Han-Jia YeLAMDA · Nanjing University

Research overview

Model Reuse

Reuse, collaboration, and evolution of large and small models.

We study how to organize existing large, small, and specialist models, tools, and agents into useful systems. A new task may call for a suitable model, a focused adaptation, or several capabilities working together. Our goal is to learn how to make these choices and improve the system as tasks and environments change.

Explore the big picture Papers & code

From resources to a working system

Understand the task.
Organize the capabilities.

A new taskClassify customer feedback.

Quality · cost · reliability

Existing resources

Large modelSmall modelSpecialistToolAgent

Use an appropriate capability

Small modelClassifyCheck outputReturn the result

A focused task may need one suitable model. The useful choice depends on evidence about its capability.

Conceptual examples of the research vision. No models are run here.

Learn how tasks and capabilities fit together, then reuse that knowledge to select, route, learn, and reason.

The research programme

Reuse, collaboration & evolution
of large and small models.

Existing models carry knowledge. Learning to use that knowledge extends the problem from designing a model to organizing a system of models, tools, and agents.

Large models offer broad capabilities, while smaller and specialized models can offer lower cost or a better fit to a particular task. A useful system should understand the task, judge which capabilities it needs, and decide when to use one model or coordinate several resources.

Our research connects four questions. They form an iterative loop: feedback from execution can improve later selection, adaptation, and organization as tasks and environments change.

A connected research agenda

A system learns how to use what it already has.

Task requirementsModels · tools · agentsAn evolving resource library
Select useful capabilities

Describe tasks and model behavior, then choose suitable resources under quality, cost, and reliability requirements.

Feedback closes the loop. Execution produces evidence that can inform later selection and updates to the resource library.

Activate and revisit these facets as the task requires. The cycle outlines a research agenda for learning model systems.

A foundation for this direction · IJCAI 2025 survey

A Unifying Perspective on Model Reuse:
From Small to Large Pre-Trained Models

The survey connects selection, adaptation, and model representations across specialized and general pretrained models. It provides an entry point to the wider model-reuse landscape.

Read the survey Explore the collection

What if the objects we represent are tasks and models?

Represent tasks and models. Learn how they fit.

A task has requirements; a model has capabilities. We can learn representations of both, then predict how well they fit before fully evaluating every candidate. These meta-representations extend representation learning from individual examples to entire tasks and pretrained models.

RepresentationAn individual example

Encode its features.

Meta-representationA task or a model

Encode its requirements or capabilities.

Model Spider · the idea

Learn where a model can help.

Historical evidence

A

Model ATask outcomes

B

Model BTask outcomes

C

Model CTask outcomes

A selector learns from model–task interactions.

Model and task tokens

Schematic model and task representationsSwitch between two image tasks. The highlighted candidate changes with the learned suitability ranking. Point positions are illustrative; geometric distance is not the scoring rule.ABCNew taskTokens describe models and tasks

01 · Score

Learned suitability

A learned scorer compares each model token with the task token.

02 · Rank

  1. A
  2. B
  3. C

Prioritize promising candidates.

03 · Refine

Shortlist A + B

Enrich the task token with candidate-specific evidence, then re-rank.

Learn a ranking, then refine the shortlist.

For image task A, this example ranks A first and shortlists A and B. Model Spider learns suitability from model–task evidence, then uses candidate-specific information to refine the ranking.

Illustrative tokens, rankings, and shortlist size. Model Spider learns a suitability score; the distances drawn here are not its scoring rule.

NeurIPS 2023

Model Spider

A meta-representation for model selection

Task and model tokens capture information beyond individual samples. A learned suitability function scores candidate models for a new task, then candidate-specific evidence refines the shortlist. The shared-space view makes model selection a problem that can itself be learned.

Represent the task and the available models, then learn which capabilities fit.

Full paper title

Model Spider: Learning to Rank Pre-Trained Models Efficiently

Paper Code

The same question arises across modalities: what evidence describes the task, and which pretrained capability is suitable?

What changes when the candidate models change?

Extend selection across modalities.

The representation-and-matching view extends to different forms of evidence: class names, language instructions, personalized style requests, or a time-series history. These examples broaden the selection problem while keeping its central question intact.

Vision–language

SWAB

NeurIPS 2024

Use target class names to select a vision-language model when target images are unavailable, while addressing modality and capability gaps.

Full paper title

Bridge the Modality and Capability Gaps in Vision-Language Model Selection

Paper Code

Language models

Capability Instruction Tuning

AAAI 2025

Combine a model’s capability representation with the user’s instruction to predict which LLM is suitable before executing the candidates.

Full paper title

Capability Instruction Tuning

Paper

Image generation

Polaris

arXiv 2026

Retrieve and integrate relevant generation checkpoints and adapters for personalized style instructions, without new training for each request.

