlled templates and programmatic evidence alignment, creating repeated patterns, artificial identifiers, and a narrow technical-project domain.The public extension reduces generator dependence, but En.QA and NarrativeQA sessions index one shared document and nearly saturate target-document recall.Source: §F.2 · p. 21

F.3 ROUTER TRANSFER ACROSS DOMAINS

  • The public router improves over raw retrieval but trails PSS clues, activates on null and OOD probes, and fails the frozen guard’s 5% transfer gate.Correct-PSS null/OOD activation is 17.86%; routing reduces BGE-M3 and BM25 retrieval relative to direct retrieval.Source: §F.3 · p. 21

F.5 STATE CAPACITY AND SCALING

  • Fixed state shape does not establish session-wide information retention, and capacity is finite while write compute and total storage grow with history and session count.Current queries, retrieved text, and generated-token KV also consume online memory.Source: §F.5 · p. 21

F.6 COST RELATIVE TO TEXT STATE

  • Text cards optimize storage, editability, and write latency; with BGE-M3 PSS gains 6.25 points, while BM25 gains 11.07 points from lexically explicit text.The primary Qwen rank-32 state is 44.04 MiB and takes 29.03 seconds to write; the parsed reference averages 65.65 raw UTF-8 bytes.Source: §F.6 · pp. 21–22

F.7 RETRIEVER AND DOMAIN INTERACTION

  • BM25 is unusually strong on PSS-320’s explicit project and version identifiers, so these results do not establish a general ranking of retriever types.Natural paraphrases, multilingual sessions, and semantically equivalent descriptions could change the results; graph-compatible performance also depends on pipeline configuration and indexing cost.Source: §F.7 · p. 22

F.8 CAPPED CONTEXT AND EFFICIENCY SCOPE; F.9 QUALITATIVE AND CAPACITY-ANALYSIS SCOPE

  • The capped full-context arm is not a true full-history oracle, efficiency measurements are narrow, and only 42 of 240 semantic-peak combinations yield identifiable breakpoints.The breakpoints do not identify a capacity law or globally optimal rank; measured efficiency lacks run-to-run variation, energy, validated FLOPs, and length-specific curves.Source: §F.8 · p. 22, §F.9 · p. 22

F.10 PRIVACY, DELETION, AND CONTINUAL LEARNING; F.11 THREATS BY VALIDITY CLASS; G.1 PER-TASK RETRIEVAL AND ANSWERING

  • Adapter controls do not test membership inference, extraction, theft, unlearning, or deletion, and the registry experiment does not demonstrate continual learning or rollback.Validity is also bounded by one language, three related model families, and three writer seeds; the public answer table reports F1 rather than Exact Match.Source: §F.10 · p. 22, §F.11 · p. 23, §G.1 · p. 23

F.4 SEPARATION OF RETRIEVAL AND ANSWER GENERATION; G.1 PER-TASK RETRIEVAL AND ANSWERING

  • The primary endpoint is evidence retrieval, not answer correctness; the reported public PSS-clue F1 is 20.00%, versus 1.03% for ambiguous queries.PSS-clue F1 is close to the parsed reference’s 21.78%; answer results remain auxiliary to the retrieval intervention.Source: §F.4 · p. 21, §G.1 · p. 23

G.2 CONVERSATIONAL-REWRITE RETRIEVAL-TO-QA EXTENSION

On the primary public sessionization, PSS improves retrieval and controlled answering over raw queries for all three readers; UltraDomain shows significant answer gains but does not pass the joint retrieval-to-QA criterion. The extension also reports a separate 262-session public audit and 112-session UltraDomain replication.

G.2 CONVERSATIONAL-REWRITE RETRIEVAL-TO-QA EXTENSION

  • The extension uses one generated rewrite per method across readers, hides history and unloads PSS, and evaluates retrieval and controlled answering with a shared top-3 evidence interface.The public track reuses a locked 262-session split; sessionized UltraDomain provides a separate 112-session replication, with answers, documents, and target IDs withheld from rewriters.Source: §G.2 · pp. 23–24
  • 35.50–35.88 points in All-Recall@3 and 4.06–10.60 points in controlled answer F1: PSS passes the joint criterion against raw queries for all three public-track readers.Both effects have pHolm = 3.0 × 10−4; PSS also beats ConvSearch-R1 in retrieval by 14.89–15.27 points for every backbone, but its F1 advantage is significant only for Qwen2.5-7B.Source: §G.2 · p. 24, Table 13 · p. 25
  • 10.63–18.69 points in controlled answer F1: PSS improves UltraDomain F1 for all readers, but its 1.79-point mean retrieval gain is nonsignificant and the joint criterion is not met.Positive means below the adjusted threshold remain descriptive, not significant retrieval-to-QA support.Source: §G.2 · pp. 24–25, Table 15 · p. 26

