Limits of Intelligence — orchestration vs raw capacity

In this overview:

Quality means the task-specific metric reported by the source and is not assumed to be comparable across studies.

Cost means recorded or estimated inference cost under a stated model, provider, token count, and pricing date.

Latency means end-to-end elapsed time under a stated execution setup.

Reliability means repeatable task completion under a declared evaluation protocol, including failure, abstention, and repair behavior.

Auditability means the presence of inspectable evidence, state, traces, or decision records; it is a system property, not a benchmark score.

Routing and cascade studies demonstrate that model selection can improve cost–performance trade-offs when models have complementary strengths and the quality estimator can identify which model is sufficient for a query.

FrugalGPT reported up to 98% cost reduction while matching the strongest individual model in its evaluated setting. This is a study-specific maximum, not a universal routing effect, and the economics depend on the model set, task distribution, token usage, and pricing date.

Later work on unified cascade routing reinforces a central limitation: selection is only as useful as the estimator used to predict answer quality.

Routing can fail when the task distribution shifts, model versions change, costs change, or the router’s confidence signal is correlated with the same errors made by the routed model. A production router should therefore be recalibrated after material model or distribution changes and compared with a strong single-model baseline.

Cost-reporting fields

• provider and exact model version;
• input and output token counts;
• pricing date;
• retry and cache policy;
• number of calls;
• failed-call treatment;
• quality metric;
• whether cost includes routing and verification overhead.

1. Test-time computation can substitute for some pretraining scale in specific reasoning regimes, but the benefit depends on prompt difficulty and base-model competence. In a FLOPs-matched MATH analysis, compute-optimal inference with a smaller model outperformed a roughly 14-times-larger model on problems where the smaller model already had non-trivial success.

This does not establish that search dominates stronger models generally. On harder tasks, and when verifiers produce false positives or misrank valid paths, additional sampling or search can plateau or reduce net utility. 2. Test-time compute needs a competence floor.

Search can improve a candidate set only when the base model produces enough useful candidates and the system can distinguish them. More sampling does not repair a verifier that systematically accepts incorrect outputs.

Verifier choice changes the quality–compute trade-off. Process-level verification has outperformed outcome-only verification in some mathematical reasoning setups, while later theoretical and empirical work shows that the advantage depends on task structure, supervision quality, search procedure, and the verifier’s false-positive and false-negative profile.

Verifier-guided search should therefore report verifier accuracy separately from end-to-end system accuracy. A stronger search procedure cannot compensate for a verifier that consistently ranks incorrect paths above valid ones.

Verifier metrics

• verifier precision and recall;
• false-positive rate;
• false-negative rate;
• calibration;
• ranking metric;
• task-domain shift;
• number of candidates;
• acceptance threshold;
• end-to-end task outcome;
• compute spent per accepted answer.

Verifier choice

Reported effects are task- and setup-dependent. Process verifiers improved selection in some MATH experiments; imperfect process and outcome verifiers can both misrank candidates and lose effectiveness as search budgets grow.

1. Conformal methods can calibrate a coverage, violation, or risk target under the assumptions of the chosen procedure and calibration data. The parameter α should be described as the relevant miscoverage or violation tolerance, not as a rejection rate. Abstention and escalation rates must be measured empirically.

Any conformal gate used in a product should be recalibrated after meaningful distribution or model changes and monitored for subgroup or domain-specific coverage failures. 2. Conformal gates calibrated to a stated violation or miscoverage tolerance, with empirical reporting of escalation, abstention, and subgroup performance. 3. Coverage is not correctness for every case.

A marginal 90% coverage guarantee does not mean every accepted answer has a 90% probability of being correct. The exact interpretation depends on the conformal target, calibration design, assumptions, and population over which the guarantee is stated.

1. Selective retrieval asks whether external retrieval is needed for a query. Self-Routing RAG reports favorable accuracy–retrieval trade-offs in the authors’ evaluated setup, but the magnitude varies by benchmark and model.

RAFT addresses a different problem: adapting a model to use domain-specific retrieved evidence and ignore distractor documents. It should be cited as a retrieval-quality and domain-adaptation method, not as the source of a retrieval-triggering result. 2. Retrieval can reduce reliability when the evidence is irrelevant, stale, contradictory, incomplete, or ranked for lexical similarity rather than answer sufficiency. A retrieval policy should therefore evaluate at least three distinct questions:

1. Should the system retrieve?
2. Did it retrieve relevant and sufficient evidence?
3. Did the generator use that evidence faithfully?
4. Selective retrieval should be compared with a strong long-context baseline, but context-window size alone is not evidence that the model can identify and reason over the relevant information reliably.

