#training-data

11 curated events

papersTODAY 04:00 UTC

Paper Proposes Exploration-Guided Prompt Scaffolding for Multimodal RL Post-Training

A new arXiv paper argues that training prompts in online reinforcement learning vary widely in how useful they are to the current policy, with some already solved and others too hard to give a dependable learning signal. The authors propose an exploration-guided prompt scaffolding method that selects or structures prompts for multimodal reinforcement post-training so rollouts are better spent. The work appears in both the cs.AI and cs.LG listings as arXiv:2609.15051v1.

reinforcement-learning/multimodal/training-data/arxiv

papersTODAY 04:00 UTC

Paper Proposes Joint Optimization of Prompts and Training Data via Failure Signals

A new arXiv paper addresses automatic prompt optimization, a technique that normally revises prompts using task feedback while leaving the training set unchanged. The authors argue that repeatedly tuning against the same examples limits feedback to already-known weaknesses, and they propose a failure-guided approach in which prompts and training data are improved together. This co-evolution aims to surface new shortcomings rather than only correcting previously identified ones.

prompt-optimization/training-data/llm/research

papersTODAY 04:00 UTC

Paper Analyzes Selection Bias When Model Edits Target Localized Spans

A new arXiv paper examines what happens when human corrections are applied only to identified editable spans of a model's output. The authors decompose the localized gradient into edited and untouched portions at a fixed checkpoint, showing that selective feedback channels can amplify relative selection bias. They also study gradient geometry, target mismatch, and importance weighting as factors in this effect.

research/fine-tuning/bias/training-data

papersTODAY 04:00 UTC

arXiv paper analyzes how AI-generated data affects dataset decomposition

A new preprint examines what happens to batch decomposition and downstream model performance when training sets mix human data with text or images produced by existing large language models. Using random datasets containing anomalies, the authors study criticality in dissimilar decomposition and undersampling techniques. The work aims to clarify the statistical behavior of datasets that are increasingly populated with synthetic samples.

research/training-data/synthetic-data/llm

industrySEP 12 08:10 UTC

Apple to Train AI Models on User Data, Reversing Earlier Privacy Pledge

Apple has changed course and now plans to use data from iPhone users to train its AI models, abandoning its previous commitment that personal information would never be used for that purpose. The shift reverses a promise the company had presented as a firm guarantee to customers.

privacy/apple/training-data/user-data

papersSEP 10 04:00 UTC

arXiv paper proposes data-centric post-training pipeline for financial reasoning

A new research paper tackles the shortage of training data suitable for reasoning-focused fine-tuning in the financial domain, noting that most available QA pairs lack explicit reasoning steps, sufficient context, or reliably checkable answers. The authors present a pipeline that mines financial text, distills it into reasoning-oriented training examples, and applies learning with verifiable answers to improve model performance on financial tasks.

reasoning/finance/training-data/distillation

papersSEP 10 04:00 UTC

Study: Multilingual Data Mixing Helps LLMs Reason in Non-English Languages

A new arXiv paper tackles the tendency of reasoning language models to think in English even when prompted in other languages, which limits access for non-English speakers. The authors show that how training data is mixed across languages is key to getting models to generalize and carry out reasoning in the user's language itself. The work offers a path toward making advanced reasoning capabilities usable beyond English.

multilingual/reasoning/llm/training-data

papersSEP 12 04:00 UTC

arXiv Paper Models Collapse When Multiple AI Systems Train on Each Other's Output

A new arXiv preprint examines how recursive training on AI-generated text leads to model collapse, extending prior work from a single model to settings where many models exchange and train on one another's outputs. The authors analyze how the dynamics play out across a multi-model ecosystem, where each participant learns from a shared pool of synthetic data. The work is a preprint and has not yet been peer reviewed.

model-collapse/synthetic-data/research/training-data

papersSEP 12 04:00 UTC

arXiv paper proposes fragility spectrum for recursive language-model training

A new arXiv preprint examines what happens when text produced by language models is fed back into their own training data, a practice linked to shrinking output diversity. The authors propose a "fragility spectrum" framework to characterize how different training protocols and data mixtures degrade under this recursive loop. The work aims to give researchers a more systematic way to compare which setups hold up and which break down.

arxiv/training-data/model-collapse/research

industrySEP 10 11:00 UTC

Mathematicians seek proof OpenAI did not train on their work

Researchers are asking OpenAI to demonstrate whether their published mathematical papers were included in the company's training data. The demand follows a public dispute over an OpenAI-credited mathematical result that critics say closely resembles existing literature. No evidence about the model's data sources has been released by OpenAI so far.

openai/training-data/copyright/research

industrySEP 11 22:58 UTC

Mecka AI in talks for Sequoia-led round at reported $500M valuation

Mecka AI, a two-year-old startup focused on robot training data, is close to securing a funding round led by Sequoia Capital that would value it at roughly $500 million. The deal would arrive just months after the company announced its Series A. Investor interest in robot training data has grown as AI firms look to apply their models to physical tasks.

funding/robotics/sequoia/training-data