Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems

Pascal Sowietzki; Steffen Holter; Yannick Metz; Emil Kasper; Raphael Baur; Mennatallah El-Assady

cs.MAOct 9, 2026

Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems

Authors: Pascal Sowietzki, Steffen Holter, Yannick Metz, Emil Kasper, Raphael Baur, Mennatallah El-Assady

Organizations: ETH Zurich, Switzerland

Abstract

To balance task performance and computational cost, agentic workflows can use a lower-cost model for execution and selectively consult a stronger model. In such executor-advisor systems, the executor decides when to seek advice, making escalation depend on self-assessment, although current language models struggle to detect where their own reasoning fails. The need for stronger reasoning emerges during execution and varies across workflow phases. We introduce ARIA, a meta-controller that learns when to consult by balancing task success against consultation cost. Our approach formulates consultation as a sequential resource-allocation problem. At workflow contact points, ARIA maps execution evidence to a compact state abstraction and selects SILENT or CONSULT, with action values learned by Monte Carlo control from cost-penalized outcomes. Across evaluated tasks of SWE-bench Verified, Terminal-Bench 2.1, and $τ^2$ -bench, ARIA recovers 44-81% of the executor-advisor success gap at less than half the advisor model's cost. ARIA achieves the best observed success-cost trade-off among ten other evaluated conditions, including routing and cascading methods, fixed and prompted consultation, and random placement at matched advice volume. The policies are compact and inspectable, with a single consultation-price parameter governing the success-cost trade-off without retraining either underlying model.

Figures & tables

Figure 1: ARIA setup. At each contact point, the meta controller chooses between Silent and Consult . Consultation adds guidance from a stronger advisor to the shared context.

Condition	Allocation rule	Tests for
Executor-only	no advisor	lower-cost reference
Advisor-only	advisor executes	stronger-model reference
FrugalGPT ( Chen et al., 2024a )	score-based cascade	cost-aware task-level cascading
RouteLLM ( Ong et al., 2025 )	pre-task routing	learned task-level routing
AutoMix ( Aggarwal et al., 2024 )	verify then escalate	adaptive task-level escalation
Plan ( Mozannar et al., 2025 )	first contact	early intervention

Table 1: Baseline design for selective consultation, covering capability references, task-level allocation, fixed schedules, LLM-based timing, uninformed placement, and learned allocation.

	SWE	Terminal	$\tau^{2}$
PGR	68.2%	44.4%	80.6%
Cost / Pro	48.1%	40.9%	34.5%
Cost Share	30.8%	29.6%	15.5%
Sav. vs. equal-succ. mix	$12.86	$7.59	$11.36

Table 3: ARIA relative to the autonomous anchors.

	$\Delta S$	$\Delta C$	$\Delta N_{c}$	$\Delta$ ER
SWE	+20.7	+$6.42	+1.95	+0.28
TB	+4.0	+$6.74	+0.82	0.00
$\tau^{2}$	-2.5	+$5.04	+5.78	-0.38

Table 4: Change from Always-

B4

B10

Figure 2: Success-cost operating points on SWE-bench for ARIA policies with $\lambda\in\{0.1,0.35,0.5\}$ , consultation-budget ablations, and reference conditions.

Figure 3: Primary failure mode across baselines and benchmarks grouped into system-design, inter-agent misalignment, and verification failures, showing different residual failure profiles.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	$\bar{c}$	$\Delta$	$\kappa$
SWE-bench	0.0490	0.1663	0.029465
Terminal-Bench	0.0683	0.2816	0.024254
$\tau^{2}$ -bench	0.0244	0.1281	0.019048

Appendix

Table 5: Benchmark-specific cost calibration for the headline policies at

\lambda=0.10

. The quantities

\bar{c}

and

\Delta

are in USD;

\kappa=\lambda\bar{c}/\Delta

is the per-consultation penalty in reward units.

	SWE-bench	Terminal-Bench	$\bm{\tau^{2}}$ -bench
Controller seeds	3	3	3
Training episodes	$\sim$ 150	162	160
Training tasks	$\sim$ 50	27	20
Approx. passes	3	6	8
Discount $\gamma$	1	1	1
MC update	$\alpha=1/N(s,a)$

Appendix

Table 6: Controller-training budgets and shared hyperparameters. Evaluation uses a frozen greedy policy without exploration.

Benchmark	Development source	Train tasks	Train episodes	Held-out test
SWE-bench	screened benchmark training pool	50-52	150-156	50 (Verified-Mini)
Terminal-Bench	39-task development partition	27	162	50 (test)
$\tau^{2}$ -bench	official telecom train split	20	160	40 (telecom test)

Appendix

Table 7: Development and evaluation splits used for policy learning and final evaluation. Train–test overlap was verified from task identifiers rather than assumed from split names.

Method	Allocation rule	Budget
Executor-only	no advisor	–
Advisor-only	advisor executes	–
FrugalGPT	score-based cascade	–
RouteLLM	pre-task routing	–
AutoMix	verify then escalate	–
Plan	first contact	$B=1$

Appendix

Table 8: Baseline contracts.

