Weak model drafts, strong model reviews: SpecAgent, EdgeWise, LEASH

TL;DR. Three papers on the same HF day (2026-10-16) make the cheap model the default worker in a long agent run and call the strong model only for some steps. They differ in what triggers the strong model. SpecAgent turns the strong model into a reviewer: the weak model drafts a run of grounded steps, the strong model verifies the whole segment in one prefill pass and repairs the first bad step. EdgeWise trains the small on-device model with step-level RL to decide for itself when to delegate to a cloud LLM. LEASH is training-free: it compares the weak agent's hidden states with stored states from past successes and failures, escalates when the state looks like a failure, then keeps watching the strong black-box agent with a shadow observer. LEASH is also on Kurate's cs.AI top-20 (#18), so it is cross-source confirmed (HF + Kurate).

flowchart LR
  T["Agent task<br/><small>long tool-use run</small>"] --> W["Weak model<br/><small>drafts every step</small>"]
  W --> G{"Trigger<br/><small>who decides escalation</small>"}
  G -->|SpecAgent| V["Strong reviewer<br/><small>one prefill, verdict</small>"]
  G -->|EdgeWise| D["Learned delegate<br/><small>RL-trained policy</small>"]
  G -->|LEASH| U["Latent uncertainty<br/><small>hidden-state kNN</small>"]
  V --> S["Strong model step<br/><small>repair or take over</small>"]
  D --> S
  U --> S
  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 T input
  class W,S core
  class G,V,D,U loop

Key findings

SpecAgent (arXiv 2610.21463).

• The weak model runs read-only tool calls while it drafts, so every drafted step is grounded in real observations. It stops at the first action that would change the environment.
• The strong model checks the drafted segment in a single prefill pass, commits the longest valid prefix, and rewrites the first rejected step. It decodes only short verdicts and repairs.
• On τ²-Bench it comes within one point of the strong model's success rate while roughly halving memory traffic and wall time. On AutomationBench and VitaBench it recovers 44-75% of the quality gap. In an edge-cloud setup on MCP-Atlas it closes 53-75% of the gap with the strong model generating at most half the tokens.

EdgeWise (arXiv 2610.17560).

• Finding: calling the cloud LLM on a small subset of key steps matches full-LLM execution.
• Step-level RL jointly optimizes task solving and delegation, with credit assigned to each delegation decision. No extra router model at inference.
• Matches or beats pure-LLM execution while cutting inference cost up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1. Also studies when to delegate, what context to share, and how often.

LEASH (arXiv 2610.22280). Cross-source confirmed (HF + Kurate cs.AI #18).

• Three layers: task-level uncertainty for early routing, step-level uncertainty for runtime escalation, and a shadow observer that keeps monitoring the strong black-box agent after handoff.
• Uncertainty = similarity of the current hidden state to hidden states from past failed vs successful trajectories. It reuses activations already produced, so no extra forward pass and no trained monitor.
• On OSWorld across four backbones, beats random escalation by 6.55-11.94 points at matched escalation rate. Qwen3.5-35B-A3B + LEASH: 73.66% at $0.41/task, matching always-on Claude Opus 5 (72.90%) at 19.2% lower cost.

Relation to prior wiki pages

• Extends the granularity ladder. The 10-16 routing page completed request → session → role → turn. These three go one finer: the step inside a turn, and SpecAgent even batches steps into verified segments.
• Fills the 10-15 gap on where to fire. Reading Right, Acting Wrong (10-15) found most wrong next actions follow a correct reading of the state and cluster on the same states. LEASH's failure-memory kNN is a direct way to detect those states; it does not cite that result.
• Speculative decoding at the trajectory level. SpecAgent is the speculative-decoding pattern (speculative-decoding: cheap draft, parallel verify) lifted from tokens to tool-calling steps. Verification by prefill instead of decode is the same cost trick.
• Confirms the asymmetric-switch result. Whose Turn (10-16) showed switching up mid-session loses the prompt cache. SpecAgent sidesteps this: the strong model reads the drafted segment once as prefill, so it pays a prefill but not a full re-decode.
• Tension. Fewer Bits, More Tries (10-15) argued retrying a small model can replace escalation on verifiable tasks. These papers assume the strong model is needed for some steps. Both may hold: retry for verifiable steps, escalate for unverifiable judgment steps.

Gaps

• No head-to-head among the three triggers on one benchmark.
• LEASH's failure memory needs labeled past trajectories in the same environment; transfer to a new task family is untested.
• SpecAgent stops drafting at the first write action, so write-heavy tasks get less benefit.

Concept page: llm-routing.