Hyeonseong Jeon

PhD Applicant (Fall 2027) | CS BS @SeoulNatlUni '25 | Exchange @uwcse '24

[Hyeonseong Jeon @yeonsumia]

Long-horizon values depend on shorter-horizon values.
But what if those shorter-horizon values are still wrong?

We introduce DCRL: a recursive divide-and-conquer approach that learns values from short to long, substantially mitigating value-error accumulation over long horizons compared with standard GCRL methods. 🧵

[Hyeonseong Jeon @yeonsumia]

@a398877956 Regarding goal representations, I think TRL authors had already shown comparison on 1B datasets (https://t.co/8WtzVp0ibj).

[Hyeonseong Jeon @yeonsumia]

Long-horizon values depend on shorter-horizon values.
But what if those shorter-horizon values are still wrong?

We introduce DCRL: a recursive divide-and-conquer approach that learns values from short to long, substantially mitigating value-error accumulation over long horizons compared with standard GCRL methods. 🧵

[Hyeonseong Jeon @yeonsumia]

@a398877956 I think that data coverage matters for navigation tasks.

[Hyeonseong Jeon @yeonsumia]

@a398877956 Sure! Could you share more details (e.g., what changes you've made, environment, training steps, etc)?

[Hyeonseong Jeon @yeonsumia]

@a398877956 Thanks for your interest! We evaluated DCRL in both oracle state-based and pixel-based settings.

[Hyeonseong Jeon @yeonsumia]

We’re releasing the code to fully reproduce the main results from our paper!

Code: https://t.co/gzWEtXaRhD

[Hyeonseong Jeon @yeonsumia]

Long-horizon values depend on shorter-horizon values.
But what if those shorter-horizon values are still wrong?

We introduce DCRL: a recursive divide-and-conquer approach that learns values from short to long, substantially mitigating value-error accumulation over long horizons compared with standard GCRL methods. 🧵

[Hyeonseong Jeon @yeonsumia]

@baymax3009 Thank you for the insightful feedback! In Figures 7 and 8, we compare our approach with a recent "explicit" quasimetric method, TMD (Myers et al., 2025).

[Hyeonseong Jeon @yeonsumia]

Thanks for your interest! I believe extending DCRL to a general off-policy setting would be an important direction. Our core idea is to learn behavior values recursively while simultaneously recovering optimal values. To this end, we may need to make some changes to standard off-policy algorithms. I’m currently exploring this direction, and I hope I’ll be able to share more soon!

[Hyeonseong Jeon @yeonsumia]

@Void_The_Null @YoungwoonLee @seohong_park Great question! Extending DCRL to stochastic real-world settings is an important direction for future work.

[Hyeonseong Jeon @yeonsumia]

TL;DR: Long-horizon goal-conditioned value learning suffers from error accumulation. Recursively learning short-segment values before long-segment values mitigates this accumulation.

For more details, check out our paper and project website.

Paper: https://t.co/dyZD8d5Oad
Website: https://t.co/ejrYFoC092

w/ @YoungwoonLee

[Hyeonseong Jeon @yeonsumia]

Across diverse goal-reaching tasks, DCRL substantially outperforms prior offline GCRL methods. On the five most challenging long-horizon OGBench tasks, it raises the best prior average score from 55 → 64, outperforming both flat and hierarchical baselines.

[Chan Hee (Luke) Song @luke_ch_song]

Do VLMs actually understand 3D space 🌎?

Or are they exploiting shortcuts hidden in natural images?

🚀 Excited to share our new work:

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

@NVIDIAAI × @SeoulNatlUni × @OhioStateCSE

🧵👇

[Ainaz Eftekhar @ainaz_eftekhar]

🎉 Excited to share that our paper “Convergent Functions, Divergent Forms” will be presented at NeurIPS 2025🤖 in San Diego!
We present LOKI, a compute-efficient framework for co-evolving robot morphologies🦾 and control policies⚙️. LOKI discovers diverse, high-performing robot designs through shared control policies, addressing key challenges of bi-level evolution such as premature convergence and behavior collapse. It’s also significantly more efficient than Quality-Diversity methods, achieving greater diversity in locomotion behaviors and stronger transfer to various downstream tasks.
👉Project page: https://t.co/JKpy1d1hMP