Research
Papers on food embeddings, physical AI, and when combining language models actually helps. All of them are also on Google Scholar.
- When More Planning Hurts: Learned Planners for World ModelsOct 2026
A 1,505-parameter learned planner for world models, and why extra planning steps don't rescue a failing plan.
bibtexhide bibtex
@misc{chen2026whenmoreplanninghurts, title={When More Planning Hurts: Learned Planners for World Models}, author={Josef Chen}, year={2026}, eprint={2610.37891}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2610.37891} } - RoboJev: An Open Vision-Language Judge for Robot Success DetectionOct 2026
An open 27B model that watches ten frames of a robot's attempt and says whether it worked: 95.2% balanced accuracy on held-out episodes.
bibtexhide bibtex
@misc{chen2026robojevanopenvision, title={RoboJev: An Open Vision-Language Judge for Robot Success Detection}, author={Josef Liyanjun Chen}, year={2026}, eprint={2610.33718}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2610.33718} } - FlavourBench: Executable Culinary Reward Maps for Language Model Evaluation and Post-TrainingAug 2026
534 cooking tasks scored against Epicure's reward maps, across 27 frontier models.
bibtexhide bibtex
@misc{chen2026flavourbenchexecutableculinaryreward, title={FlavourBench: Executable Culinary Reward Maps for Language Model Evaluation and Post-Training}, author={Josef Chen and Erim Hayretci}, year={2026}, eprint={2608.20574}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2608.20574} } - Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent ControlAug 2026
When concurrent agent work fills a GPU, and why routing on device beats a host round trip.
arXiv:2608.12123 · pdf · dataset
bibtexhide bibtex
@misc{chen2026readycohortsboundinggpu, title={Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control}, author={Josef Liyanjun Chen}, year={2026}, eprint={2608.12123}, archivePrefix={arXiv}, primaryClass={cs.DC}, url={https://arxiv.org/abs/2608.12123} } - When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier ModelsJun 2026
Combining LLMs rarely beats the single best model. A co-failure ceiling, measured across 67 frontier models.
bibtexhide bibtex
@misc{chen2026whendoescombininglanguage, title={When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models}, author={Josef Chen}, year={2026}, eprint={2606.27288}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2606.27288} } - Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing SoJun 2026
Robot flash memory as depreciating capital, with a shadow price on every write.
bibtexhide bibtex
@misc{chen2026memoryasawasting, title={Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing So}, author={Josef Liyanjun Chen}, year={2026}, eprint={2606.18144}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2606.18144} } - AEGIS: A Backup Reflex for Physical AIJun 2026
Detecting imminent manipulation failure from a weak policy's internal states, then switching to a stronger one just in time.
bibtexhide bibtex
@misc{chen2026aegisabackupreflex, title={AEGIS: A Backup Reflex for Physical AI}, author={Josef Chen}, year={2026}, eprint={2606.06660}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2606.06660} } - AURA: Action-Gated Memory for Robot Policies at Constant VRAMJun 2026
Write to memory only when it would change the next action.
bibtexhide bibtex
@misc{chen2026auraactiongatedmemory, title={AURA: Action-Gated Memory for Robot Policies at Constant VRAM}, author={Josef Chen}, year={2026}, eprint={2606.02775}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2606.02775} } - Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM DecodeMay 2026
Why faster GPUs disappoint at batch-1: launch overhead, not bandwidth.
bibtexhide bibtex
@misc{chen2026memoryboundbutnot, title={Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM Decode}, author={Josef Chen}, year={2026}, eprint={2605.30571}, archivePrefix={arXiv}, primaryClass={cs.AR}, url={https://arxiv.org/abs/2605.30571} } - Epicure: Navigating the Emergent Geometry of Food Ingredient EmbeddingsMay 2026
Ingredient embeddings trained on 4.14M recipes, from chemistry to recipe context.
arXiv:2605.22391 · pdf · explorer
bibtexhide bibtex
@misc{radzikowski2026epicurenavigatingtheemergent, title={Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings}, author={Jakub Radzikowski and Josef Chen}, year={2026}, eprint={2605.22391}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2605.22391} } - Epicure: Multidimensional Flavor Structure in Food Ingredient EmbeddingsApr 2026
Fifteen independent dimensions of taste, texture, geography and culture, recovered from embedding space.
arXiv:2604.22776 · pdf · explorer
bibtexhide bibtex
@misc{radzikowski2026epicuremultidimensionalflavorstructure, title={Epicure: Multidimensional Flavor Structure in Food Ingredient Embeddings}, author={Jakub Radzikowski and Josef Chen}, year={2026}, eprint={2604.22776}, archivePrefix={arXiv}, primaryClass={cs.CY}, url={https://arxiv.org/abs/2604.22776} }