Planner–executor–verifier architectures have improved task completion in several benchmarked agent systems, but the benefit depends on the planner, tool interface, verifier quality, rollback policy, and task structure. Verification hooks create opportunities for repair; they do not guarantee that the system detects the right failure or restores a valid state.

Agent-reliability metrics

• task success;
• pass^k or repeated-run reliability;
• policy compliance;
• tool-call validity;
• state correctness;
• recovery success;
• number of actions;
• human intervention;
• end-to-end latency and cost.

1. Static routing by task or content type

Useful when categories are stable and easy to identify. It fails when difficulty varies substantially within the category or the taxonomy becomes stale. 2. Learned routing

Estimate expected task utility for each model, including predicted quality, cost, latency, and failure or abstention penalties. The target and calibration method should be stated; a confidence score is not automatically a probability of correctness. 3. Post-hoc acceptance

Run a cheaper model first and escalate when an acceptance score fails. The main risk is a false-positive verifier that accepts a plausible but incorrect answer. 4. Unified cascade routing

Combine ex-ante model selection with post-hoc acceptance. The benefit depends on a quality estimator that remains valid across models and query types. 5. All routing policies should be re-evaluated after a model upgrade, provider change, prompt change, material price change, or shift in the live query distribution.

1. A controlled store for approved terminology, units, schemas, and normalization rules. Version it and record which rule affected each output.
2. State that connects an output or decision to the source passages, tool results, or records used to produce it. Evidence links improve inspectability but do not by themselves prove that the output is correct.
3. A retrieval layer for prior records or long-tail entities. Evaluate freshness, duplication, contamination, access control, and whether retrieved state is still valid for the current task.
4. • treating retrieved context as correct merely because it was retrieved;

• allowing obsolete state to persist without expiry or version checks;
• writing model-generated claims back into memory without validation;
• losing provenance when records are summarized or merged;
• assuming a long context window removes the need for retrieval evaluation;
• using one similarity score as both relevance and evidence-sufficiency signal.

Reported component effects across studies

Study-specific component comparisons

Effect: can improve majority-vote or selection metrics on tasks with extractable answers, at higher token and latency cost.

Boundary: more samples can amplify false-positive acceptance when the verifier is weak.

Early-stop self-consistency

Reported result: 33.8–84.2% fewer samples across six evaluated reasoning benchmarks with comparable task performance.

Boundary: strongest fit is for tasks where candidate answers can be compared or normalized reliably.

Calibrates a stated coverage or risk target under the method’s assumptions. Efficiency, retention, abstention, and subgroup behavior must be measured.

Peer-reviewed conference/journal Peer-reviewed workshop Preprint Official benchmark/report inAi synthesis Illustrative figure

1. Western Electric control-chart rules
2. These rules are one possible drift signal, not an AI-specific guarantee. Their increased sensitivity also raises false-alarm frequency.
3. can inform may reduce should be evaluated in the measured product setting
4. cuts cost improves auditability stabilizes output

A compact expandable glossary would improve the page.

Cascade

A sequence that begins with a cheaper or smaller model and escalates when an acceptance rule is not met.

Router

A policy that selects a model, tool, or workflow before or during execution.

Verifier

A model, program, test, or rule that scores, ranks, or accepts candidate outputs.

Outcome verifier / outcome reward model

A verifier that evaluates the final result.

Process verifier / process reward model

A verifier that evaluates intermediate reasoning steps or state transitions.

Selective retrieval

A policy that decides whether and where to retrieve external evidence rather than retrieving for every query.

Conformal calibration

A family of methods that calibrate coverage or risk targets under stated assumptions using held-out calibration data.

Abstention

A decision not to answer or act when an acceptance condition is not met.

Repair budget

A declared limit on retries, revisions, tool calls, or escalation attempts.

Test-time compute

Additional computation used during inference, such as sampling, search, revision, or verification.

Pareto frontier

A set of configurations for which improving one objective requires worsening another. The objectives and constraints must be explicitly defined.

---

Routing and cascades

1. Hu et al. RouterBench: A Benchmark for Multi-LLM Routing System. arXiv, 2024.
   Standardized router evaluation and a large recorded-outcome dataset.

