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

Peiyang Liu; Xi Wang; Di Liang; Wei Ye

cs.LGAug 24, 2026

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

Authors: Peiyang Liu, Xi Wang, Di Liang, Wei Ye

Organizations: National Engineering Research Center for Software Engineering, Peking University, China · Peking University, China · Tencent, China

Abstract

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.

Figures & tables

Figure 1. The context-allocation problem in one picture. (a) A fixed inference budget of

k\times T

evidence slots can be packaged as one wide context (

k{=}24

T{=}1

) or as many narrow contexts that rotate through fresh evidence (

k{=}2

T{=}12

). Both consume

24

retrieved documents, yet the second packaging covers

+0.144

more of the answer space (Table 4 ). (b) Turning that allocation law into a scheduler: on a frozen held-out frame, Ascp beats every selection-style baseline by

+0.033

+0.081

portfolio recall (all BH

q<.001

); the two grey arms are reference configurations whose remaining gaps are not significant (Table 10 ). (c) None of this is measurable with relevance proxies. Swapping the padding of the evaluation pool from off-query distractors to same-query hard negatives that entail no answer leaves the causal leave-one-out probes almost unchanged (

-0.01

AUC), while BM25 and query–document cosine fall by more than

0.5

to chance (Figure 3 ).

		vs. PR@ $T$			vs. groundedness
Family	Measure	$n$	$r$	$p$	$r$	$p$
coverage	Evidence coverage rate	23682	+0.097	2.8e-50	-0.167	9.7e-148
	Utilisation concentration	23682	-0.023	3.0e-04	-0.146	2.7e-112
diversity	Semantic diversity	23682	+0.089	6.1e-43	-0.111	9.3e-66
	Distinct-1	23682	+0.106	3.7e-60	-0.100	1.7e-53
	Distinct-2	23682	+0.128	2.4e-87	-0.134	9.4e-96

Table 1. Within-query correlations across systems (

n{=}2395

task–model–query groups, 23682 observations), centered within group. Lower panel: partial correlations between portfolio recall and one measure after residualising against paired coverage or output-diversity.

\ddagger

marks a measure whose sign is not constant across tasks, so its pooled value should not be read alone.

Figure 2. The closed-loop measurement and orchestration pipeline. The scheduler transforms a retrieved pool into a curated context per generation round. Subsequently, our causal probe isolates true evidence utilization via

k+1

parallelizable teacher-forced forward passes, completely bypassing autoregressive decoding overhead. These causal signals dynamically populate the utilization matrix, forming the exact feedback loop that governs subsequent submodular scheduling and attribution-steered cognitive decoding. Three linked panels describing the measurement pipeline. The first panel shows scheduling policies. The second panel shows the causal ablation probe costing k+1 teacher-forced passes. The third panel shows the utilization matrix and dynamic levers.

Figure 3. Probe validation and controls on constructed pools. (a) – (b) AUC for recovering the designed answer-bearing documents, by context width, under off-query and same-query padding; the ordering of causal estimators and relevance proxies reverses with the negative population. (c) Leave-one-out against ContextCite by padding regime, fixed protocol. (d) cover@

m

gain over a random ranking with

95\%

cluster-bootstrap intervals, a target invariant to which answer-equivalent document was labelled necessary. (e) Transfer of the off-query

95

th-percentile operating point to same-query passages that contain and entail no gold answer; only fixed-target curves are false-positive rates, since under free generation the padding may legitimately shape the response. Five panels. Panels (a) and (b) plot estimator AUC against context width for off-query and same-query padding. Under off-query padding the three relevance proxies sit near one and above both leave-one-out curves; under same-query padding BM25 and query-document cosine fall to or below chance while the leave-one-out curves are unchanged. Panel (c) compares leave-one-out with ContextCite by padding regime under the fixed protocol. Panel (d) shows set-coverage gain over a random ranking with 95 percent intervals: the three counterfactual estimators are clearly positive and the three relevance proxies straddle zero. Panel (e) shows exceedance of the off-query threshold on same-query near-misses under fixed and free targets.

