Teach an AI model toforecastyourbusinesstopredictworldeventstoreasonwithdatatoprovenewtheoremstocontinuouslylearntohavegreattaste

Bridgewater used Tinker to train Qwen3-235B to recognize what matters to its investors: relevant financial news, signals of interest-rate changes, and useful analysis buried in lengthy documents.

Across six financial information-filtering tasks, the trained model achieved 84.66% accuracy on held-out documents versus GPT-5.5’s 78.2%, approximately 30% fewer errors at one-fourteenth the inference cost.

Lightning Rod Labs used Tinker to train gpt-oss-120b to predict clinical events from patient notes (including whether a patient would receive dialysis, a blood transfusion, or mechanical ventilation later in their hospital stay).

In a retrospective evaluation on 500 held-out questions, the trained model reduced prediction error by 27% and expected calibration error by 69% compared with the prompted base model, achieving performance comparable to GPT-5.

Mantic used Tinker to train gpt-oss-120b to forecast geopolitical, economic, and technological events. The team paired roughly 10,000 resolved forecasting questions with research gathered in advance, then used reinforcement learning to reward probability estimates that better matched real-world outcomes.

On unseen questions from the Metaculus Q2 2025 benchmark, the model’s rescaled log score rose from 38.6 to 45.8.

Monoid Labs used Tinker to train DeepSeek V3.1 to research questions, weigh evidence, and predict real-world outcomes.

On a backtest of 100 resolved questions from the Metaculus Spring 2026 AI Benchmark, the trained model achieved approximately 34% lower forecasting error than untrained DeepSeek V3.1 and 22% lower error than Gemini 3.1 Pro High, measured by Brier score.

Glean used Tinker to train Waldo, an agentic search model built on NVIDIA Nemotron 3 Nano, to find the enterprise information needed to answer users’ questions.

Integrated into Glean’s full system with adaptive reasoning, Waldo helped reduce latency by approximately 50% and token usage by 25% while maintaining comparable answer quality.

Chroma used Tinker to turn gpt-oss-20b into Context-1, a specialized search agent. Through supervised fine-tuning and reinforcement learning, it learned to break questions into targeted searches, find supporting evidence, and discard irrelevant material to make room for further research.

Across its evaluations, Chroma also reported retrieval quality comparable to frontier models, at a fraction of the cost and up to 10× faster inference.

Exa used Tinker to train Qwen3-4B to search the web, connect evidence across multiple sources, and answer complex factual questions.

On 2WikiMultihopQA, training raised the model’s accuracy from 50.0% to 83.9%, a 33.9-percentage-point improvement. The trained 4B model also outperformed Qwen3-235B, which scored 77.4%, with both models using the same Exa search tools.

Long-running tasks can produce feedback after the model has already changed. Trajectory used Tinker to develop SDPO++, a self-distillation method that keeps training stable with these delayed results—even when most attempts fail and only one rollout is available per task.

On APEX-Agents, GPT-OSS-120B’s pass rate rose from 5% to 25%—a fivefold gain on the benchmark and nine percentage points above the original SDPO method.

MIT researchers developed Self-Distillation Fine-Tuning (SDFT), which uses expert demonstrations to guide a model’s learning from its own responses. Tinker helped the team rapidly prototype ideas.

On tool use, SDFT achieved 70.6% accuracy versus 63.2% with standard supervised fine-tuning. Average performance on prior capabilities stayed near the base model’s level, while standard fine-tuning caused a substantial decline

Mercor builds RL environments for AI labs and validates them by measuring how much models improve through training. Using Tinker, it trained Nemotron 3 Ultra on 1,240 tasks spanning investment banking, law, and consulting, using realistic document workspaces, tools, and expert-written rubrics.

After 60 RL steps, the model fully solved twice as many tasks on the held-out 480-task APEX-Agents benchmark: 110 versus 55.

AfterQuery uses Tinker to test how its post-training datasets improve models across coding, knowledge work, and tool use.

For terminal-based tasks, AfterQuery combined supervised fine-tuning on successful agent trajectories with reinforcement learning in Harbor environments, rewarding test completion and efficient solutions. The pipeline raised gpt-oss-20b’s Terminal-Bench 2.0 pass rate from 3.1% to 17.0%, more than a fivefold improvement

Architect Labs trains LLMs to turn English chip-design specifications into functional Verilog code. Starting with Kimi-K2.5, the team used reinforcement learning (RL) to teach the model to write Verilog that is both syntactically and semantically correct, guided by a training curriculum of gradually increasing difficulty.

The resulting model, Architect v0.1, raised pass@1 from 48.4% to 54.5% on the 198-problem CVDP benchmark.

Krea builds foundational image models and creative tools. It uses Tinker to train custom models for prompt expansion and reward modeling, supporting reinforcement learning and evaluation for its image models.

For its latest prompt-expansion pipeline, Krea trained Qwen 3.8 27B to produce more concise reasoning. The fine-tuned model matched Gemini 3.8 Flash’s performance in 90% of evaluated cases while running 15–30× faster at 1/10th the inference cost.

Stanford researchers developed Counterfactual Simulation Training (CST), rewarding reasoning that helps another model predict how answers would change under modified inputs.

In a released experiment using Tinker, the team trained GPT-OSS-20B to make these influences easier to detect. After two training rounds, monitoring accuracy increased from 68.2% to 83.5% on 800 held-out examples, an improvement of 15.3 percentage points.

Axiom Math builds AI for formal mathematics, where every claim has a checkable ground truth. AxiomProver uses Tinker to train via RL, conjecturing, proving, and verifying its own work.

In December 2025 it solved all 12 problems on the Putnam, the world’s hardest college-level math competition: 8 by the end of the competition window and the remaining 4 in the days after, every proof formally verified in Lean.