Year: 2026
Type: conference-paper
Abstract: We introduce Output-Space Search (OS-Search), which turns LLM generation into endpoint search.An outer loop selects a target z * in a frozen encoder-defined 3D output space Z, and a retrieval-grounded policy trained with sequence-level RL generates outputs whose coordinates land near z * under standard autoregressive decoding.This enables parallel sweeps and black-box optimization in Z without path-dependent token/program search.On stories, sweeping Z text yields 3.1× higher LLM-scored diversity than prompt-chaining.On code, Bayesian optimization over Z code improves an objective withheld from the controller under matched inference budgets while preserving validity.tive depends on the endpoint rather than the full trajectory) because z * specifies an intended endpoint region in Z.The outer loop searches over targets z * rather than token paths or program edits.We implement this interface with Output-Space Search (OS-Search), which keeps representation and actuation modular.Representation.We define fixed coordinates z(x) for task outputs using a frozen encoder E and a frozen linear projection (Sec.3, App.A).Actuation.We train a retrieval-grounded policy conditioned on a requested target z * so that, via sequence-level RL, its generations land near that request in Z while also self-reporting ẑ.Use.The resulting controller exposes a black-box actuator F that maps (p, z * ) to a sample x, enabling parallel sweeps and black-box optimization over targets in Z without decoding-time guidance.We evaluate OS-Search on both stories (parallel diversity sweeps) and code (outer-loop black-box optimization in Z), and report target-tracking and calibration diagnostics.Contributions.We make three contributions.• We propose OS-Search: a fixed, external, encoder-defined output space Z together with a z * -conditioned controller trained with sequence-level RL to (approximately) hit requested coordinates and self-report ẑS .• For stories, we build an anchored Z text and show that anchoring yields a stable reference axis z 1 , which we validate via a monotonic relationship between requested z * 1 and an automated templatedness proxy (EQ-Bench Slop-Score).Sweeping a small grid in Z text yields large embarrassingly-parallel multi-branch diversity gains.• For code, we build an unanchored Z code and show it supports black-box optimization: Bayesian optimization (BO) over z * improves a withheld executable objective under matched valid-program budgets.
Author: Tobias Materzok
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166832458