The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search

Peiyang Liu; Xi Wang; Di Liang; Wei Ye

Source-linked AI summary

The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search

Peiyang Liu, Xi Wang, Di Liang, Wei Ye

arXiv:2608.23252v1Aug 24, 2026cs.LGcs.CLcs.IR

TL;DR

RAG lacks rigorous measurement of evidence utilization and principled allocation of fixed context budgets for diverse portfolios. The paper introduces causal leave-one-out attribution and evaluates sequential context allocation with a deconfounded factorial design. It reports that hard-negative evaluation exposes proxy failures, context width has elasticity −0.68, and sequential generations yield 16.8 to 20.5 percentage points of portfolio-recall gains.

Problem
RAG needs rigorous measurement of what evidence language models utilize and how fixed context budgets should be allocated for portfolio generation.
Method
The paper combines causal leave-one-out probing, a deconfounded context-allocation experiment, and an attribution-steered closed-loop submodular scheduler.
Results
16.8 to 20.5 percentage points of portfolio-recall gains are reported for sequential allocation, while hard-negative tests expose traditional proxy failure and measure width elasticity at −0.68.
Takeaways & Limitations
The findings support sequential, feedback-driven orchestration rather than static context maximization for generative search.
Takeaways & Limitations
Sequential generation incurs higher autoregressive latency and probe overhead, and open-loop context rotation can waste budget because relevance density decays down ranked lists.

Abstract

from arXiv · show

As Retrieval-Augmented Generation (RAG) shifts toward diverse portfolio generation, it is stymied by two critical bottlenecks: flawed measurement of evidence utilization, and suboptimal context budget allocation. We resolve both sequentially. To resolve measurement, we expose a pervasive ``diagnostic illusion'': standard relevance proxies fail catastrophically on hard negatives. We replace them with an efficient causal leave-one-out probe that accurately isolates generative reliance and formally calibrates the structural dilution of LLM attention. To resolve allocation, we deploy this causal probe in a deconfounded factorial grid. We prove that the prevailing strategy of monolithic context widening is an architectural trap penalized by relevance decay. Instead, allocating compute iteratively across multiple sequential generations drives transformative portfolio recall gains of 16.7--20.5 absolute percentage points, scaling robustly up to 32B models. Finally, we unify these solutions into a deployable closed-loop submodular scheduler. Augmented by an attribution-steered contrastive decoder to override LLM attention inertia, our architecture systematically forces fresh evidence integration. By dominating classical open-loop baselines, we establish sequential, feedback-driven orchestration as the definitive paradigm for generative search. Our code, data, and causal measurement instruments are available at https://github.com/PeiYangLiu/ascp.

1 Introduction

The paper addresses how to measure evidence utilization and allocate fixed inference budgets for generative portfolios. It proposes causal attribution, sequential narrow contexts, and closed-loop orchestration as solutions.

Context-allocation problem: Generative portfolio construction requires deciding whether a fixed budget should fund one wide context or multiple sequential narrow contexts.The two configurations differ in how they expose retrieved evidence across generations.
Causal measurement: A leave-one-out counterfactual probe measures generative reliance efficiently by holding the generated response fixed during evaluation.This enables teacher-forced audits without the bottleneck of autoregressive decoding.
Diagnostic illusion: Standard relevance proxies can appear near-perfect on off-query distractors but collapse to chance on same-query hard negatives, whereas the causal probe remains robust.Hard negatives are topically dense documents that contain no actual answers.
Context-width dilution: −0.68 (0.02) is the calibrated width elasticity showing that evidence attribution dilutes as context width expands.The paper presents this dilution as an inherent generative behavior under controlled diagnostic isolation.
Sequential allocation: 16.8 to 20.5 percentage points of absolute portfolio coverage gains result when compute is allocated across multiple narrower sequential generations.The sequential approach is reported to breach the extraction ceiling of monolithic single-pass models up to the 32B scale, despite higher latency and probe overhead.
Closed-loop orchestration: An attribution-steered submodular scheduler uses causal feedback to drive closed-loop context orchestration and override attention inertia.The architecture is evaluated against selection-style baselines and is presented as an inference-time scaling framework for generative search.

