4B Router Beats Opus 5.5: Route Anything Framework Dynamically Selects Models Across Modalities
Researchers from NUS, UCB, and HKUST propose Route Anything, a unified routing framework where a 4B parameter model selects among diverse candidates (LLMs, agents, robot policies) per input, outperforming fixed strong models like Opus 5.5 on Terminal-Bench (77.78 vs 69.44) and cutting API costs by 96% with minimal quality loss.
Researchers from the National University of Singapore, InfRec Cardinal AI Lab, UC Berkeley, and Hong Kong University of Science and Technology introduce Route Anything , a research framework that places generative models and robot policies into a shared quality‑cost space so the system can pick the most suitable candidate for each input.
Core Question: Can Model Complementarity Surpass the Strongest Single Model?
Traditional routers aim to reduce cost by using cheaper models to approximate strong ones. This work asks whether combining models can actually exceed the best single model. On the AAII benchmark, a 4B‑parameter router achieved 77.78 on 36 shared Terminal‑Bench inputs, while a fixed Opus 5.5 call scored 69.44 .
Large‑Scale Empirical Study
The team evaluated 25 benchmarks , 452 candidate configurations , and 374,401 scored runs covering text generation, code/office agents, and robot policies. Two baselines were compared:
Fixed baseline : pick the best candidate per benchmark on training data and always use it at test time.
Oracle : after seeing all candidates' actual results, choose the highest‑scoring one per input.
On 24 benchmarks with 19,814 complete candidate results, the fixed baseline averaged 82.60 while Oracle reached 87.98 — a 5.4‑point gap . The gap was especially large for video generation ( 18.9 points) and robot manipulation ( 18.4 points).
Capability‑Tree Router for Multi‑Modal Tasks
A unified capability‑tree router was trained on Qwen3.5‑4B . It reads the task text and output requirements, predicts the relevant capability node, and looks up a node‑to‑candidate mapping table built during training. The tree contains 157 leaf nodes and handles five output modalities: text, code/office agents, image, video, and audio. The 4B selector chooses the model; the chosen model executes the task.
Robotics: Visual Outcome Prediction Router
For robot policies, task text alone is insufficient — object positions, camera views, and scene disturbances affect success. A separate visual outcome prediction router takes the instruction, three camera feeds, and robot state, then predicts success probabilities for seven vision‑language‑action policies and selects one to execute.
Training signal design mattered: labeling only the single "champion" policy per scene discarded many successful alternatives. Using inference time to break ties, 94% of champion labels were decided by speed, causing 24,556 of 31,408 successful runs to be treated as errors. Switching to per‑policy outcome prediction (each policy's success/failure contributes to learning) raised robot success rate from 42.8% (champion‑only classifier) to 71.6% .
Performance Gains Across Benchmarks
The unified router was evaluated on 16 benchmarks (3,444 shared test inputs) and scored 79.98 , beating the per‑benchmark fixed‑candidate baseline ( 79.68 ). Notable improvements:
Terminal‑Bench : 69.44 → 77.78
VBench‑2.0 (video): 62.17 → 64.28
OneIG (image): 83.20 → 84.30
The router kept the same candidate for 83%–86% of inputs; gains came from the minority of inputs where switching helped.
Robotics Simulation Results (RoboTwin)
On 2,000 simulated scenes from 10 unseen tasks, overall success rose from the best single policy's 70.45% to 71.55% (+1.1 pp). Average per‑scene policy inference time dropped from 14.51 s to 11.05 s ( 24% reduction ). The router learned to prefer PatchWAM in clean scenes and ME‑Dex in randomized scenes.
Cost‑Aware Routing: CodeRouterBench
An independent CodeRouterBench compared eight models on 9,999 code tasks (2,919 test). Using Opus 4.6 as reference, a state router trained with zero quality loss tolerance achieved only 0.44 points lower quality on test, while average cost per task fell from $0.011655 to $0.000452 — a 96.1% reduction . Quality‑per‑dollar reached 25.5× the reference. For image generation, allowing ≤0.5 point quality change cut cost by 26% with a 0.47 point quality drop.
Conclusion and Future Directions
Route Anything demonstrates that a small selector can dynamically route across heterogeneous candidates — LLMs, agents, robot policies — and exceed fixed strong baselines while drastically reducing cost. The work shows that capability categorization captures only ~8% of the Oracle gap; most complementarity lies within the same task type across different inputs. Preserving every candidate's success and failure signals is crucial for learning these fine‑grained routing decisions. As model and policy libraries grow, accurate dispatch of existing capabilities becomes a key path toward general AI systems.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
