SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing
Summary
SeLMRoute introduces an LLM routing framework that separates candidate-independent semantic evidence extraction from performance learning, achieving 72.08% average accuracy on LLMRouterBench across 15 datasets and 20 candidate models, outperforming the strongest fixed candidate (69.23%) and enabling both performance-oriented and cost-aware routing decisions.
View Cached Full Text
Cached at: 10/01/26, 04:23 PM
Paper page - SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing
Source: https://huggingface.co/papers/2609.34736
Abstract
Largelanguagemodel(LLM)routingaimstoselectthemostsuitablemodelforeachincomingquery.Mostexistingrouterslearnthisdecisiondirectlyfromqueryembeddings,modelrepresentations,preferencedata,orclustersofsimilarexamples.Suchapproachescanbeeffective,yettherepresentationusedforroutingrarelystateswhataqueryactuallyrequires.WeintroduceSeLMRoute,aroutingframeworkthatseparatestheextractionofcandidate-independentsemanticevidencefromthelearningofcandidateperformanceandtheapplicationofdeploymentobjectives.Adecisionmodelfirstevaluatesasetofinterpretablequestionsaboutthequery,suchasitsreasoningrequirementsanduseofexternalknowledge,witheachjudgmentretainedasaprobabilitydistribution.Theresultingprobabilisticsemanticstateisusedbyalightweightsupervisedroutertoestimatecandidatemodelperformance.Routingobjectivesareappliedafterperformanceestimation,whichallowsthesamesemanticstatetosupportperformance-orientedandcost-awaredecisions.OntheLLMRouterBench(15datasets,20candidatemodels,11,481queries),SeLMRouteachievesanaverageaccuracyof72.08%pm0.45,whilegroupedfive-foldout-of-foldevaluationreaches72.64%,comparedwith69.23%forthestrongestfixedcandidate.Therepresentationachievesthehighestmeanperformanceamongtheevaluatedsemantic,dense,lexical,anddomain-levelrepresentations.Inaseparate13-modelperformance-costsetting,SeLMRouteimprovesperformanceinallfivegroupedsplits,withameanPerfGainof2.66%.Ourcodeisavailableathttps://github.com/Indigma-Innovations/SeLMRoute.
View arXiv pageView PDFGitHub0Add to collection
Get this paper in your agent:
hf papers read 2609\.34736
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.34736 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.34736 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.34736 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
SLMs as Multi-Agent Routers: A Progressive SFT and Reinforcement Learning Approach
This paper proposes training small language models as multi-agent routers via progressive supervised fine-tuning and reinforcement learning, achieving better retrieval relevance and lower latency compared to LLM baselines that route on intent alone.
LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers
LLMRouter presents a unified formulation of LLM routing as a sequential decision process, along with an open-source infrastructure and benchmark (xRouteBench) for developing, evaluating, and deploying LLM routers. Empirical results show learned routers achieve 14.6% relative improvement over the strongest fixed-model baseline.
Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing
The paper introduces RouteFM, a foundation model for LLM routing that learns reusable routing capabilities via episodic pretraining across heterogeneous environments, allowing a frozen router to adapt to new domains, modalities, and candidate pools through behavioral context alone. RouteFM outperforms the strongest baseline by 2.23 quality points on MMR-Bench with only eight observations per candidate, supporting a 'pretrain once, route anywhere' paradigm.
SEAR: Segment-Evidence-Aware Routing for Weak-to-Strong Multilingual Speech MCQ
This paper presents the SEAR system for multilingual speech MCQ, achieving 90.92% accuracy on the MLC-SLM Challenge through segment-evidence-aware routing, supervised fine-tuning, and reinforcement learning.
Backbone-Adaptive Evidence Routing for Robust Pairwise LLM Judging
BAER introduces a backbone-adaptive evidence routing method for robust pairwise LLM judging, achieving higher accuracy across benchmarks by dynamically selecting evidence mechanisms per condition.