2 Related Work

Prior work diversified ranked results and developed single-response RAG systems, but generative portfolio coverage and evidence utilization remain underexamined. This paper connects diversification, causal attribution, and inference-time scaling through sequential evidence exposure.

Diversified retrieval: Classical information retrieval diversifies ranked lists to address underspecified intent, rewarding aspect coverage and penalizing redundancy.This tradition assumes a human consumes the ranked results.
Diversity-aware RAG: RAG shifts the consumer from a human to a generative model, motivating systems that pack distinct information into a limited prompt for one comprehensive answer.Related pipelines include MMR-like penalties and sliding windows.
Sequential evidence exposure: The paper formalizes repeated generator queries so distinct document exposure drives portfolio coverage across sequential readings.The analysis suggests sliding-window gains primarily arise from increasing the distinct documents reaching the generator.
Evidence utilization: Most existing RAG systems optimize a single response rather than measuring how much retrieved evidence is consumed across successive generations.The paper contrasts this with an evidence-accumulation regime.
Causal measurement: Standard proxies and judges fail catastrophically on hard same-query negatives, making causal attribution a prerequisite for portfolio-coverage claims.The paper uses sentence embeddings as a similarity baseline and argues that causal probes are needed for valid attribution.
Inference-time scaling: Token-level diversity controls manipulate textual variety while leaving evidence utilization largely static, whereas this work redistributes grounding across multiple rounds.The scheduling and decoding interventions target over-used and under-used evidence.

3 Problem Formulation and Evaluation Metrics

This section formalizes portfolio recall and evidence coverage as complementary objectives for sequential RAG, while identifying textual diversity as a misleading proxy for evidence diversity.

3.1 The End-to-End Goal: From Single Answers to Generative Portfolios: Portfolio Recall (PR@T) measures the fraction of gold-standard answer units recovered by any response in a T-response portfolio.The portfolio is optimal when the union of sequential responses covers as much underlying truth as possible.
3.1 The End-to-End Goal: From Single Answers to Generative Portfolios: Sequential diversification is valuable because the gap between portfolio recall and the best single-response recall quantifies additional recovered truth.
3.2 The Diagnostic Signal: Quantifying Genuine Evidence Coverage: Evidence Coverage Rate (ECR) measures the percentage of unique exposed documents that the model actually utilizes across rounds.ECR uses a causal attribution matrix and a utilization threshold to quantify extraction efficiency.
3.2 The Diagnostic Signal: Quantifying Genuine Evidence Coverage: ECR penalizes scheduling policies that inject ignored documents, thereby isolating scheduling precision rather than exposure volume.
3.3 The Evaluation Trap: Textual Diversity vs. Evidence Diversity: The textual diversity confound conflates different wording with different facts, allowing repeated reliance on one document to appear as knowledge diversity.
3.3 The Evaluation Trap: Textual Diversity vs. Evidence Diversity: Textual novelty can harm reliability: aggressive novelty systematically loses source grounding and correlates negatively with lexical groundedness (r=−0.167).
3.3 The Evaluation Trap: Textual Diversity vs. Evidence Diversity: ECR tracked an identity-blinded judge’s assessment of informational coverage across 858 document-level judgments with ρ=0.654.

4 The Causal Attribution Probe and Its Validation

The paper replaces observational relevance proxies with an intervention-based leave-one-out probe that measures generative reliance efficiently and exposes failures under hard negatives.

An intervention-based causal probe is required because embedding similarity and lexical overlap are confounded by query relevance.
The probe asks how much less likely an already-generated response becomes when a document is ablated from its context.
Per-token log-likelihood drops quantify counterfactual sensitivity and structural reliance on individual documents.
The fixed-response probe uses parallelizable teacher-forced forward passes, bypassing autoregressive decoding overhead.
Off-query padding reverses the apparent ranking: cosine and BM25 approach AUCs of 1.000 at k=24, while deletion LOO reaches 0.829.The near-perfect relevance-proxy performance is attributed to construction leakage.
Deletion LOO covers 0.792 of the answer set, a +0.144 absolute gain over random ranking with p<0.001.The gain remains globally positive across context widths and identifies compact sufficient sets under redundancy.
Protocol isolation yields a calibrated width elasticity of −0.68 (0.02), showing that utilization dilution is structural rather than measurement error.Positional displacement was ruled out because deletion and length-matched replacement produced similar elasticities, −0.104 versus −0.112.

