Knowin Research · Embodied intelligence

KnowinRouter

Expert–Evidence Alignment for Embodied Vision–Language Models

Choose the expert and the evidence together. Verify what comes back. Let the same expert answer once.

Read the paper ↗Code ↗Data Coming soon Model Coming soon

Knowin's first internal Mixture-of-LoRA (MoL) solution: independent capability adapters on a shared Qwen3-VL-8B backbone. An internal milestone, not an academic priority claim.

ONE BOUNDED ROUTEIllustrative

01 / JOINTLY SELECT EXPERT + EVIDENCE

GeneralSpatial ReasoningSpatiotemporal

KnowledgeGroundingDepthOCR

02 / ACQUIRE ONCE · VERIFY PACKETS

✓ Useful knowledgeAmbiguous grounding

One answer · same expertOriginal observation + retained evidence

If all evidence is rejected, keep the original input and the selected expert. No online answer judge. No retry loop.

61.4

Embodied QA mean

6 benchmarks · native metrics · Q1

74.7

General VLM mean

10 benchmarks · separate suite · Q4

+1.75

Expert selection alone

9 benchmarks · no external evidence · Q2

Paper-reported results under distinct protocols. This code release has not reproduced these measurements. Availability and evidence status ↓

01 / The method

Route capability.
Route evidence.

The same evidence can help one expert and distract another. KnowinRouter scores the complete expert–knowledge–tool tuple before making any external call.

U(x, e, a) = B(x, e) + G(x, a) + Γ(x, e, a) − λC(a)Evidence-free competence + shared evidence value + expert–evidence compatibility − acquisition cost

Paper framework: offline independent LoRA adaptation, evidence-null diagnostics and paired outcomes; online joint route proposal, one acquisition round, packet subset verification and one fixed-expert answer. — Method overview from the supplied paper. Offline outcomes supervise the Utility Router and Packet Verifier. Online verification uses packet attributes, not generated candidate answers.

01 / CAPABILITY SIGNATURES

Measured, overlapping strengths

General, Spatial Reasoning and Spatiotemporal experts share the frozen backbone. Evidence-free diagnostic outcomes describe capabilities without imposing a hierarchy.

02 / COMPLETE ROUTE UTILITY

One decision under a budget

Choose one expert, optional knowledge and any admissible tool subset. Explicit null actions preserve an evidence-free path for every expert.

03 / PACKET VERIFICATION

Revise evidence, keep the expert

Inspect schema, coverage, confidence, ambiguity and provenance. Compare valid subsets with null while holding the already-paid acquisition cost fixed.

02 / Reported evidence

Results with their context.

Cross-system comparisons, matched expert controls and the separate ablation protocol answer different questions. Their scores are not interchangeable.

Six benchmarks, six native metrics

Q1 · Higher is better. The manuscript combines reproduced, official API and externally reported scores; cross-system protocol identity is not established.

Paper-reported Q1 comparison: four plotted systems per benchmark across six benchmarks. All 24 labeled scores and 10 model names are transcribed in the following table. — Original comparison figure, with the project label updated. All plotted comparisons are transcribed below. The reported KnowinRouter equal-weight mean is 61.4.

Embodied QA (Q1) · all 10 models and 24 scores shown in Figure 3
Model	UniEQA	OpenEQA	VSI-Bench	ERQA	RoboVQA	Where2Place
KnowinRouter (8B)	71.6	58.1	64.1	46.5	63.0	65.0
Qwen3-VL-8B-Instruct	—	—	61.2	42.8	55.3	63.0
InternVL3.5-14B	70.6	—	60.5	—	—	—
Pelican1.0-VL-7B	70.3	57.2	—	—	62.6	—
Gemini-2.5-Pro	60.6	—	—	—	—	49.6
MiMo-Embodied-7B	—	—	—	—	—	63.6
OpenAI o3	—	—	57.8	41.5	52.2	—
Qwen-VL-Max	—	52.2	—	—	—	—
InternVL3-38B	—	—	—	41.7	—	—
GPT-4o	—	51.1	—	—	—	—

Paper-reported native scores; higher is better. — means the score is not shown in Figure 3, not zero. The manuscript provides four plotted systems per benchmark, not a complete model-by-benchmark matrix. No averages are computed from partial rows. Download transcribed values (JSON).

General vision–language performance

Q4 · Complete Table 1 from the supplied manuscript. All scores are higher-is-better. Scroll horizontally on small screens. Cross-system averages do not establish matched before/after retention.

Paper-reported general VLM results · full comparison, including all non-leading columns
Model	MMBench v1.1	MMStar	MMMU	AI2D	MathVista-mini	TextVQA	ChartQA	DocVQA	HallusionBench	POPE	Avg.
LLaVA-OV-7B	80.8	61.7	48.8	82.0	63.2	77.6	80.0	87.5	47.3	87.5	71.6
MiniCPM-V2.6	78.0	57.5	49.8	82.1	60.6	80.1	79.3	90.5	48.1	87.0	71.3
Cambrian-8B	75.9	57.6	42.7	79.7	49.0	71.7	73.3	77.8	41.6	86.5	65.6
InternVL2-8B	79.4	61.5	51.2	83.8	58.3	77.4	83.3	91.6	45.0	86.4	71.8
InternVL2.5-8B	82.2	62.8	56.0	84.5	64.4	78.2	84.7	92.9	49.3	88.5	74.4
Qwen2-VL-7B	80.5	60.7	54.1	83.0	58.2	84.3	83.0	93.9	42.4	87.3	72.7
Qwen3-VL-8B	74.9	42.9	47.0	46.7	32.8	83.1	83.5	95.8	51.3	88.3	64.6
SpatialBot-3B	58.2	39.1	31.4	56.7	35.0	50.2	48.1	56.3	36.2	81.5	49.3
SpatialRGPT-7B	72.3	52.8	43.5	74.6	49.1	70.5	68.4	78.0	41.0	85.7	63.6
ManipVQA-7B	69.5	50.2	40.3	71.5	46.8	67.0	65.3	75.2	38.8	83.7	60.8
LEO-7B	70.8	51.4	41.7	72.0	47.3	68.8	66.7	76.5	39.5	84.9	62.0
KnowinRouter (8B)	81.9	63.2	57.8	84.3	61.5	83.9	84.2	93.8	47.9	88.3	74.7

