ARC 1.0
A decision model for routing AI agents across cloud and local models.
ARC—the Adaptive Routing Controller—reads an agent’s conversation and produces signals for choosing a model and its reasoning effort. Applications combine those signals with their available models, prices, and policies to decide how to handle the next request.
Technical report · Documentation · Raydock · Rayline
What ARC does
ARC separates predictions from policy. It predicts success and token usage; the application supplies model prices and decides how to balance quality, cost, and local execution. Updating prices does not require retraining those predictors when the model and serving configuration stay fixed.
| Signal | What it enables |
|---|---|
| Model success | Compare eligible models by their predicted probability of completing a task. |
| Token usage | Estimate uncached input, cache reads, cache writes, and output; apply the prices relevant to your deployment. |
| Reasoning quality and usage | Compare the expected benefit and cost of supported reasoning settings within a model. |
| PII detection | Identify the presence and types of sensitive information in the supplied input. |
| Request categorization | Classify requested operations and domains, allowing multiple labels. |
| Developer-defined signals | Supply a question and answer options to the general-purpose question head and obtain probabilities for the answers. |
The question head is trained primarily for routing decisions, alongside broader decision tasks. For example: Does this request need a stronger model? Which of these models should answer? Will the proposed model succeed?
Choose a policy
Sage and Swift are preset configurations of the same routing model. Their utility policy chooses the eligible model with the highest predicted success value minus expected cost:
Utility = V × predicted success − expected cost
| Configuration | Behavior |
|---|---|
| ARC Sage · V = 4.5 | Places greater value on successful completion. |
| ARC Swift · V = 0.1 | Places greater weight on reducing cost. |
| Prefer-local | Uses an eligible local model when it clears a chosen predicted-success threshold; otherwise applies the configured fallback policy. |
You can supply your own value of success, model prices, eligibility rules, and local threshold. A preset expresses a preference; realized quality and cost depend on the requests, model pool, and execution setup.
Evaluation
The following results reproduce the evaluations in the technical report. They cover the paper’s routing configurations and head checkpoints.
Hybrid edge–cloud summary
Across six benchmarks, the paper’s hybrid configuration retains 97.5% of Opus 5.5’s quality at 92.9% lower cost. This uses DeepSeek V4.1 Flash and a multiple-DGX-Spark deployment setup, with a 0.85 prefer-local threshold. Quality averages per-benchmark ratios; cost pools spending across matched tasks.
Cost optimization against single models
The upper bars show task success; the lower bars show cost per task relative to the paired baseline (100%). Blue denotes ARC; gray denotes the single model.
ARC Sage vs Opus 5.5
Exact values
| Benchmark | Opus 5.5 success | ARC Sage success | ARC cost / baseline |
|---|---|---|---|
| DeepSWE | 69.9% | 68.7% | 7% |
| Terminal-Bench 2.1 | 91.0% | 92.4% | 58% |
| SWE-Atlas QA | 61.6% | 62.4% | 13% |
| MMMU-Pro Vision | 82.2% | 82.9% | 39% |
| GPQA Diamond | 89.9% | 94.1% | 21% |
| Tau3 Banking | 42.8% | 33.2% | 17% |
Comparison with other routers
RouteBench’s Model Routing Index combines quality, speed, and cost on a 0–10 scale; higher is better. The nominal weights are 60%, 20%, and 20%. Mean requires all five benchmark scores. Missing results remain unavailable (—).
Exact values
| Router | DeepSWE | SWE-Atlas QA | Terminal-Bench 2.1 | Tau3 Banking | MMMU-Pro Vision | Mean |
|---|---|---|---|---|---|---|
| ARC Sage | 9.6 | 9.1 | 10.0 | 7.7 | 10.0 | 9.3 |
| ARC Swift | 9.7 | 8.5 | 10.0 | 7.5 | 10.0 | 9.1 |
| OpenRouter Auto | 8.8 | 8.6 | — | 9.0 | 8.8 | — |
| OpenRouter Auto Code | 8.5 | 8.7 | — | 8.9 | 9.7 | — |
| OpenRouter Jev (medium) | 9.0 | 8.2 | — | 7.4 | 9.4 | — |
| Pareto | 9.9 | 9.1 | 10.0 | 8.0 | 8.9 | 9.2 |
| Fugu Max | 4.7 | 7.0 | 8.2 | 6.5 | 9.0 | 7.1 |
| Fugu Ultra v2 | — | 7.6 | — | 6.3 | 8.1 | — |
Opus vs ARC prefer-local
Prefer-local uses a predicted-success threshold of 0.85 and V = 4.5 for fallback. The deployment setup uses DeepSeek V4.1 Flash across multiple NVIDIA DGX Spark systems.
Exact values
| Benchmark | Opus 5.5 success | ARC prefer-local success | ARC cost / Opus |
|---|---|---|---|
| DeepSWE | 69.9% | 72.3% | 1.1% |
| Terminal-Bench 2.1 | 90.7% | 90.5% | 75.7% |
| SWE-Atlas QA | 61.6% | 62.6% | 1.6% |
| Tau3 Banking | 42.8% | 35.3% | 17.3% |
| GPQA Diamond | 89.5% | 92.5% | 12.3% |
| MMMU-Pro Vision | 82.2% | 82.7% | 40.9% |
PII detection and request categorization
Evaluation on an internal dataset with users separated across training, calibration, and test partitions. Labels use weak supervision with probabilistic models, rules, and human review. PII uses ROC-AUC; categorization uses example-averaged F1. The release includes all three specialized heads. The lighter blue bars use the full input; the darker blue bars use the latest message.
