NanoHorizon
LiveTrain the strongest Craftax policy in 30 minutes from a Qwen3.5-2B starter. Compare receipt-backed submissions, achievement frequencies, and replayable policy traces.
Open challenge →Synth Research
Research at Synth
We study how language-model agents improve, operate across long horizons, and produce evidence that can be independently verified. We publish open challenges, evaluations, systems, reports, and field notes.
01
Challenges
Open research problems with fixed contracts, public starters, held-out evaluation, and receipt-backed leaderboards.
02
Evaluations
Frozen benchmark contracts, comparable costs, and rollout-level evidence for understanding how models and agents actually behave.
GameBench Code Policies
NewExecutable policies across frozen game environments, paired with source, replayable visual rollouts, and measured candidate results.
Open evaluation →Craftax
LiveLong-horizon survival, crafting, exploration, and combat with cost curves and replayable terminal traces.
Open evaluation →DungeonGrid
PilotFog-of-war coordination across single- and multi-agent parties, measured against inference cost.
Open evaluation →Shrimp Farm
LiveA 730-day operational benchmark with strict turn, budget, and output contracts.
Open evaluation →Evaluation reports and methodology notes are being prepared for publication.
03
Optimization & Learning
Methods for improving prompts, policies, and long-horizon agents through measured search and replay.
Synthesis in progress
GEPA
Reflective prompt evolution, Pareto candidate coverage, and compute scaling across stable task contracts.
Active research
GELO
Archive-based optimization that returns to useful trajectory states before exploring the next frontier skill.
Report
GELO: Archive-Based Optimization for Long-Horizon Agents
Go-Explore Long-Horizon Optimizer (GELO) is a checkpoint-native optimizer for language agents in sparse, long-horizon environments. GELO maintains an archive of high-value trajectory states, opens scoped theme searches from those checkpoints, and promotes candidates only when replayed heldout evidence improves the archive. In Craftax, NetHack, and Crafter, GELO turns undifferentiated prompt mutation into targeted frontier search over skills such as furnace placement, coal and iron recovery, and dungeon coverage.
Jun 11, 2026
Report
Scaling Train Time Compute for Gepa
Scaling train-time compute for GEPA across public task containers — same-container comparisons, coverage curves, and proposer scaling evidence.
Jun 2, 2026
04
Systems & Open Source
Technical reports on the open and internal systems that make agent research reproducible, inspectable, and extensible.
Public building blocks from Synth Laboratories.
View the GitHub organization ↗01GitHub
Cookbooks
Runnable GEPA, HealthBench eval, local MLX SFT on Craftax, and CISPO examples with measured receipts.
Explore the repository ↗02GitHub
Optimizers
An Apache-2.0 Rust optimization platform with a public GEPA engine and Python API, plus the SDK, CLI, and runbooks for hosted GELO jobs.
Explore the repository ↗03GitHub
Containers
An MIT-licensed Python SDK and HTTP task contract for datasets, mutable prompt programs, rollouts, scoring, traces, checkpoints, and resume.
Explore the repository ↗05
Field Notes
Working observations, negative results, reproductions, and early findings that are useful before they become formal reports.
Field Notes publications are in preparation.