The gains are selective. Under the matched control, the General expert remains stronger on UniEQA (−3.37) and OpenEQA (−2.43). KnowinRouter also does not lead every column in the general VLM table. These results motivate retaining General as an explicit route.

03 / Complementarity & ablation

Useful together.
Useful selectively.

Expert choice accounts for the largest component effect. Knowledge and perception supply different information, and can also introduce conflicting or ambiguous evidence.

Ablation scores: full 60.2; without router 54.1; random 52.4; without knowledge 57.9; without perception 57.4; without both 55.8; heuristic labels 58.3; joint end-to-end 57.1; joint expert update 56.2. — Q3 · Separate frozen protocol, full-system reference 60.2. Editable redraw from the paper; plot source.

Evidence has an expert-dependent value.

From the evidence-free variant (55.8), knowledge alone recovers 1.6 points and perception alone 2.1. Together they recover 4.4: a paper-reported complementarity of 0.7 points under this ablation protocol.

−6.1without the router

+0.7joint evidence interaction

Routing also helps without tools.

In the separate nine-benchmark evidence-null control, question-conditioned expert selection reaches 60.24 versus 58.49 for the best fixed expert. Reported gain: +1.75, with media-grouped 95% CI [1.25, 2.26].

Inspect original complementarity and sensitivity figures

Original expert complementarity figure: capability ranks reverse by benchmark and remain distinct across candidate training checkpoints. — Paper Figure 4. Pointing/Grounding is a diagnostic candidate, not a fourth deployed expert; localization becomes a shared tool action.

Original ablation and sensitivity figure, including LoRA rank, expert count, answer judge and label coverage sweeps. — Paper Figure 5. The LoRA-rank sensitivity sweep is distinct from the main specialist training rank of 128.

04 / A bounded route, step by step

What reaches the expert?

Illustrative · not a model prediction
The paper's mug example explains the decision sequence. This interactive trace uses synthetic packets; it makes no claim about a measured answer.

Illustrative mug-placement example from the paper showing capability mismatch, missing evidence, uncertain packet utility, and a joint route with verification. — Conceptual example from the supplied paper. Scene annotations and flow labels are illustrations, not newly generated model outputs.

Select a complete route

The router jointly selects the Spatial Reasoning expert, a reference-frame template and grounding. The finite budget is checked before any provider runs.

The expert stays fixed after this step

mode: illustrative / synthetic
expert: Spatial Reasoning
knowledge: reference-frame template
tools: [grounding]
answer_generations: 0

The actual CPU demo follows the same bounded API with explicit mock components. Run the demo and tests ↗

05 / Authors & institutions

The team behind KnowinRouter.

Weilin Ruan^1,2Yingjie Xu^2,3Yangkai Wei²Bohan Zhou^1,2Shuaibo Li¹Harold Haodong Chen^3,4Kaiwen Zhou²Yinchuan Li²Ying-Cong Chen³James Cheng^1,*

¹ The Chinese University of Hong Kong · ² Knowin.AI

³ The Hong Kong University of Science and Technology (Guangzhou)

⁴ The Hong Kong University of Science and Technology

The Hong Kong University of Science and Technology (Guangzhou)

The Hong Kong University of Science and Technology

* corresponding author · rwlinno@gmail.com

06 / Open resources

Read it. Inspect it. Build on it.

Code, tests and a weight-free CPU demo are released. Original research predictions, logs, data and weights remain unavailable.

Available now

Public source, install/configuration paths, unit tests, training and evaluation entry points, the project preprint and a minimal LaTeX upload package.

PDF ↗LaTeX source ↓Live CI status ↗

Coming soon

HF model and data repositories are documentation-only placeholders. There are no downloadable weights or datasets and no promised release date. View the collection ↗

Original PDF figures are reused with attribution to the supplied manuscript. Editable sources are available for the newly drawn numerical charts and project logo. The original figure-authoring files were not supplied. The source package is prepared for arXiv/Overleaf; no arXiv identifier is claimed.

Cite the project preprint

@misc{ruan2026ektrouter,
  title={EKT-Router: Expert--Evidence Alignment
    for Embodied Vision--Language Models},
  author={Ruan, Weilin and Xu, Yingjie and
    Wei, Yangkai and Zhou, Bohan and Li, Shuaibo
    and Chen, Harold Haodong and Zhou, Kaiwen
    and Li, Yinchuan and Chen, Ying-Cong and Cheng, James},
  year={2026},
  howpublished={Project preprint},
  url={https://knowin-29.github.io/KnowinRouter.github.io/}
}