Figure 4. Width elasticity of the thresholded utilisation statistic

q

, with

95\%

intervals. (a) By calibration criterion: thresholds that do not themselves shrink with width agree on the protocol-clean value, whereas free-generation and threshold-free criteria estimate a different quantity. (b) By natural-pool redundancy stratum: the decline survives in every low-redundancy stratum, so redundancy modulates but does not explain it. Two forest plots of the width elasticity of the thresholded utilisation statistic. The left panel varies the calibration criterion: criteria whose threshold does not shrink with width cluster near minus 0.68, while free-generation and threshold-free criteria are much closer to zero. The right panel varies the natural-pool redundancy stratum; every stratum remains substantially negative.

policy	ECR	used	offered	PR@ $T$
Ascp scheduler	0.626	6.76	10.55	0.300
deep rotation	0.375	9.37	25.00	0.303
vanilla RAG	0.578	2.89	5.00	0.237
MMR	0.587	3.13	5.28	0.239
DPP-RAG	0.526	4.71	8.83	0.274
xQuAD	0.571	2.85	5.00	0.238

Table 2. Policy dependence with one ordinary decoder. All arms use the same explicit sampler; full custom decoder excluded. ECR uses fixed-response threshold and equal task weights. Across the 8 evaluated policies, ECR spans a range of 0.251, with the vast majority of paired ECR contrasts surviving strict BH-FDR correction. Offered is a treatment, not an outcome.

Task	gold units	$\rho(1..5)$	$\rho(11..30)$	head-5 share	head-5 share	$\mathrm{d}\log\rho/\mathrm{d}\log i$
	per query			(any)	(new)
ASQA	3.4	0.469	0.211	0.290	0.713	-0.40
QAMPARI	13.6	0.295	0.170	0.247	0.398	-0.27
RECIPES	20.7	0.890	0.870	0.170	0.391	-0.01

Table 3. Relevance density

\rho(i)

: probability retrieved rank

i

attests a gold answer, from gold annotations over

200

queries/task. New head-5 share counts only answer mass not attested earlier. ELI5 omitted: free-text claims lack rank-level lexical attestation.

Figure 5. Deconfounded width–count grid. Solid lines represent rotating through fresh evidence; dashed lines repeatedly sample the same fixed context. Three panels show portfolio recall by width, recall by generation count for rotation and fixed-context controls, and their paired difference with 95 percent confidence intervals.

$k$	$T$	$kT$	offered		PR@ $T$		$\Delta$ PR@ $T$	95% CI
			rotation	fixed	rotation	fixed
2	1	2	2.0	2.0	0.204	0.204	+0.000	$[+0.000,+0.000]$
2	5	10	10.0	2.0	0.328	0.238	+0.090 †	$[+0.068,+0.113]$
2	12	24	24.0	2.0	0.397	0.257	+0.140 †	$[+0.113,+0.169]$
5	1	5	5.0	5.0	0.218	0.218	+0.000	$[+0.000,+0.000]$
5	5	25	25.0	5.0	0.339	0.266	+0.073 †	$[+0.052,+0.094]$

Table 4. Deconfounded width–count grid: rotation cycles disjoint rank windows; fixed repeats top-

k

for

T

samples.

\Delta

PR@

T

is paired within query and bootstrapped by query, isolating fresh-evidence gain. Pool

N=400

keeps

kT\leq 288<N

with no repeats; earlier

N=30

made

T{=}12

offer the same

24

–

30

docs. ∗ / † :

p<0.05

p<0.01

condition	count T12 vs T1 at k24	count T12 vs T1 at k2	width k24 vs k2 at T12	width k24 vs k2 at T1	rot over fix at k24T12
asqa/llama	+0.204 †	+0.289 †	-0.061 ∗	+0.023	+0.066 †
asqa/qwen	+0.144 †	+0.229 †	-0.042	+0.043	+0.074 †
qampari/llama	+0.187 †	+0.141 †	+0.139 †	+0.093 †	+0.117 †
qampari/qwen	+0.134 †	+0.115 †	+0.058 †	+0.039 †	+0.082 †

Table 5. Decisive deconfounded-grid contrasts by task and generator, as within-query paired portfolio-recall differences. Last column: rotation over matched fixed-context control. ∗ / † :

p<0.05

p<0.01

estimate	$T=1$	$T=5$	$T=12$
asqa/llama	$+0.036$	$-0.033$	$-0.053$
asqa/qwen	$+0.047$	$-0.000$	$-0.038$
qampari/llama	$+0.081$	$+0.148$	$+0.134$
qampari/qwen	$+0.041$	$+0.066$	$+0.057$
grand mean	$+0.051$	$+0.045$	$+0.025$
between-condition SD	$+0.025$	$+0.083$	$+0.090$