G.3 PAIRED EFFECTS, CONTROLS, AND AUDIT

  • No dagger is assigned: the public router diagnostic fails its all-baselines rule because the parsed-reference comparison is negative, and no cross-track combined p-value is constructed.The four PSS-320 state-versus-raw/card comparisons use 256 paired sessions and have pHolm = 3.99996 × 10−5, so both state rows in Table 2 receive daggers.Source: §G.3 · p. 25, §G.2 · p. 25
  • The final aggregate contains 4,192 rows for 262 expected sessions; among answerable sessions the router selects parsed reference 158 times, raw 104 times, and PSS-RRF zero times.These counts explain why the guard improves abstention while sacrificing much of the correct-state retrieval gain.Source: §G.3 · p. 26

H CROSS-MODEL AND ULTRADOMAIN DETAILS

Correct session states produce substantial gains over wrong and shuffled states in the public evaluation, while UltraDomain’s high raw and parsed-reference ceilings constrain the preregistered decision. The section also describes synthetic session construction, state writing, causal controls, and guarded routing.

H CROSS-MODEL AND ULTRADOMAIN DETAILS

  • 31.30–33.59 points are the public correct–wrong gaps, and 47.71–48.09 points are the correct–shuffled gaps across Qwen, Mistral, and Phi-3.UltraDomain’s direction is consistent, but high raw and parsed-reference ceilings prevent the complete preregistered decision from passing.Source: H CROSS-MODEL AND ULTRADOMAIN DETAILS · p. 26, Table 16 · p. 27
  • Phi-3’s 24 MiB state obtains the strongest PSS-320 scores in both added retrievers, but this does not establish a causal architecture advantage or cross-model state interchangeability.Backbone and state properties change together; the experiment supports method-level portability, not a globally optimal rank.Source: H CROSS-MODEL AND ULTRADOMAIN DETAILS · p. 26, §J.6 · p. 29

I.1 GENERATION CONTRACT

  • PSS-320 builds synthetic technical-project sessions with active and archived configurations, source documents, ordered histories, and six probes; calibration and test source groups are disjoint.Null/conflict probes have empty evidence groups and gold rewrites of UNSUPPORTED; automated checks report zero issues, with sampled human validation of construction criteria.Source: §I.1 · p. 27, §I.2 · p. 27, §I.3 · p. 27

J.1 BACKBONE AND ADAPTER PLACEMENT

  • The selected Qwen2.5-7B writer uses rank-32 LoRA in the final eight layers, first training on ordered-history language modeling and then on clue readout with retrieval-aligned targets.The retrieval-aligned target includes project, current configuration/component, checkpoint, and route terms; training excludes evaluation-probe answers.Source: §J.1 · p. 28, §J.2 · p. 28, §J.3 · pp. 28–29

J.7 CAUSAL STATES

  • Correct, matched-wrong, shuffled, and zero-reset states use the same backbone, prompts, retriever, and scoring code to test whether effects depend on the session-specific state.For the guarded arm, correct, matched-wrong, and shuffled states obtain 0.5063, 0.3313, and 0.3852 All-Recall@4, respectively; the reconstructed zero-reset value is provisional.Source: §J.7 · p. 29, §M.1 · p. 32

K.2 MODEL FITTING

  • Guarded PSS routes among raw, card, and PSS-RRF using observable retrieval features, with calibration completed before held-out evaluation; action-conditioned accuracy is not a causal estimate.The router uses cross-fitting and fixed thresholds; claim scope remains limited by synthetic construction, mechanical public sessionization, and reader and router limitations.Source: §K.2 · p. 30, §K.3 · p. 31, §K.4 · p. 31

N SEMANTIC CATEGORY ANALYSIS

PSS clues are strongest relative to the card for categories requiring recovery of specific project, configuration, version, or relation terms, while the card performs better for global-theme retrieval. Router action counts describe allocation, not action-specific causal quality.