Exact values
| Signal | Metric | Full input | Latest message |
|---|---|---|---|
| PII presence | ROC-AUC | 0.919 | 0.960 |
| Operation | F1 | 0.775 | 0.787 |
| Domain | F1 | 0.834 | 0.820 |
Model-routing questions
Our internal Router Question Bench contains 3,056 questions curated from agent tasks with counterfactual trajectories across models. Agreement is the percentage of questions for which the model’s highest-probability answer matches the target’s highest-probability answer.
Agreement values by question family
| Question family | ARC | Jev 1.14 |
|---|---|---|
| Cheapest adequate tier | 59.8% | 10.7% |
| Tier with cost preference | 55.9% | 22.5% |
| Task difficulty | 49.7% | 23.4% |
| Model at stated value | 45.5% | 23.2% |
| Need for a strong model | 80.0% | 62.9% |
| Success score | 47.1% | 32.3% |
| Will this model succeed? | 71.2% | 58.4% |
| Which model should answer? | 40.4% | 28.4% |
| Overall | 56.2% | 32.7% |
Additional Router Question Bench metrics
| Metric | ARC | Jev 1.14 |
|---|---|---|
| Calibration error ↓ | 0.079 | 0.31 |
| Selected-model task success ↑ | 71.6% | 65.4% |
| Cost per attempt (USD) ↓ | 0.77 | 1.19 |
| Option-order flips ↓ | 19.0% | 9.3% |
General decision capabilities
Decision Index 0.3 reports chance-corrected skill (0–100; higher is better). Overall uses the suite’s category weights.
Exact values
| Category | ARC | Jev 1.14 | StartLux 0.8B |
|---|---|---|---|
| Knowledge | 24.6 | 57.1 | 28.5 |
| Language | 49.8 | 61.2 | 55.5 |
| Retrieval | 47.1 | 63.5 | 66.5 |
| Tools | 72.0 | 73.8 | 74.6 |
| Arts | 51.0 | 39.2 | 51.1 |
| Overall | 46.9 | 60.7 | 53.8 |
Distribution of model usage
Model-selection shares, averaged with equal weight across the paper’s six benchmarks. These are selection shares, not token or spending shares. DeepSeek is local in the prefer-local column; the other selections are cloud models.
Exact model shares
| Model | ARC Sage | ARC Swift | ARC prefer-local |
|---|---|---|---|
| Opus 5.5 | 11.9% | 0.0% | 12.3% |
| Sonnet 5.5 | 4.9% | 0.0% | 3.6% |
| Haiku 5.5 | 0.4% | 25.1% | 0.2% |
| GPT-6.1 Sol | 46.8% | 24.9% | 35.7% |
| GPT-6 Luna | 0.0% | 10.9% | 0.0% |
| GPT-6 Astra | 0.4% | 0.2% | 1.3% |
| DeepSeek Flash | 34.8% | 38.4% | 46.8% |
| GLM-5.3 Flash | 0.7% | 0.5% | 0.1% |
Use ARC
Connect an agent
Use the Rayline documentation to connect an existing agent to the hosted ARC service. The hosted model IDs are arc-sage-1.0 and arc-swift-1.0.
Raydock is an open-source edge–cloud gateway for connecting agent tools to local models and cloud services. Rayline’s vLLM Semantic Router fork provides the cloud routing integration.
Work with the weights
The release includes routing, reasoning, PII, request-classification, and question-head weights. It uses the custom ArcImagePolicyForPooling architecture, built on Qwen3.5-0.8B.
The public vllm-arc plugin is a work in progress. It provides encoder integration and Stage-1 scoring. End-to-end serving, the decision/policy API, and caller-facing question and content-signal APIs are not yet supported. Stock vLLM does not load this architecture by itself. Follow the plugin’s installation instructions for the supported vLLM fork; use the hosted API for an operational integration today. Join our Discord to collaborate on maturing our open-source serving pathways!
Prices are application inputs. The package contains no price snapshot. Predictions of success and token usage can be scored without prices; a routing policy combines those predictions with your rates. Preset values are recorded in presets.json.
Architecture and release contents
ARC shares a Qwen3.5-0.8B backbone across its signals. Specialized heads read the frozen lower twelve layers. The general-purpose question head uses all twenty-four layers, with the upper twelve fine-tuned for decision questions.
| Component | Included in the release |
|---|---|
| Encoder and routing heads | config.json and model.safetensors, including model-success, token-usage, reasoning, PII, and request-classification weights. |
| Question head | Fine-tuned upper layers, question readout, and inference manifest in question_head/. |
| Input preparation | Tokenizer, processor, and chat-template files. |
| Policies | Preset values in presets.json; prices are supplied at runtime. |
| Validation | goldens.jsonl and goldens.signals.jsonl are synthetic test fixtures with expected routing and signal outputs. They contain no user or production traffic. |
| Release records | Component hashes and provenance, plus LICENSE and NOTICE. |
Training
ARC learns from agent trajectories, tasks with verifiable outcomes, and human comparisons. Weak supervision combines task outcomes, probabilistic models, rules, and human review. The routing predictors learn success and usage separately; the question head learns structured assessments from questions and answer options. The curated training corpus and production training system are proprietary.
The technical report describes the architecture, training approach, cost-optimization evaluations, developer-defined signals, and hybrid edge–cloud routing. Its comparisons depend on the stated model pool, policy, prices, and benchmark population.
License and contact
ARC 1.0 is released under Apache 2.0. See LICENSE and NOTICE for the release terms and attribution, including Qwen3.5-0.8B and Nemotron-PII (CC-BY-4.0).
Built by the Rayline Team. Questions and integration feedback: team@rayline.ai.
- Downloads last month
- -