Table 6. Hierarchical rotation width contrast

k=24

minus

k=2

, paired within query. Condition rows are partially pooled task–generator effects (Paule–Mandel); last rows predict a new condition. With

K=4

, interval uses Higgins–Thompson–Spiegelhalter

t_{K-2}

, not normal. At

T=12

the mean is small and prediction interval wide;

K=4

limits precision.

contrast	ASQA		QAMPARI
	$\Delta$	$n$	$\Delta$	$n$
$k{=}48$ vs $k{=}24$ , $T{=}1$	$-0.033$ ∗	120	$+0.004$	120
$k{=}96$ vs $k{=}24$ , $T{=}1$	$-0.031$	30	$+0.040$	29
$k{=}48$ vs $k{=}24$ , $T{=}5$	$-0.030$	58	$+0.002$	59
$T{=}5$ vs $T{=}1$ , $k{=}24$	$+0.089$ †	60	$+0.112$ †	60
$T{=}12$ vs $T{=}1$ , $k{=}24$	$+0.101$ †	60	$+0.157$ †	60

Table 7. Long-context Qwen2.5-7B-Instruct-1M width–count contrasts at widths beyond the

7

–

8

B grid;

k=96

is about

11

K prompt tokens. Entries are within-query paired PR@

T

differences; ∗ / † :

p<0.05

p<0.01

. Generation-count gains remain positive (

+0.085

+0.157

); no width gain beyond

k=24

is positive/significant, and ASQA worsens.

task	$\|A\|$	head-5	$\rho$ top-5	width $\Delta$	width $\Delta$	count $\Delta$
		share		at $T{=}1$	at $T{=}12$	at $k{=}2$
ASQA	3.4	0.713	0.469	$+0.033$	$-0.052$ ∗	$+0.259$ †
QAMPARI	13.6	0.398	0.295	$+0.066$ †	$+0.099$ †	$+0.128$ †
ELI5	–	–	–	$+0.011$	$+0.010$ †	$+0.242$ †
HotpotQA	1.0	0.630	0.169	$+0.333$ †	$-0.046$ ∗	$+0.450$ †

Table 8. Relevance density versus gold-set size over four task profiles; the multi-hop HotpotQA replaces the recipe stress test here so that gold-set size and relevance density vary independently.

|A|

mean gold answers,

\rho

rank-band answer probability, head-5 concentration. At

T{=}12

, top-5 share tracks width: only dispersed QAMPARI (0.398) gains, not ASQA (0.713) or HotpotQA (0.630) despite gold sets

3.4

1.0

; with one generation, multi-hop gains

+0.33

. ELI5 string

\rho

is not estimable. ∗ / † :

p<0.05

p<0.01

task	generator	count $T$ at k2	count $T$ at k24	width $k$ at T12	budget (2,12) vs (24,1)	rot over fix at k24T12
eli5	llama	$+0.228$ †	$+0.156$ †	$-0.087$ †	$+0.242$ †	–
eli5	qwen	$+0.256$ †	$+0.292$ †	$+0.072$ ∗	$+0.219$ †	–

Table 9. ELI5 factorial: small claim gold set but dispersed BM25 profile, separating relevance density from gold-set size; HotpotQA extension in Table 8 . Entries are within-query paired PR@

T

differences; ∗ / † mark

p<0.05

p<0.01

by paired bootstrap.

baseline	baseline PR	Ascp PR	$\Delta$ PR	BH $q$	$\Delta$ ECR	$\Delta$ ground
Ascp without steering	0.300	0.309	+0.009	$0.102$	+0.056	+0.029
deep rotation	0.303	0.309	+0.006	$0.343$	+0.307	+0.050
vanilla RAG	0.237	0.309	+0.072	$<.001$	+0.104	+0.003
MMR	0.239	0.309	+0.070	$<.001$	+0.095	+0.011
DPP-RAG	0.274	0.309	+0.035	$<.001$	+0.156	+0.011
xQuAD	0.238	0.309	+0.071	$<.001$	+0.111	+0.002