5 Theoretical Bounds and Empirical Laws of Evidence Consumption

The paper contrasts an idealized independent-consumption baseline with empirical evidence showing attention inertia, relevance-sensitive scheduling, and dilution under wider contexts.

The toy model assumes each document is independently utilized with fixed probability q across kT available document slots.
Under that baseline, maximum coverage occurs only when no document is repeated across rounds, suggesting blind fresh-document rotation.
Empirical utilization is clustered: attention inertia biases models to ignore new conflicting information after fixation on an earlier concept, with p=6×10^-31.
Naive deep rotation can feed low-quality documents and degrade task performance instead of improving coverage.
Deep rotation achieves ECR=0.375, whereas the causal-feedback submodular scheduler reaches ECR=0.626 by penalizing redundancy and forcing fresh evidence.
The measured slope ranges from −0.43 to −0.67 across answer-attestation redundancy levels, while steep decay persists in every low-redundancy stratum.
Context widening exhibits sub-linear utilization decay, with a calibrated canonical elasticity near −0.68 rather than a flat attention capacity.The exact coefficient varies with task complexity and pool redundancy.
The resulting architecture allocates budgets across narrower sequential windows, using causal feedback to maximize extraction efficiency.

6 System Architecture: The Closed-Loop Orchestration Framework

The framework replaces open-loop context rotation with a closed-loop scheduler that uses causal utilization feedback to select fresh evidence, while contrastive decoding counters attention inertia.

Architecture: Ascp dynamically orchestrates which evidence enters the context and how forcefully the decoder extracts it.The scheduler acts as the system’s brain, while the decoder addresses generative extraction.
Baseline: Rotate exposes disjoint rank windows across rounds but cannot tell whether the LLM used documents or penalize redundant information.Its blind descent through the ranking wastes budget on lower-ranked noisy documents as relevance density decays.
Scheduler: Ascp clusters documents into knowledge facets and greedily maximizes a monotone submodular objective that discounts already-utilized facets.The scheduler uses leave-one-out causal feedback after each round to favor unexplored evidence and impose diminishing returns.
Scheduler: Replacing causal feedback with embedding similarity collapses the scheduling gains, showing that causal sensitivity is central to Ascp’s operation.The ablation reports that similarity-based feedback produces no meaningful utility beyond the naive open-loop assumption.
Decoder: The attribution-steered decoder subtracts probability mass associated with over-used documents and shifts it toward fresh evidence under an adaptive-plausibility constraint.This intervention targets LLM attention inertia while bounding the shift to reduce hallucination risk.
Deployment: The architecture incurs higher autoregressive latency and probe overhead, although teacher-forced probe passes can be parallelized.The probe requires k+1 forward passes per generation, executed on already-generated text.

7 Experimental Methodology and Setup

The evaluation freezes retrieval and evidence pools, then compares allocation and orchestration strategies across matched tasks, models, settings, and paired inference runs.

Evaluation control: The zero-leakage framework freezes the retriever and retrieved document pools so performance differences reflect scheduling, prompting, and decoding interventions.All systems receive the same raw evidence.
Benchmarks: The benchmarks cover ambiguous QA, broad answer-set generation, explanatory QA, and a non-English recipe-adaptation stress test.The tasks vary in answer dispersion and relevance density, including QAMPARI’s 21 annotated entities per query and the recipe task’s slope of −0.01.
Retrieval and metrics: ASQA, QAMPARI, ELI5, and HotpotQA use preserved top-N=30 retrieval pools, with HotpotQA following the same matched-evidence pipeline.ELI5 claims are evaluated using NLI entailment rather than string containment.
Models: Experiments use open-weight 7B–8B generators, with Qwen2.5 additionally scaled to 14B and 32B models.Semantic facets are induced with English and multilingual MPNet encoders.
Statistical protocol: Primary evaluation uses fresh 100-query-per-task data across four tasks, three generators, and two seeds with paired inference and crossed-bootstrap significance testing.Benjamini-Hochberg FDR controls multiplicity within designated testing families.

