PISA-E, PolyRoute, and the Hydrogen/Carbon Orchestrator

TL;DR

Three routing results in one window, at three layers. PISA (2611.10193) estimates each model's capability on a prompt by combining soft cluster memberships with per-model predictors; PISA-E extends it to choosing an ensemble, both which models and how many, under a cost/quality dial you set at inference time with no retraining. On four routing benchmarks PISA has the best cost-performance hypervolume among single-model routers in- and out-of-distribution; PISA-E beats the best ensemble methods by 2.09 (ID) and 2.87 (OOD) points at the quality end, and at cheaper settings matches them while cutting cost about 36% and 25%. PolyRoute (2611.13843) is the first fixed yardstick for typed-decision encoders (yesterday's decision models) outside English: 12 languages, 9 scripts, code-mixed text such as Hinglish. The English checkpoint scores 61.2%, the shipped multilingual one 77.3%, the best fine-tune 96.6%, and plain TF-IDF plus logistic regression 91.9%. Google's Orchestrator in Antigravity routes inside one session between Gemini 4 Carbon (coordinator) and the new Hydrogen (workers and verifiers), reaching near-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster.

flowchart LR
  Q["Prompt<br/><small>incoming request</small>"] --> C["Soft clusters<br/><small>prompt membership</small>"]
  C --> E["Capability estimates<br/><small>per-model predictors</small>"]
  D["Cost dial<br/><small>set at inference</small>"] --> S
  E --> S["Ensemble select<br/><small>members and size</small>"]
  S -->|cheap| One["One small model<br/><small>36% cheaper point</small>"]
  S -->|quality| Ens["Ensemble<br/><small>+2.1 to +2.9 pts</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class Q,D input
  class C,E core
  class S loop
  class One,Ens exit

Key findings

• PISA-E's ensemble-success surrogate avoids ensemble-specific training, so new models can be added by fitting one predictor.
• PolyRoute: a 3.7-point gap separates TF-IDF (91.9%) from the best fine-tune (96.6%), and the shipped multilingual decision encoder (77.3%) loses to TF-IDF. Typed-decision encoders are not yet robust multilingual routers out of the box.
• On code-mixed text the English checkpoint (85%) is nominally ahead of the multilingual one (79%, p=0.134, not significant).
• Orchestrator: the first hyperscaler-shipped intra-session routing between a frontier and an execution model with a published quota saving (24%).

Gaps

• PISA-E's benchmarks are offline routing datasets; no live-traffic or latency cost of running several models.
• PolyRoute's texts and labels are all LLM-generated, and the split resolves only differences above about 4 points.
• Orchestrator numbers are Google's own, on one benchmark.

How it relates to prior wiki pages

• Extends LLM routing: Route Anything (11-05) found a perfect router gains 5.4 points over the best single model while learned routers gain 0.3. PISA-E attacks the gap from a different side: ensembles raise the ceiling, and adaptive cardinality keeps the cost bounded.
• Tests the 11-12 decision-model wave (Gemini Decision 1): PolyRoute is the first public check of those encoders off-English, and the shipped multilingual checkpoint underperforms a 1990s baseline. Kurate's cs.AI #17 this week (Nudging Decision Models, whether digital nudges steer generation-free agents) is a second early robustness probe.
• Confirms the role-routing pattern (Cohere lead-plus-workers 11-08, AgentBaton 11-10): Orchestrator is Google productizing "frontier lead, cheap workers".