Table 10. Frozen held-out comparison: steering chosen once by one-standard-error on development, then tested on disjoint queries (

100

per task), four tasks, three generators, two decoding seeds. Contrasts are paired on query/generator/seed/prompt/width/count/code. Sampling: temperature

0.7

, top-

p=1

, top-

k

disabled, repetition penalty

1

. Crossed bootstrap, equal task weights, BH-FDR for PR@

T

; ECR fixed-response. Rounded differences may shift

0.001

; audit requires 2400 complete cells.

baseline	baseline PR	scheduler PR	$\Delta$ PR	BH $q$
deep rotation	0.303	0.300	-0.003	$0.717$
vanilla RAG	0.237	0.300	+0.063	$<.001$
MMR	0.239	0.300	+0.061	$<.001$
DPP-RAG	0.274	0.300	+0.026	$0.002$
xQuAD	0.238	0.300	+0.062	$<.001$
PM-2-RAG	0.228	0.300	+0.072	$<.001$

Table 11. Scheduler-only held-out comparison: Ascp without steered decoding, so both sides use ordinary generation and isolate evidence scheduling. Crossed bootstrap over decoding seeds and within-task queries; equal task weights, fixed three-model set; BH-FDR covers the eight PR@

T

contrasts.

contrast	$\Delta$ PR@ $T$	seed- $t_{4}$ CI	$p$	$\Delta$ ECR	$\Delta$ ground	$\Delta$ words
full decoder $-$ off	+0.0107	$[+0.0072,+0.0142]$	$0.001$	+0.0555	+0.0275	-1.0
contrast, APC held on	+0.0116	$[+0.0045,+0.0187]$	$0.011$	+0.0583	+0.0232	-0.6
APC only	-0.0009	$[-0.0086,+0.0069]$	$0.775$	-0.0028	+0.0042	-0.4
contrast, APC held off	+0.0187	$[+0.0145,+0.0229]$	$<.001$	+0.1012	+0.0366	-0.2
APC, contrast held on	-0.0080	$[-0.0107,-0.0052]$	$0.001$	-0.0457	-0.0092	-0.8
contrast $\times$ APC	-0.0071	$[-0.0163,+0.0021]$	$0.098$	-0.0429	-0.0134	-0.5

Table 12. Held-out decoder decomposition with identical sampling: temperature

0.7

, top-

p=1

, top-

k

disabled, repetition penalty

1

. APC is adaptive-plausibility constraint; contrasts isolate contrast direction, APC filtering, and complement. Intervals/

p

: two-sided

t

over five seed means (

df=4

), equal task weights, fixed three-model set; audit needs all 6000 frames; BH-FDR covers PR@

T

model	task	$(2,12)$	$(24,1)$	$\Delta$	95% CI
qwen14	asqa	0.569	0.401	$+0.168$	$[+0.107,+0.233]$
qwen14	pooled	0.400	0.266	$+0.134$	$[+0.093,+0.175]$
qwen14	qampari	0.231	0.131	$+0.100$	$[+0.050,+0.157]$
qwen32	asqa	0.560	0.380	$+0.180$	$[+0.116,+0.250]$
qwen32	pooled	0.429	0.291	$+0.138$	$[+0.098,+0.181]$
qwen32	qampari	0.298	0.202	$+0.097$	$[+0.055,+0.140]$

Table 13. Budget-matched scaling validation on 14B and 32B models. Entries represent paired portfolio recall, strictly contrasting the multi-round iterative generation

(2,12)

against monolithic wide-context generation

(24,1)

. The absolute structural superiority of multi-round evidence allocation robustly holds at scale, maintaining massive positive bounds across tasks.

Figure 6. Cost versus quality across every grid cell (Qwen/ASQA, A100, probe removed). (a) Total tokens per query; (b) Sequential latency. Solid lines represent rotation; dashed lines represent the matched fixed-context arm. The annotated pair is the budget-matched