8 Experimental Results

The factorial experiment crosses context width and generation count to isolate whether inference budget is better spent widening one prompt or generating sequentially.

Factorial design: The study crosses k∈{2, 5, 12, 24} with T∈{1, 5, 12} to deconfound context width from sequential generation count.The grid is evaluated symmetrically over four tasks and two generators with fixed-context controls.
Factorial design: The deconfounded width–count analysis compares rotating through disjoint evidence windows with repeatedly sampling the same top-k context.Its paired portfolio-recall contrasts isolate the gain from fresh evidence.

transformative gains, whereas expanding context width represents an architectural trap fundamentally

Across tasks and model scales, sequential narrow-window generation consistently outperforms monolithic context widening because widening pushes the model toward lower-yield evidence. Causal-feedback scheduling further improves recall and token efficiency by selecting fresh evidence.

Width versus count: Sequential generation delivers robust portfolio-recall gains across widths, while width expansion’s marginal return collapses as rounds increase.Iterative gains remain positive across task-generator combinations, whereas widening from 2 to 24 documents falls from +0.050 at T=1 to +0.024 at T=12.
Width versus count: +0.144 portfolio recall is obtained against the single-pass (24, 1) baseline from the same evidence footprint.Even the expensive (24, 12) configuration only marginally exceeds the narrow-iterative baseline.
Fresh evidence: At (24, 12), rotation grounds the portfolio on 50.9 documents after exposing 288, versus 10.3 grounded documents under fixed-context resampling.The result supports the conclusion that repeated sampling cannot create genuine diversity from a stagnant prompt.
Scale and capacity: +0.101 and +0.157 gains persist when T rises from 1 to 12 at k=24 with a 1-million-token capacity, while expanding beyond k=24 yields no significant benefit.On ASQA, increasing k to 48 actively degrades performance.
Task dependence: On HotpotQA, twelve sequential generations reduce the width effect from +0.333 at T=1 to −0.046 by assembling multi-hop evidence across rounds.The iterative (2, 12) allocation dominates the monolithic (24, 1) allocation under matched hardware budgets.
Relevance decay: Width utility tracks head-5 answer-mass share: widening remains viable for QAMPARI at 0.398 but becomes toxic for ASQA at 0.713 and HotpotQA at 0.630.Low-yield tail documents dilute attention in concentrated tasks.
Closed-loop scheduling: Ascp reaches 0.309 terminal portfolio recall, gaining +0.033 over Carriage and up to +0.081 over selection baselines.The pure scheduler retains gains of +0.024 to +0.072 over the evaluated selection methods.

9 Conclusion

The paper identifies flawed evidence-use measurement and static context allocation as central RAG bottlenecks, then proposes causal auditing and sequential, feedback-driven orchestration. Its results support iterative generation and closed-loop scheduling over monolithic context widening, while documenting theoretical guarantees and empirical robustness boundaries.

Causal measurement: Standard IR proxies collapse on same-query hard negatives, whereas an intervention-based causal probe maintains robust discrimination and audits genuine evidence consumption.The probe also calibrates generative-attention dilution, yielding a canonical width elasticity of −0.68 under controlled isolation.
Context allocation: 16.8 to 20.5 absolute percentage points: sequential generation drives major portfolio recall gains, while wider contexts suffer rank degradation and unstable returns.The factorial analysis identifies sequential generation count T as the true causal allocation factor.
Closed-loop orchestration: The attribution-steered submodular scheduler uses causal feedback to allocate context toward fresh evidence and dominates evaluated open-loop selection baselines.Its objective is strictly monotone non-decreasing and submodular for any fixed causal feedback state.
Closed-loop orchestration: An orthogonal contrastive decoder counters attention inertia, while multi-round orchestration maximizes the generative yield of retrieved pools.The full attribution-steered decoder adds +0.0107 absolute evidence-consumption performance in ablation.
Scope and validation: Theoretical selection-versus-rotation regimes depend on relevance decay, but empirical utilization is over-dispersed and cross-round correlated by structural redundancy.The mixed-effects test rejects policy invariance with χ2 = 171.4, p = 6 × 10^-31, and finds over-dispersion by a factor of 1.39.
Scope and validation: Under a severely constrained N = 30 pool, budget-matched gains shrink to ∼+0.065, although cross-seed standard deviation remains ≤0.011.The restricted pool starves sequential iterations of fresh evidence, compressing structural gaps while preserving low variance.

