# Synthetic Data Generation

NVIDIA has packaged world-foundation-model training, optimization and inference into Cosmos Framework, with agent-callable skills and checkpoints that run out of the box. Developers previously had to fit a model to their own sensors, action space and data, then get it into an application, two separate rework efforts. The released Cosmos 3 Nano simulation recipes cover driving, robotics and surgery: a Transfer-Auto reference run on four GPUs returns seven 720p videos and 99 LiDAR sweeps, while Sim-Auto, distilled into a causal few-step model, generates observations step by step from driving-policy input and returns them to the policy for closed-loop evaluation.

NVIDIA has packaged world-foundation-model training, optimization and inference into Cosmos Framework, with agent-callable skills and checkpoints that run out of the box. Developers previously had to fit a model to their own sensors, action space and data, then get it into an application, two separate rework efforts. The released Cosmos 3 Nano simulation recipes cover driving, robotics and surgery: a Transfer-Auto reference run on four GPUs returns seven 720p videos and 99 LiDAR sweeps, while Sim-Auto, distilled into a causal few-step model, generates observations step by step from driving-policy input and returns them to the policy for closed-loop evaluation.

For targeted instruction-tuning data selection, QSGS organizes candidate examples by gradient similarity into a graph and applies a submodular graph cut that pulls the subset toward the target task while penalizing internal redundancy, then adds gradient-guided synthesis to fill directions the pool cannot cover. On MMLU, BBH and TydiQA across Llama-2, Llama-3 and Mistral, QSGS at a 0.5% budget (1,350 of 270,000 examples) matches or exceeds full-data performance on Mistral-7B and beats the strongest baseline on every benchmark.

For targeted instruction-tuning data selection, QSGS organizes candidate examples by gradient similarity into a graph and applies a submodular graph cut that pulls the subset toward the target task while penalizing internal redundancy, then adds gradient-guided synthesis to fill directions the pool cannot cover. On MMLU, BBH and TydiQA across Llama-2, Llama-3 and Mistral, QSGS at a 0.5% budget (1,350 of 270,000 examples) matches or exceeds full-data performance on Mistral-7B and beats the strongest baseline on every benchmark.

Handwritten Meitei Mayek has been recognised one character at a time, yet its most frequent confusions depend on where a character sits in a word. Training words are composed from isolated character images and read whole by a convolutional encoder, a bidirectional LSTM and CTC, giving 0.24% character and 1.57% word error on 5,000 synthetic test words; on 100 unseen real words written on a tablet by one native writer, character error is 2.96% and word error 16.0%.

Handwritten Meitei Mayek has been recognised one character at a time, yet its most frequent confusions depend on where a character sits in a word. Training words are composed from isolated character images and read whole by a convolutional encoder, a bidirectional LSTM and CTC, giving 0.24% character and 1.57% word error on 5,000 synthetic test words; on 100 unseen real words written on a tablet by one native writer, character error is 2.96% and word error 16.0%.

A meta model pre-trained offline on synthetic Wiener-Hammerstein systems predicts a new system's output zero-shot from a short input-output context, matches state-of-the-art accuracy on nonlinear identification benchmarks, and cuts an out-of-distribution continuous stirred tank reactor's prediction error from 1.249 to 0.242 by adapting only the input side.

A meta model pre-trained offline on synthetic Wiener-Hammerstein systems predicts a new system's output zero-shot from a short input-output context, matches state-of-the-art accuracy on nonlinear identification benchmarks, and cuts an out-of-distribution continuous stirred tank reactor's prediction error from 1.249 to 0.242 by adapting only the input side.

The authors present ToolRACER, a three-stage synthetic data pipeline coordinating user, assistant, and tool emulation agents that injects user-originated and environment-originated adversarial behaviors, validates traces through four evaluators plus a refinement loop, and yields ToolRACERBench: six domains, 55 personas, 5.6K trajectories with about 66% failure-prone scenarios; models fine-tuned on it improve end-to-end agentic accuracy on τ-bench, BFCLv3, and ACEBench over baseline training sets.

The authors present ToolRACER, a three-stage synthetic data pipeline coordinating user, assistant, and tool emulation agents that injects user-originated and environment-originated adversarial behaviors, validates traces through four evaluators plus a refinement loop, and yields ToolRACERBench: six domains, 55 personas, 5.6K trajectories with about 66% failure-prone scenarios; models fine-tuned on it improve end-to-end agentic accuracy on τ-bench, BFCLv3, and ACEBench over baseline training sets.

The authors introduce InvestigationWorlds: they retrieve 100 real U.S. federal civil cases with decided summary judgment motions from PACER and use an attorney-validated generation pipeline to synthesize role-tagged documents around the original record, so that one corpus admits multiple coherent but incompatible factual readings of which only one matches the court-adopted hypothesis; evaluating six frontier agents on 100 cases, they find agents often commit to incorrect hypotheses despite retrieving relevant evidence, with the highest case-solved rate only 33.3%.

The authors introduce InvestigationWorlds: they retrieve 100 real U.S. federal civil cases with decided summary judgment motions from PACER and use an attorney-validated generation pipeline to synthesize role-tagged documents around the original record, so that one corpus admits multiple coherent but incompatible factual readings of which only one matches the court-adopted hypothesis; evaluating six frontier agents on 100 cases, they find agents often commit to incorrect hypotheses despite retrieving relevant evidence, with the highest case-solved rate only 33.3%.