24

-slot comparison. Crucially, the two cost axes disagree on the preferred packaging, illustrating why no single cell serves as a universal recommendation. Two panels plotting portfolio recall against measured cost on a logarithmic axis, one line per context width, solid for rotation and dashed for the matched fixed-context arm. The left panel uses total tokens per query and the right uses sequential latency. Narrow rotation curves dominate on tokens, while wide single-generation cells are cheapest in latency.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

contrast	seed 0	seed 1	seed 2	mean	s.d.
count T12 vs T1 @k2	0.141	0.137	0.154	0.144	0.0090
count T12 vs T1 @k24	0.159	0.160	0.141	0.153	0.0109
width k24 vs k2 @T12	0.097	0.099	0.090	0.095	0.0049
width k24 vs k2 @T1	0.092	0.094	0.097	0.094	0.0024
rot over fix @k2T12	–	–	–	–	–
rot over fix @k24T12	–	–	–	–	–

Appendix

Table 14. Seed replication of the factorial under a severely constrained

N=30

retrieval pool. Unlike the deep

N=400

pool in the main text (which prevents evidence exhaustion), this artificially restricted setting starves the sequential iterations of fresh evidence, naturally compressing the absolute structural gaps (e.g., the budget-matched gain shrinks to

\sim+0.065

). Crucially, despite this compressed effect size, the cross-seed standard deviation remains negligible (

\leq 0.011

). This proves that algorithmic stochasticity cannot explain our allocation laws, even under extreme resource starvation.

Figure 7. Judge meta-evaluation over

179

portfolios and

858

document-level judgements. (a) Identifying the supporting document: both the probe and the embedding proxy track the judge’s profile equally well overall, but the probe is significantly better at identifying the primary source. (b) A comparison of probe-derived and judge-derived coverage measures. (c) Correlation with the judge’s assessment of textual novelty. A response-only judge evaluates textual novelty by heavily tracking the distinct-

2

metric rather than actual evidence coverage. Three bar panels of Spearman correlations against an independent judge. The first compares the leave-one-out probe with an embedding proxy on document-level agreement, where the probe is ahead on top-1 agreement. The second shows probe-derived coverage measures against judge-derived ones. The third shows correlation with judge textual-novelty ratings, where distinct-2 leads.

Variant	ASQA		QAMPARI		ELI5		Recipes
	PR@ $T$	ECR	PR@ $T$	ECR	PR@ $T$	ECR	PR@ $T$	ECR
Ascp (full)	0.493	0.331	0.162	0.324	0.273	0.459	0.238	0.371
$-$ attribution feedback	0.465	0.320	0.163	0.309	0.248	0.459	0.242	0.398
$-$ submodular selection	0.470	0.264	0.158	0.265	0.257	0.408	0.220	0.319
$-$ steered decoding	0.488	0.325	0.161	0.321	0.278	0.465	0.231	0.378
steered decoding only	0.398	0.150	0.107	0.122	0.232	0.177	0.221	0.197

Appendix

Table 15. Component ablations and probe substitutions (qwen). Attribution-feedback removal leaves open-loop scheduling; probe-substitution rows keep the scheduler fixed and swap only the evidence-use signal.

Figure 8. One-at-a-time sensitivity sweeps conducted on the disjoint ASQA development split to ensure reproducible hyperparameter stabilization. The explicitly circled point indicates the selected optimal value for each distinct architectural parameter. The strictly frozen evaluation configuration deployed uniformly across primary experiments is

\beta=0.3

\beta_{\mathrm{doc}}=0.3

\lambda=0.25

\kappa=0

\alpha=0.5

k=5

, and

N=30

. Eight small panels, one per scheduler hyper-parameter, plotting portfolio recall against the parameter value on the ASQA development split. The best value in each panel is circled. All curves are shallow, so no parameter dominates the result.

$k$	$T$	commit PR@ $T$	integrate PR@ $T$	$\Delta$	95% CI
2	1	0.203	0.253	$+0.050$	$[+0.031,+0.070]$
2	5	0.326	0.394	$+0.068$	$[+0.049,+0.088]$
2	12	0.393	0.453	$+0.060$	$[+0.041,+0.080]$
5	1	0.219	0.292	$+0.073$	$[+0.053,+0.095]$
5	5	0.336	0.406	$+0.070$	$[+0.052,+0.091]$
5	12	0.411	0.475	$+0.065$	$[+0.044,+0.087]$

Appendix

Table 16. Utility-side factorial under commit versus integrate instructions. Integrate asks for every supported interpretation and allows enumeration. Differences are integrate minus commit, paired within query (

n=314

–

480

per cell), testing allocation under the instruction most favourable to one wide response.

