Nudge: feedback-driven decision tuning for small models, +14.8% on RouterBench
sockpuppet17263 · reddit · 2026-10-17
Nudge is an open-source Python library that sits on System One models like Jev or Rune and shifts future picks toward what worked, using outcome feedback — no fine-tuning, no GPUs, just one SQLite file.
How it works: each call logs the model's odds and chosen option; you report outcomes later. To disambiguate situations, it asks yes/no, 1–5, or multiple-choice questions about text already read (proposed by an LLM or written by you), discarding non-discriminative ones. It fits a per-option Bayesian linear model, picks via Thompson Sampling, and keeps a 5% exploration floor.
On RouterBench (5,000 real prompts, quality-minus-cost routing), Rune alone scored 0.581 vs 0.667 with Nudge — a 14.8% gain. v0 research software; pip install nudge-policy.
More from coding & agent
- TwelveLabs' video intelligence agent Jockey hits GA, plugs into ChatGPT and Claude via MCP — lmoroney · 2026-10-17
- Google demo: Gemini rebuilds ancient ruins into explorable 3D worlds via Antigravity + Blender MCP — jnack · 2026-10-17
- Mixpanel MCP Server Lets AI Agents Query Funnels and Retention via Natural Language — modelcontextprotocol · 2026-10-17
- MCP Connector Ties AI Assistants to UK's Largest FOI Platform WhatDoTheyKnow — modelcontextprotocol · 2026-10-17
- Codex mobile remote-control pain: server-relay design ignores at-home phone use — tinyfool · 2026-10-17
- Dev uses a custom Skill to cut a product launch video, extending AI agent workflows — AlchainHust · 2026-10-17