2. Chen, Zaharia, and Zou. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. TMLR / arXiv, 2023–2024.
   Learned cascades and study-specific cost reductions.

3. Dekoninck, Baader, and Vechev. A Unified Approach to Routing and Cascading for LLMs. ICML 2025.
   Unified cascade routing and the importance of quality estimators.

4. Uddin and Bauer. Conformal LLM Routing with Distribution-Free Safety Guarantees. ACL Student Research Workshop, 2026.
   Post-vintage conformal routing result with explicit violation and confidence parameters.

Test-time compute and verifier limits

1. Snell et al. Scaling LLM Test-Time Compute Optimally Can Be More Effective Than Scaling Model Parameters. ICLR 2025.
   Difficulty-dependent compute allocation and conditional smaller-versus-larger model result.

2. Stroebl, Kapoor, and Narayanan. The Limits of Inference Scaling Through Resampling. Preprint, revised 2026.
   False-positive verifier ceilings and utility-limited resampling.

3. Yu, Li, and Wang. Scaling Flaws of Verifier-Guided Search in Mathematical Reasoning. Preprint, 2025.
   Diminishing and reversing verifier-guided-search gains.

4. Lightman et al. Let’s Verify Step by Step. ICLR 2024.
   Process supervision in mathematical reasoning.

5. Jia, Rakhlin, and Xie. Do We Need to Verify Step by Step?. ICML 2025 / preprint record.
   Theoretical challenge to universal process-supervision superiority.

6. Li et al. Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning. ICLR 2024.
   Reduced sampling under the authors’ reasoning-benchmark setup.

Uncertainty and conformal methods

1. Cherian, Gibbs, and Candès. Large Language Model Validity via Enhanced Conformal Prediction Methods. NeurIPS 2024.
   Conformal LLM-output validity, conditionality, and utility trade-offs.

2. Wang et al. ConU: Conformal Uncertainty in Large Language Models with Correctness Coverage Guarantees. Findings of EMNLP 2024.
   Conformal uncertainty for open-ended generation.

3. Kapoor et al. Calibration-Tuning: Teaching Large Language Models to Know What They Don’t Know. UncertaiNLP 2024.
   Calibration-oriented fine-tuning; use the exact workshop venue.

Retrieval and long context

1. Zhang et al. RAFT: Adapting Language Model to Domain Specific RAG. 2024.
   Domain-specific retrieval-augmented fine-tuning and distractor handling.

2. Wu et al. Self-Routing RAG: Binding Selective Retrieval with Knowledge Verbalization. Preprint, revised 2026.
   Selective source routing and study-specific accuracy–retrieval trade-offs.

3. Liu et al. Lost in the Middle: How Language Models Use Long Contexts. TACL 2024.
   Position-sensitive degradation in long-context use.

4. Modarressi et al. NoLiMa: Long-Context Evaluation Beyond Literal Matching. ICML 2025 / arXiv.
   Long-context degradation without easy lexical matching.

5. Bai et al. LongBench v2. ACL 2025.
   Realistic long-context reasoning across multiple task categories.

Multi-agent aggregation and evaluation

1. Wang et al. Mixture-of-Agents Enhances Large Language Model Capabilities. ICLR 2025.
   Multi-model aggregation and benchmark-specific automated preference results.

2. Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023.
   Utility and known judge biases.

3. Dubois et al. Length-Controlled AlpacaEval. COLM 2024.
   Length-bias mitigation in automated preference evaluation.

4. Zheng et al. Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates. ICLR 2025.
   Limits and gaming risks in automated preference benchmarks.

Agents and long-horizon reliability

1. Yao et al. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. 2024.
   Tool-use, rule-following, and repeated-run reliability.

2. METR. Task-Completion Time Horizons of Frontier AI Models. Updated May 2026.
   Long-horizon task reliability on a defined software-oriented suite.

Inference latency

1. Chen et al. Cascade Speculative Drafting for Even Faster LLM Inference. NeurIPS 2024.
   Cascaded speculative drafting; speedup is setup-dependent.

2. Narasimhan et al. Faster Cascades via Speculative Decoding. ICLR 2025.
   Joint treatment of cascades and speculative decoding.

Monitoring terminology

1. NIST. Engineering Statistics Handbook — variables control charts and WECO rules.
   Definition of Western Electric rules and their false-alarm trade-off.