B.2 Multiplicity Control and Redundancy Stratification

The analysis controls multiple-comparison risk and stratifies retrieval pools to test whether evidence-utilization dilution persists independently of redundancy. Results show steep decay across low-redundancy conditions, while redundancy modulates but does not create the penalty.

Multiplicity Control: 108 of 123 executed statistical tests survive Benjamini-Hochberg FDR correction unchanged.The correction used strict BH-FDR control at α=0.05.
Redundancy Stratification: The factorial analysis stratifies 2,880 query-level observations across six context widths using entailment, answer attestation, and embedding similarity.The redundancy measures exhibit loose internal correlation of r=0.12–0.35.
Redundancy Stratification: 0.575 to 0.164: under strict pairwise entailment filtering, utilized fraction q declines as context width increases from k=2 to k=24.The corresponding elasticity is −0.528 (0.021).
Redundancy Stratification: 0.561 to 0.204: zero duplicated answer-bearing passages still produce utilization decay across context widths.The corresponding elasticity is −0.434 (0.024).
Redundancy Stratification: −0.667: intense answer attestation steepens the decay slope, but redundancy cannot independently manufacture the context-expansion dilution penalty.The log k × interaction is highly significant at p=2 × 10^-7.
Redundancy Stratification: A steep decay persists in every low-redundancy stratum, although natural-pool free-generation slopes are shallower than the protocol-clean −0.68.The relevant evidence is the persistence of decay rather than its absolute magnitude.

C Reproducibility and Experimental Artifacts

The paper documents controlled artifacts, frozen configurations, validated data construction, extensive evaluation grids, and reproducibility procedures. Across settings, sequential generation remains above matched single-response baselines, while flat relevance density favors wide-context scheduling.

Reproducibility Controls: The frozen configuration β=0.3, βdoc=0.3, λ=0.25, κ=0, α=0.5, k=5, and N=30 was selected through one-at-a-time sensitivity sweeps.Hyperparameters were stabilized on a disjoint ASQA development split and locked before primary held-out evaluations.
Dataset Construction: Validated pools contain at least m+2 answer-bearing documents and twenty-four verified distractors, with deterministic shuffling preserving labels across k+1 ablations.Answer-bearing documents require contiguous normalized answer aliases, and distractors are sampled from unrelated queries.
Experimental Scale: 188 unified execution runs and 11,579 individually scored generation rows support the factorial evaluations.A standard single evaluation run operates at 126 seconds per query.
Cross-Cultural Stress Test: −0.01 relevance-density slope: the cross-cultural recipe setting is nearly flat, insulating wide-context configurations from deep-rank decay.Unconstrained rotational scheduling achieves peak portfolio extraction in this isolated regime.
Sequential Generation: +0.142 to +0.200: increasing generation count from T=1 to T=12 continues to yield large gains under the integrative instruction.The integrate prompt raises marginal recall by +0.050 to +0.098 but does not alter the factorial scaling laws.
Sequential Generation: +0.109 to +0.035: the matched single-response baseline trails the sequential T=12 portfolio as context width expands from k=2 to k=24.The deficit remains statistically significant across all 480 paired queries.

D.3 Comprehensive System Ablations: The Necessity of Causal Feedback

Component ablations show that attribution feedback, submodular selection, and steered decoding contribute distinct structural benefits. Replacing the causal evidence-use signal causes scheduling efficacy to collapse, making causal feedback central to bounded-budget coverage.

Component Ablations: The ablation grid isolates the marginal utility of attribution feedback, submodular selection, and steered decoding across four evaluated tasks.These structural substitutions provide the component-level comparison for the Ascp architecture.
Probe Substitutions: Replacing the causal probe with embedding similarity or uniform-utilization assumptions triggers a systemic collapse in scheduling efficacy.The scheduler is held fixed while only the evidence-use signal is substituted.
Causal Feedback: Counterfactual causal sensitivity is the irreplaceable driver of maximizing evidence coverage within a bounded generative budget.This conclusion mirrors the diagnostic illusion identified earlier in the evaluation pipeline.