LookThere checkpoints

Checkpoints for LookThere! Sparse Vision by Reinforced Selection (NeurIPS 2026).

Code: https://github.com/SreehariRamMohan/lookthere

Checkpoints

All models are built on DINOv2 ViT-B/14. Accuracy is measured on the validation set used in the paper. The model picks patches at random from its learned policy, so the accuracy can change by a few tenths of a point between runs.

file	task	config	accuracy @ sparsity
billiards/max-left-right/sparsity90/model.pt	Billiards (max-left-right)	configs/billiards/billiards_90_rl.yaml	96.7 @ 90.0%
billiards/max-left-right/sparsity99/model.pt	Billiards (max-left-right)	configs/billiards/billiards_99_rl.yaml	93.8 @ 99.0%
tt100k/sparsity95/model.pt	TT-100K traffic signs	configs/tt100k/tt100k_95_rl.yaml	95.9 @ 94.8%
tt100k/sparsity99.6/model.pt	TT-100K traffic signs	configs/tt100k/tt100k_99_rl.yaml	90.0 @ 99.6%
birdsnap/sparsity90_distillinit/model.pt	Birdsnap	configs/birdsnap/birdsnap_95_rl.yaml	88.0 @ 90.2%
birdsnap/sparsity95_distillinit/model.pt	Birdsnap	configs/birdsnap/birdsnap_95_rl.yaml	85.4 @ 95.1%
birdsnap/sparsity97_distillinit/model.pt	Birdsnap	configs/birdsnap/birdsnap_95_rl.yaml	80.5 @ 97.1%
imagenet/sparsity90/model.pt	ImageNet-1K	configs/imagenet/imagenet_90_rl.yaml	81.6 @ 90.0%
imagenet/distill_pretrain/model.pt	ImageNet-1K (distilled from DINOv2)	configs/imagenet/imagenet_distill_dinov2.yaml	no accuracy (67.6% sparsity)

The Birdsnap models start from imagenet/distill_pretrain and are fine-tuned without a warmup.

Warmup checkpoints

These are the Billiards models after the warmup, which trains on all patches before RL. Use them to start RL training yourself. They also save the optimizer state, so training picks up exactly where it stopped.

file	task	accuracy (all patches)
billiards/max-left-right/warmup/model.pt	max-left-right	97.5
billiards/rightmost-number/warmup/model.pt	rightmost-number	98.1
billiards/left-color-min-max/warmup/model.pt	left-color-min-max	98.7

SHA256SUMS.txt lists a checksum for every file.

Usage

pip install -U huggingface_hub
hf download SreehariRamMohan/lookthere --local-dir lookthere_ckpts

# Evaluate a model.
python train.py eval configs/billiards/billiards_90_rl.yaml ckpt=lookthere_ckpts/billiards/max-left-right/sparsity90/model.pt

# Run RL training from a warmup checkpoint.
python train.py rl configs/billiards/billiards_95_rl.yaml init=lookthere_ckpts/billiards/max-left-right/warmup/model.pt

Each file loads with torch.load(path, weights_only=True). It holds the weights of the selector and the extractor, plus the settings the model needs at evaluation time. The ImageNet 90% model adds 3 to every selector score (logit_offset). This value is saved in the file and used automatically.

License

Apache 2.0

Citation

@inproceedings{rammohan2026lookthere,
  title={LookThere! Sparse Vision by Reinforced Selection},
  author={Sreehari Rammohan and Yousef Yassin and Anthony Fuller and Junfeng Wen and Carl Vondrick and Evan Shelhamer},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2609.04698}
}