B

denotes the maximum number of within-trajectory advisor consultations per episode and therefore applies only to consultation-based conditions.

	SWE-bench	Terminal-Bench	$\bm{\tau^{2}}$ -bench
ARIA consults/episode	2.780	2.460	3.042
Consultation probability $p$	0.49	0.62	0.172
Random consults/episode	2.527	2.417	2.670
Random / ARIA volume	90.9%	98.2%	87.8%

Appendix

Table 9: Calibration and realized consultation volume of the Random placement control relative to the headline ARIA policy (

\lambda=0.10

Model	Input	Cached input	Output	Reasoning	Tier threshold	Above threshold
gemini-3-flash-preview	0.50	0.05	3.00	3.00	none	none
gemini-3.1-pro-preview	2.00	0.20	12.00	12.00	200,000 prompt tokens	in 2.00 / cached 0.20 / out 6.00

Appendix

Table 10: Pricing used for the reported evaluation costs, in USD per million tokens. The final column reports the incremental surcharge applied above the listed context threshold. Thus, for gemini-3.1-pro-preview , the effective rates above 200K prompt tokens are

4.00 for fresh input,

0.40 for cached input, and $18.00 for output and reasoning tokens.

Benchmark	Condition	Ep.	Calls	Input	Cached	Output	Reason.	Total	Cached %	Cost/run
SWE-bench Verified	Executor-only	150	2,440	36.825	12.397	4.251	0.880	54.354	25.2	11.48
SWE-bench Verified	Advisor-only	150	3,367	28.609	43.244	2.164	2.886	76.903	60.2	45.90
SWE-bench Verified	FrugalGPT	196	3,470	43.958	21.112	4.850	1.675	71.596	32.5	22.61
SWE-bench Verified	RouteLLM	150	2,734	29.913	21.392	3.524	1.924	56.753	41.7	25.03
SWE-bench Verified	AutoMix	257	4,945	66.469	66.846	5.954	3.133	144.438	50.1	59.68
SWE-bench Verified	Plan	150	2,736	31.402	9.571	3.782	0.435	45.190	23.4	10.90

Appendix

Table 11: Token usage and realized model cost for all headline evaluation conditions. Token columns are totals across three evaluation seeds in millions of tokens. “Input” denotes fresh input excluding cached prompt tokens; “Cached %” is the cached share of total prompt tokens. “Cost/run” is the mean realized cost across the three evaluation seeds, corresponding to Table 2. For cascades, the episode column counts billed component executions rather than unique benchmark tasks.

	SWE-bench Verified		Terminal-Bench 2.1		$\bm{\tau^{2}}$ -bench Telecom
Condition	Succ. (%)	Cost ($)	Succ. (%)	Cost ($)	Succ. (%)	Cost ($)
Executor-only	$18.0\pm 4.0$	$11.48\pm 2.30$	$21.3\pm 3.1$	$9.76\pm 1.90$	$72.5\pm 4.3$	$5.07\pm 0.04$
Advisor-only	$62.0\pm 7.2$	$45.90\pm 5.30$	$51.3\pm 1.2$	$60.99\pm 5.05$	$98.3\pm 2.9$	$22.50\pm 1.22$
FrugalGPT	$31.3\pm 2.3$	$22.61\pm 1.44$	$22.0\pm 4.0$	$11.10\pm 2.89$	$78.3\pm 2.9$	$7.02\pm 0.40$
RouteLLM	$30.0\pm 5.3$	$25.03\pm 2.35$	$28.0\pm 3.5$	$21.66\pm 0.85$	$88.3\pm 3.8$	$9.61\pm 0.12$
AutoMix	$52.0\pm 3.5$	$59.68\pm 5.75$	$22.0\pm 2.0$	$16.51\pm 4.21$	$97.5\pm 2.5$	$24.98\pm 0.64$

Appendix

Table 12: Task success and realized model cost across three evaluation seeds. Entries report mean

\pm

sample standard deviation (

n=3

Figure 4: Task success and consultation frequency for ARIA , Heuristic , and Random across all three benchmarks.

Figure 5: Training dynamics across three benchmarks and three controller seeds. Top: per-episode consultation rates (faint points) and smoothed trajectories (solid lines). Bottom: cumulative changes in the greedy tabular policy. Dotted vertical lines mark the last observed greedy-policy change for each seed, not a certified convergence point.

Figure 6: Learned action values for SWE-bench Verified across three independently trained controller seeds. Left: compact numerical tables reporting state visit counts and action values, where $Q_{S}=Q(s,\textsc{Silent})$ and $Q_{C}=Q(s,\textsc{Consult})$ . State notation is phase/evidence/prior consultation : F/E/R/D denote Frame , Explore , Reason , and Deliver ; N/C/X denote None , Clean , and Error . Structurally unreachable states are omitted, while reachable states with $N=0$ are retained. Right: the corresponding learned value tables for the three seeds visualized as heatmaps. Deployment selects the action with the larger value within each state; ties resolve to Silent .