Full paper title

Polaris: Scaling Up Instruction-Guided Image Generation Towards Millions of Personalized Style Needs

Paper

Time-series foundation models

One-Embedding-Fits-All / ZooCast

arXiv 2025

Represent forecasting tasks and candidate models in a shared embedding space to select a useful forecaster efficiently in a zero-shot setting.

Full paper title

One-Embedding-Fits-All: Efficient Zero-Shot Time Series Forecasting by a Model Zoo

Paper

At the level of individual requests, selection becomes routing. Execution cost makes quality–cost trade-offs central.

From model selection to LLM routing

Ready-to-use models still need the right selection.

Available foundation modelsChoose for each request

When existing models can already perform a task through prompts or context, the system can select among them for each request. This raises two questions: how to make a reliable routing decision, and how to make the router itself easy to deploy.

Which model is sufficient, and how do we know?

Route for quality, cost, and reliable decisions.

Routing chooses a model for each request, aiming to retain strong performance with less execution cost. Two linked questions determine whether it works: does the learning objective lead to the right quality–cost decision, and does the supervision describe capability reliably?

NeurIPS 2026

When LLM Routers Overpay

Problem 1 · align the objective with the decision

A larger feasible budget can lead routers to over-select strong, expensive models. The work studies how small score errors near a decision boundary can change model rankings, motivating decision-aware routing.

Permission to spend more does not establish a need for more capability.

Full paper title

When LLM Routers Overpay: Strong-Model Over-Selection under Loose Budgets

Conference paper
A looser budget can produce unnecessary strong-model calls.

On RouterBench, the learned routers increasingly call GPT-4 as the budget grows, while the oracle uses it for a much smaller share of requests. The gap reveals strong-model over-selection.

Figure 1, earlier arXiv preprint

EMNLP 2026

DARS

Problem 2 · make capability evidence reliable

One sampled answer is a noisy observation of capability. DARS uses repeated decoding and meaning-preserving query rewrites to construct supervision that reflects expected performance, cost, and variability.

Better supervision improves the evidence behind a routing policy.

Full paper title

From Sampled Outcomes to Capability Distributions: Rethinking Supervision for LLM Routing

Paper Code Data
A single sample can produce an unstable routing label.

Across GPQA, MATH-500, and DROP-800, sampled outcomes and the selected winner can vary. The chart also measures disagreement with distribution-aware (DA) labels under the paper's diagnostic protocol.

DARS, Figure 2(a) Full diagnostic figure

DARS · from an answer to capability evidence

One answer is a sample. Capability is a distribution.

Follow how repeated observations of the same candidate produce richer supervision for a router.

Keep the intent.

Original requestReturn the invoice total.

Rewrite ACompute the invoice total.

Rewrite BFind the invoice's total.

The wording changes; the question stays the same.

Observe the model.

MSame candidate

Repeat decoding for the original and rewritten requests.

Repeat the process for each candidate.

Record behavior.

Illustrative response records

Response assessmentUsage record

One outcome reveals only part of the picture.

Aggregate the observationsDescribe capability, cost, and uncertainty.

Expected quality

How well does the model answer?

Expected cost

What does an attempt consume?

Variability

How much can outcomes change?

Distribution-aware routing supervision

Use richer capability evidence when learning which model to choose.

Train the router on a broader view of capability.

Quality, cost, and variability from repeated observations provide richer supervision than treating one sampled response as the model's capability.

Conceptual example. Requests and response symbols illustrate the procedure; they do not show experimental samples or measured quality and cost.

Even a good routing policy can be costly to rebuild. Can the router itself become a capability we simply reuse?

Can routing be ready to use in a new environment?

Make the router ready to use.

A foundation model brings a useful capability that can be applied directly to a new task. Yet many routing methods still require a router to be trained for the current workload and candidate models. We address this deployment barrier in two complementary ways: RouteFM makes the routing capability itself ready to use after pretraining; SaveRouter makes a router quick to construct with little feedback.

Lower the barrier to deployment

Reuse a pretrained router. Or train one with less feedback.

Conventional routing

Fit for the environment

New environmentA workload and model pool
Collect & trainFit a task-specific router
Start routingRefit when conditions require it

RouteFM

A ready-to-use routing model

Pretrain onceLearn across routing tasks
Provide contextBehavior of current candidates
Route directlyRouter parameters stay fixed

SaveRouter

Quick setup with little feedback

Sparse feedbackInformative query–model outcomes
Train the routerUse the limited supervision
Start routingAccount for setup and serving cost

RouteFM uses behavioral context, rather than updating its parameters at deployment. SaveRouter still trains a router, but reduces the feedback that must be acquired. These are schematic workflows, not training-time measurements.

RouteFM · the idea

New candidates. The same router.

Input · new request

Summarize customer feedback.

Input · current pool & behavior

Each candidate is described by its observed behavior.

Frozen
RouteFMSame parameters θ
Infer suitability
for this request

Output · routing choice

ModelB

Selected from
A · B · C

Illustrative choice

Change the request

Change the candidate pool

The request and the evidence determine the choice.

