xRouteBench — LLM Routing Benchmark

xRouteBench is a benchmark for training and evaluating LLM routers — systems that pick the best LLM from a candidate pool for each incoming query, trading off performance vs. price cost.

Every query in each scenario was executed against all 18 candidate LLMs, recording each model's response, task performance, token usage, and latency. A router learns from the train split which model to pick, and is evaluated on test.

Scenarios (configs)

Config	Domain	Train rows / queries	Test rows / queries	Metric
llmrouter_generic	13 classic NLP benchmarks (MMLU, GSM8K, MATH, MBPP, ARC, …)	80,802 / 4,487	67,122 / 3,729	em_mc, GSM8K, MATH, code_eval, f1
memory_locomo	Long-conversation memory QA (RAG top-k=5)	15,930 / 885	5,652 / 314	f1
memory_longmemeval	Long-term memory eval (RAG top-k=5)	4,986 / 277	1,818 / 101	f1
timeseries	Time-series understanding (7 sub-tasks)	17,568 / 976	2,286 / 127	mc
video	Egocentric video QA (Charades-Ego)	3,618 / 201	486 / 27	em
multimodal_geometry3k	Geometry math (multimodal)	8,640 / 480	1,098 / 61	em
multimodal_mathvista	Visual math reasoning	14,400 / 800	1,800 / 100	em, em_mc
personalized	Personalized preference (LLM-judge; chat-format queries)	2,464 / 2,235	308 / 303	llm_judge
llm_candidates	The 18-model candidate pool with pricing	18 models	—	—

Schema (routing data)

Each row = one (query, candidate model) pair:

Field	Type	Description
task_name	str	Sub-task the query belongs to (e.g. gsm8k, mbpp)
query	str	Full input prompt (personalized: JSON-encoded chat messages)
ground_truth	list/str	Reference answer(s)
metric	str	Scoring metric for this row
choices	str	Options for multiple-choice items (JSON-encoded)
task_id	str	ID within the sub-task
model_name	str	Candidate LLM this row was executed with
response	str	The model's actual response
token_num	int	Total tokens
input_tokens / output_tokens	int	Token breakdown (for price computation)
response_time	float	Latency in seconds
performance	float	Task score of this model on this query (0–1)
embedding_id	int	Index into precomputed query-embedding files (not included here)

Raw query configs (*_queries)

Each scenario also ships its raw query set (no model executions) as a <scenario>_queries config with train / valid / test splits — use these to run your own candidate models from scratch. Fields: task_name, query, ground_truth, metric, choices, task_id (+ conversation_id/category for the memory scenarios). The memory queries are the RAG top-k=5 turn-pair variant used in the published experiments. Note: the valid split exists only here; the routing-data configs have train/test.

Candidate pool & pricing (llm_candidates config)

18 models spanning $0.05–$1.25 per 1M input tokens (25× spread) served via Together AI / NVIDIA NIM. Row cost = input_tokens × input_price/1e6 + output_tokens × output_price/1e6.

Usage

from datasets import load_dataset

ds = load_dataset("ulab-ai/xRouteBench", "llmrouter_generic")   # any config name above
train, test = ds["train"], ds["test"]

pricing = load_dataset("ulab-ai/xRouteBench", "llm_candidates")["train"]

Composite reward for cost-aware routing (GraphRouter-style):

reward = α · norm(performance) − β · norm(price_cost)

Notes

• Query embeddings (.pt) are not included; they can be regenerated from the query field with any sentence encoder.
• The memory scenarios use the RAG top-k=5 turn-pair context variant.