Figure 7: Learned action values for Terminal-Bench 2.1 across three independently trained controller seeds. Left: compact numerical tables reporting state visit counts and action values, where $Q_{S}=Q(s,\textsc{Silent})$ and $Q_{C}=Q(s,\textsc{Consult})$ . State notation is phase/evidence/prior consultation : F/E/R/D denote Frame , Explore , Reason , and Deliver ; N/C/X denote None , Clean , and Error . Structurally unreachable states are omitted, while reachable states with $N=0$ are retained. Right: the corresponding learned value tables for the three seeds visualized as heatmaps. Deployment selects the action with the larger value within each state; ties resolve to Silent .

Figure 8: Learned action values for $\tau^{2}$ -bench Telecom across three independently trained controller seeds. Left: compact numerical tables reporting state visit counts and action values, where $Q_{S}=Q(s,\textsc{Silent})$ and $Q_{C}=Q(s,\textsc{Consult})$ . State notation is phase/evidence/prior consultation : F/E/R/D denote Frame , Explore , Reason , and Deliver ; N/O/F denote None , OK , and Failed . Structurally unreachable states are omitted, while reachable states with $N=0$ are retained. Right: the corresponding learned value tables for the three seeds visualized as heatmaps. Deployment selects the action with the larger value within each state; ties resolve to Silent .

Position	Evidence	Prior consult	SWE	TB	$\bm{\tau^{2}}$	All
Frame	None	No	0/3	2/3	2/3	4/9
Explore	None	No	1/3	0/3	1/3	2/9
	None	Yes	0/3	1/3	0/3	1/9
	Clean/OK	No	2/3	1/3	2/3	5/9
	Clean/OK	Yes	2/3	0/3	1/3	3/9
	Error/Fail	No	1/3	1/3	–	2/9

Appendix

Table 13: Cross-seed agreement of the learned greedy consultation policies. Entries report the number of seeds whose frozen table selects Consult in a given state. A dash indicates that the state was never visited during training for that benchmark. For

\tau^{2}

, OK and Failed correspond to the Clean and Error evidence classes used on SWE-bench and Terminal-Bench.

Figure 9: Propensity-weighted effect of Consult versus Silent within each state cell. Points show Hájek IPW estimates with 95% episode-cluster bootstrap intervals. Marker shape indicates the greedy action of the reference $Q$ -table; circled markers use the episode-level outcome because the one-step outcome is constant or unavailable. Faded intervals include zero. Parentheses report the number of observed decisions in each cell.

Figure 10: Secondary failure modes across baselines and benchmarks grouped into system-design, inter-agent misalignment, and verification failures, showing different residual failure profiles.

Explore similar work

AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows
Sep 19, 2026Rudrendu Kumar Paul, Sourav NandyLLM Agent Routing Cost-Aware Inference
Boston University, Boston, MA, USA · University of Texas at Austin, Austin, TX, USA
Certified Predictive Value-of-Advice Gating for Cost-Aware Language-Model Guidance in Reinforcement Learning
Aug 26, 2026Ibne Farabi Shihab, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al AhsanLLM-Guided RL Cost-Aware Inference
Department of Computer Science, Iowa State University · Standard Chartered, Bangladesh · Department of Computer Science and Engineering, BRAC University, Bangladesh
Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task Allocation
Jul 10, 2026Kaiji Zhou, Aleš Leonardis, Yue FengLLM Agent Orchestration LLM Agent Routing
University of Birmingham

Sep 19, 2026, cs.AIJ/K move · Enter open · S save

AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows

Rudrendu Kumar Paul, Sourav Nandy

Boston University, Boston, MA, USA · University of Texas at Austin, Austin, TX, USA

LLM Agent Routing Cost-Aware Inference Agentic Workflows

Open paper AskPDF arXiv

Enterprise agentic systems that route every trajectory step to a frontier model waste 60-80% of their inference budget on subtasks that smaller models handle equally well. Existing routing solutions optimize single-turn query assignment but ignore a property unique to agentic workflows: subtask complexity varies widely within a single trajectory. A planning step may require frontier-class reasoning while a subsequent formatting step needs only a 7B model. We formalize step-level model routing as a sequential assignment problem over agent trajectories and propose AgentRouter, a lightweight classifier (12M parameters, <5ms overhead per step on an A100 GPU) that maps each trajectory step to one of four model tiers using five features extractable at routing time. Trained on 50,000 annotated agent trajectory steps spanning planning, coding, research, and data analysis tasks, AgentRouter achieves 72% cost reduction relative to frontier-only baselines, retaining 97.3% of frontier-only quality (less than 3% degradation in end-to-end task completion); per-step routing accuracy reaches 91% on minimal-complexity steps and 85% on efficient-tier steps, with 76-82% on the harder mid-range and frontier tiers. On the same benchmarks, RouteLLM and FrugalGPT (applied per-step) achieve only 31% and 44% cost reduction respectively, because their single-turn training signal misses trajectory-level quality dependencies.

Abstract

Explore similar work

AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows

Certified Predictive Value-of-Advice Gating for Cost-Aware Language-Model Guidance in Reinforcement Learning

Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task Allocation