This example selects B to summarize text from the A–B–C pool. Change the request or pool to explore different illustrative choices; RouteFM's parameters stay fixed.

Fictional candidates, evidence, and choices illustrate the data flow; no model is run here. RouteFM's paper evaluates transfer across specified tasks, model pools, and context budgets.

arXiv 2026

RouteFM

Pretrain once. Route without retraining.

RouteFM is a foundation model for routing. Episodic pretraining teaches it to infer model capabilities from behavioral evidence. At a new deployment, the pretrained router uses context about the current candidates to handle new tasks and model pools, with its parameters kept fixed in the evaluated transfer settings.

Provide behavioral context and route with frozen parameters.

Full paper title

Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing

Paper Project Code Models

arXiv 2026

SaveRouter

Set up a router quickly with little feedback

SaveRouter constructs a router from a small set of informative query–model outcomes. It selectively acquires feedback and shares evidence across related queries, reducing the supervision needed before deployment. It also accounts for that setup expense together with later serving savings.

Use sparse feedback to shorten setup, then check when serving savings repay that cost.

Full paper title

Routing Should Pay for Itself: Sparse Supervision for Economical LLM Routing

Paper Project Code

Capability estimates can guide the next learning step as well as the next model call.

Can capability-aware allocation improve the model itself?

Use routing knowledge to guide learning and reasoning.

Routing is also a way to allocate useful computation and information inside a learning system. V₀ and V₀.₅ use transferable capability estimates to support reinforcement learning. DLMR applies a related allocation idea during vision-language reasoning: decide which memory to retrieve and how much evidence to inject.

Value models for reinforcement learning

V₀ / V₀.₅

Estimate a value prior, then use evidence where it matters

A shared idea connects these studies: reuse a generalist estimate of a policy’s capability to allocate learning computation. V₀ predicts success before rollouts; V₀.₅ combines a value prior with observed rollouts to estimate a learning baseline.

Policy history + a new prompt

A reusable value prior

Estimate expected performance before generation.

V₀

ICML 2026

Guide rollout allocation

Predict initial-prompt success from policy history, then use the estimate to guide rollout budgets.

Full paper title

V₀: A Generalist Value Model for Any Policy at State Zero

Conference paperarXivProjectCode

V₀.₅

COLM 2026

Build a reliable learning baseline

Combine the prior with sparse observed rollouts. Reliability checks support an advantage baseline and guide additional sampling.

Full paper title

V₀.₅: Generalist Value Model as a Prior for Sparse RL Rollouts

PaperCode

Use prior capability knowledge to guide computation, and calibrate it with fresh evidence.

Each branch studies a different learning decision, supported by its own paper and evaluation.

ICML 2026 · Spotlight

DLMR

Route memory inside the reasoning process

Separate visual and reasoning memories supply complementary evidence. A learned router and lightweight injector decide which memory and how much information to reuse, while the base multimodal model remains frozen. Here, routing controls internal evidence for reasoning.

The allocation principle extends from choosing an external model to choosing useful information inside one.

Full paper title

Dual-Latent Memory Routing for Vision-Language Reasoning

Conference paper Code

A wider model system must decide not only what to use, but when to invoke it and how to improve its resources.

What should a resource-aware model system learn next?

From selection to coordinated, evolving systems.

The longer-term goal connects selection with scheduling, adaptation, and evolution. Existing benchmarks provide concrete starting points; reliable autonomous improvement remains a research question.

Future work

Select and schedule resources

Go beyond choosing a capable model or tool: decide when calls should run and which can safely run in parallel under resource constraints. PeakBench separately evaluates dependency planning and physical scheduling, providing a starting point for this wider allocation problem.

How should resource requirements change both the selected tool and its invocation schedule?

PeakBench · paper

Future work

Adapt, coordinate, and evolve

Extend the resource pool to models, tools, memories, and agents. Combine selection with targeted adaptation and execution feedback, and investigate recursive self-improvement (RSI) while retaining useful capabilities.

What evidence should trigger an update, and how can we verify that the whole system improves?

Shared tools, evaluation platforms, and benchmarks make these questions easier to investigate reproducibly.

Continue with the work

Read it. Run it. Build on it.

Choose an entry point for your own research or application.

Toolbox · FCS 2026

ORBIT

A common pipeline for optimal routing and budgeted inference, supporting reproducible implementation and comparison.

Paper & code

Evaluation platform

RouteJudge

Compare routing decisions reproducibly under shared candidates and budgets, including user preferences that a single quality score can miss.

Platform Paper

Benchmark · ECCV 2026

MMR-Bench

Evaluate routing across multimodal requests and model pools, connecting capability, quality, and execution cost.

Code & benchmark Paper

Further reading & collections

A Unifying Perspective on Model Reuse · IJCAI 2025 Model Reuse Tutorial · AAAI 2026 Awesome Model Reuse ZhiJian model reuse toolbox

This content was generated with a large language model and reviewed by a human.

Method figure

Enlarged method figure