instruction	$k$	$n$	portfolio	structured	gap	95% CI
commit	2	480	0.398	0.289	$+0.109$ †	$[+0.084,+0.135]$
	5	480	0.423	0.335	$+0.088$ †	$[+0.063,+0.114]$
	12	480	0.426	0.371	$+0.055$ †	$[+0.028,+0.082]$
	24	480	0.416	0.381	$+0.035$ †	$[+0.009,+0.061]$
integrate	2	480	0.398	0.295	$+0.103$ †	$[+0.077,+0.130]$
	5	480	0.423	0.342	$+0.081$ †	$[+0.056,+0.108]$

Appendix

Table 17. Enumerated single-response baseline versus matched

T{=}12

portfolio, paired within query. The single response gets the portfolio decode-token budget and must list distinct supported interpretations with grounding, stopping rather than padding. It remains below portfolios at every width/instruction (

+0.109

+0.029

), with narrowing gaps as context widens. ∗ / † :

p<0.05

p<0.01

Iterative RAG

See all 9 papers

Know Before You Fetch: Calibrated Retrieval-Budget Allocation for Retrieval-Augmented Generation
Jun 29, 2026Zhe Dong, Fang Qin, Manish Shah +1Retrieval-Augmented Generation Selective Prediction
University of Maine at Presque Isle · Stanford University · Independent Researcher
Less can be More: Relieving RAG Bottlenecks via Evidence Frontloading and Pressure-Adaptive Budgeting
Aug 25, 2026Weibin Cai, Reza ZafaraniRetrieval-Augmented Generation Adaptive RAG
Data Lab, EECS Department Syracuse University Syracuse, NY, USA
Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models: From In-Context Prompting to Causal Retrieval-Augmented Generation
Apr 3, 2026Prakhar Bansal, Shivangi AgarwalKnowledge Augmentation for Language Models Retrieval-Augmented Generation

Jun 29, 2026J/K move · Enter open · S save

Know Before You Fetch: Calibrated Retrieval-Budget Allocation for Retrieval-Augmented Generation

Zhe Dong, Fang Qin, Manish Shah, Yicheng Wang

University of Maine at Presque Isle · Stanford University · Independent Researcher

Retrieval-Augmented Generation Selective Prediction Adaptive RAG Adaptive Retrieval Language Model Calibration

Open paper AskPDF arXiv

Retrieval-augmented generation (RAG) typically retrieves a fixed number of passages for every query. This is wasteful when the reader already knows the answer, and it can be harmful when irrelevant or partially relevant passages distract the reader. We formulate adaptive RAG as calibrated retrieval-budget allocation: given a query, decide whether to answer closed-book, retrieve a compact context (k=1), retrieve a full context (k=5), or abstain. The contribution is a probability interface rather than a new raw uncertainty signal. We calibrate sequence log-probability and prefix-logit uncertainty signals into probabilities of correctness, then use these probabilities for graded context selection, selective abstention, and explicit latency/token trade-offs. Across core QA experiments on TriviaQA, Natural Questions, and MS MARCO, with auxiliary PopQA motivation and Qwen/Llama family checks, diagnostic out-of-fold calibration improves probability quality dramatically: for sequence log-probability, ECE drops from 0.275 to 0.062 on TriviaQA, 0.643 to 0.009 on NQ, and 0.711 to 0.031 on MS MARCO. Graded retrieval improves full-context and passage-budget frontiers for both our signal and TARG-style prefix entropy/margin, while retrieval-call AUC remains essentially tied with binary gating because k=1 is still a retrieval call. Held-out train/validation/test threshold experiments report deployable operating points. At matched-accuracy frontier operating points, a measured cost model reveals that gating is not universally faster: it increases latency by about 27% on Qwen3-8B but saves about 8% on Qwen3-32B. These results support a nuanced view of adaptive RAG: calibrated confidence is best understood as a reusable interface for allocating retrieval budget under task and system constraints.

Abstract

Iterative RAG

Explore similar work

Know Before You Fetch: Calibrated Retrieval-Budget Allocation for Retrieval-Augmented Generation

Less can be More: Relieving RAG Bottlenecks via Evidence Frontloading and Pressure-Adaptive Budgeting

Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models: From In-Context Prompting to Causal Retrieval-Augmented Generation