N SEMANTIC CATEGORY ANALYSIS

  • PSS clue retrieval’s strongest card-relative gains occur in coreference, exact fact, multi-hop, and entity binding; the global-theme card is better by five points.These categories respectively involve recovering project, component, configuration, version, or relation-chain terms before retrieval.Source: N SEMANTIC CATEGORY ANALYSIS · p. 32
  • The null/OOD row is a retrieval diagnostic, not a safety result; safety is measured by the router’s conditional false-activation rate.Null queries have no gold evidence groups, and an unguarded rewrite can still retrieve lexically related chunks.Source: N SEMANTIC CATEGORY ANALYSIS · p. 32, Table 27 · p. 33
  • PSS-RRF is selected most often for temporal revision and frequently for coreference and entity binding, while the card has the highest count for global theme.Because the router allocates actions using observable confidence, outcome means across selected action groups are confounded by sample difficulty.Source: N SEMANTIC CATEGORY ANALYSIS · p. 33, Table 28 · p. 33

O RETRIEVER INTERACTION

Correct PSS clues improve retrieval across completed backends, with several passing joint corrected comparisons against raw, wrong, and shuffled states. Retrieval behavior varies by backend, and answer-side interpretation is constrained by the reader’s zero Exact Match results.

O RETRIEVER INTERACTION

  • 35.71, 30.36, 28.57, 35.71, and 35.71 points are the gains over raw for BGE-M3, Contriever, Qwen3-Embedding-4B, and compatible HyDE/RQ-RAG pipelines, respectively.These five backends pass the joint Holm-corrected rule against raw, wrong, and shuffled states; BM25 and RAPTOR gains do not pass the joint control rule. Across all seven rows, PSS is not significantly different from the parsed reference.Source: Table 29 · p. 33, O RETRIEVER INTERACTION · p. 34
  • The entity-relation graph has the strongest gold-query score, while BM25 has the strongest PSS-clue score; the hierarchical retriever nearly matches BM25 with clues.BGE-small underperforms on every path, but BGE-M3 reaches 0.3391 versus 0.1242 for raw retrieval, an encoder-specific result rather than a general dense-retrieval ranking.Source: O RETRIEVER INTERACTION · p. 33
  • Raw-plus-PSS reciprocal-rank fusion performs below PSS-clue retrieval for all four retrievers, because weak raw-query hits can displace required evidence groups.This motivates a calibrated action selector rather than unconditional hybrid retrieval.Source: O RETRIEVER INTERACTION · p. 33

P.1 READER CEILINGS

  • All listed arms have zero Exact Match, so retrieval-to-answer transition metrics based on Exact Match have no positive baseline cases.Gold evidence also underperforms gold-query RAG; possible explanations include answer/citation formatting, evidence rendering or amount, and reader-prompt mismatch.Source: Table 30 · p. 34, §P.1 · p. 34

P.2 SAME-QUERY, SAME-EVIDENCE STATE ABLATION

The same-query, same-evidence analysis finds no reliable decoder-side memory effect, while the capacity diagnostic cannot support a stable capacity law. Across the merged sections, results characterize retrieval, efficiency, statistical evidence, and important evaluation boundaries.

P.2 SAME-QUERY, SAME-EVIDENCE STATE ABLATION

  • No category exceeds unloaded, wrong, and shuffled controls simultaneously by 0.05, so the state is not a reliable decoder-side memory.The same-query, same-evidence ablation isolates answer-stage effects from retrieval effects.Source: §P.2 · p. 34, Table 31 · p. 34

Q SEMANTIC-PEAK CAPACITY DIAGNOSTIC

  • Only 42 of 240 semantic-peak combinations (17.5%) have an identifiable breakpoint, insufficient to fit a stable capacity law.The remaining combinations are flat, noisy, or do not cross the registered 90% threshold within observed levels.Source: Q SEMANTIC-PEAK CAPACITY DIAGNOSTIC · p. 35

R RQ4: INFERENCE, STORAGE, AND BREAK-EVEN ACCOUNTING

  • 67.64% lower total latency and 98.97% fewer online prompt tokens characterize Qwen PSS versus the available mixed-length full-history profile.The 29.03-second offline write amortizes after an estimated 43 repeated reads against average full-history prefill, but never against the 32-token card; these measurements are workload-specific.Source: R RQ4: INFERENCE, STORAGE, AND… · p. 36, §R.3 · p. 36

