Builders & open source
Alexander Toshev
Listed with Apple on the RLTL;DR paper (September 2026), Apple
Alexander Toshev is a machine learning researcher at Apple, listed with Apple on papers through September 2026. Toshev's recent Apple work covers training agents, including RLTL;DR, a 2026 method in which an agent writes short feedback insights after failed attempts and conditions later attempts on them. Toshev is also a co-author of DataComp-LM, a testbed for controlled dataset experiments built around a standardized Common Crawl corpus, and of MobileCLIP2, which improves the training of low-latency image-text models.1234
At a glance
- Current role
- Listed with Apple on the RLTL;DR paper (September 2026)
- Company
- Apple
- Sector
- Company
- Fields
- Agents, Computer vision, Datasets and data
- Education
- Penn
- Work listed
- 3
- Sources cited
- 7
- Profile reviewed
Research
Alexander Toshev's early deep learning papers addressed computer vision. DeepPose, which Toshev wrote with Christian Szegedy in 2013, frames human pose estimation as regression of body-joint positions with a cascade of deep neural networks. In 2014 Toshev co-authored Show and Tell, a neural image caption generator built on a deep recurrent architecture that combines advances in computer vision and machine translation to generate sentences describing an image.56
More recent co-authored papers cover training data and image-text models. DataComp-LM (2024) offers a standardized Common Crawl corpus, OpenLM-based pretraining recipes and a suite of downstream evaluations for controlled dataset experiments, and MobileCLIP2 (2025) improves the multi-modal reinforced training of low-latency image-text models, with pretrained models and data-generation code released.34
Agents
Several recent papers co-authored by Toshev concern agents. LLaRP (2023) adapts a frozen pre-trained language model into a policy for embodied visual tasks trained with reinforcement learning. A 2026 paper uses language models as simulated world models to generate stateful training data for API-calling agents, and RLTL;DR (2026) has an agent write short feedback insights after failed attempts and learn to internalize them.721
Work and contributions
DataComp-LM: In search of the next generation of training sets for language models Published research
Co-author, listed with Apple, of the 2024 DataComp-LM paper, which provides a standardized corpus, pretraining recipes, and downstream evaluations for dataset experiments at model scales from 412M to 7B parameters.3
RLTL;DR: Self-improvement by Internalizing Self-generated Feedback Published research
Co-author of the 2026 Apple paper introducing RLTL;DR for self-improvement in reinforcement learning with verifiable rewards.1
MobileCLIP2: Improving Multi-Modal Reinforced Training Published research
Co-author, listed with Apple, of the 2025 MobileCLIP2 paper on improved multi-modal reinforced training for low-latency image-text models.4
Timeline
- Posted DeepPose, a deep neural network method for human pose estimation, with Christian Szegedy.5
- Co-authored Show and Tell, a neural image caption generator.6
- Co-authored LLaRP, which adapts large language models into policies for embodied tasks.7
- Co-authored DataComp-LM, a testbed for language model training-set experiments.3
- Co-authored MobileCLIP2 on improved reinforced training for low-latency image-text models.4
- Co-authored a paper on scaling stateful supervision for API-calling agents with simulated world models.2
- Co-authored RLTL;DR, a self-improvement method built on self-generated feedback.1
Where they work
- Apple 27 people in the wiki
Related people
- Fartash FaghriAlso at Apple
- Niko GrupenAlso at Apple
- Vaishaal ShankarAlso at Apple
- Sachin MehtaAlso at Apple
- Vipul Ved PrakashAlso at Apple
- Matt DeitkeComputer vision · Datasets and data
- Jerry LiuAgents · Datasets and data
- Stefano SoattoAgents · Computer vision
- Tsung-Yi LinComputer vision · Datasets and data
- Piotr DollárComputer vision · Datasets and data
Suggest a correction or removal
Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.
References
7 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026
- RLTL;DR: Self-improvement by Internalizing Self-generated Feedback (arXiv 2609.37633)
- Simulate to Generalize: Scaling Stateful Supervision for API-calling Agents using LLM World Models (arXiv 2607.16900)
- DataComp-LM: In search of the next generation of training sets for language models (arXiv 2406.11794)
- MobileCLIP2: Improving Multi-Modal Reinforced Training (arXiv 2508.20691)
- DeepPose: Human Pose Estimation via Deep Neural Networks (arXiv 1312.4659)
- Show and Tell: A Neural Image Caption Generator (arXiv 1411.4555)
- Large Language Models as Generalizable Policies for Embodied Tasks (arXiv 2310.17722)