S.2 BINARY PAIRED TESTS

  • For H2, the paired difference interval [0.0898, 0.1344] exceeds the −0.02 non-inferiority margin; the reported exact and Holm-adjusted p-value is 1.36 × 10−10.For H1, 134 of 256 sessions were correct only under guarded state and zero only under raw retrieval; the Holm-adjusted p-value was 1.84 × 10−40.Source: §S.2 · p. 37

S.4 MULTIPLICITY AND EXPLORATORY ANALYSES

  • The corrected primary family comprises H1 and H2; category, retriever, efficiency, and other listed analyses are exploratory, and H3 lacks bootstrap intervals for each retrieval-control gap.Aggregate stage success does not establish identical denominators for every auxiliary metric.Source: §S.4 · p. 37, §S.5 · pp. 37–38

U QUALITATIVE AND ERROR-ANALYSIS PROTOCOL

  • No completed manual error labels are reported, so the proposed explanations for card, LoRA, and reader failures remain unvalidated hypotheses.The planned taxonomy covers scope or entity binding, revision conflicts, evidence misses, router errors, and unsupported generation.Source: U QUALITATIVE AND ERROR-ANALYSIS… · p. 39

V.1 INFORMATION-MATCHED TEXT CONTROLS

  • On PSS-320, PSS clue retrieval reaches 50.08% All-Recall@4 versus 39.45% for the compact card, a 10.63-point gain.Guarded PSS reaches 50.63%, an 11.18-point gain over the card.Source: §V.1 · p. 40

V.2 ONLINE PARAMETER READOUT AND PSS-DERIVED TEXT REUSE

  • Online PSS rewriting improves BGE-M3 and BM25 All-Recall@4 over base-only rewriting by 21.04 and 28.96 points, respectively.On BGE-M3, it also exceeds fixed PSS text-readout rewriting by 15.62 points (pH = 0.00035); the BM25 effect is 12.71 points (pH = 0.0826).Source: §V.2 · p. 41

V.3 QUESTION-CONDITIONED READOUT FROM DISTINCT HISTORICAL RECORDS

Question-conditioned PSS reads selected historical records from a compiled state without supplying the persistent text bank to the readout model. In the tested settings, it improves evidence retrieval over several alternatives, while controlled answering and serving-cost analyses distinguish evidence effects from interface and workload costs.

V.3 QUESTION-CONDITIONED READOUT FROM DISTINCT HISTORICAL RECORDS

  • At K = 16, PSS reaches 83.75% All-Recall@4 with BGE-M3 and 95.00% with BM25, exceeding all-record concatenation by 65.00 and 60.00 points.Compared with record-wise RRF, PSS gains 67.50 and 55.00 points while reducing evidence queries from sixteen to one; wrong-group PSS reaches 1.25% and 0.00%.Source: §V.3 · p. 43, Table 40 · p. 43
  • At K = 16, PSS has the highest BGE-M3 mean among evaluated non-reference methods, 6.25 points above the strongest reported text score; Grok selection reaches 97.50% with BM25.PSS reaches 95.00% with BM25, exceeding all four retrieval-based selector configurations and bank rewriting.Source: §V.3 · p. 44, Table 41 · p. 44
  • At K = 16, mean bank-rewriting memory inputs are 1367.69 tokens, while mean generated lengths are 63.60 tokens for PSS and 16.89 for bank rewriting.The cost profile is workload- and hardware-specific; its boundary excludes common final answering, offline model/index construction, and disk state-cache misses.Source: §V.3 · p. 44, §V.5 · p. 46

V.4 EVIDENCE-CONTROLLED ANSWERING

  • With a fixed explicit gold rewrite and the same frozen, unloaded reader, PSS evidence improves F1 over raw-query evidence by 6.03 points (95% CI [2.64, 9.23], pH = 0.0214).This controlled 16-session result isolates evidence access; when each method’s retrieval string is instead used as the reader’s question, PSS-minus-raw is −6.89 F1 points.Source: §V.4 · pp. 44–45, Table 43 · p. 45

V.5 COMPLETE ONLINE ACCOUNTING

  • Pure PSS takes 1.341 seconds on average versus 2.619 seconds for full-history rewriting, a 48.8% reduction under the measured workload.Complete Guarded PSS takes 2.630 seconds, using two generations and three retrievals; its total prompt count is 138.125 tokens.Source: Table 44 · p. 45, §V.5 · p. 46
Loading 2610.33047v1…