---
tags: notes, Projects
---
# Algonaute 2025 models
A little walk through how do I evaluate code base for scientific projects.
*Disclaimers*: I only read the TRIBE paper and the summary paper
[Insights from the Algonauts 2025 Winners](https://arxiv.org/abs/2508.10784v1) in full.
## Links to the GitHub repos
[GitHub List](https://github.com/stars/htwangtw/lists/algonautes2025)
:first_place_medal: :moyai: TRIBE: https://github.com/facebookresearch/algonauts-2025
:second_place_medal: :vibration_mode: VIBE: https://github.com/Bicanski-NCG/VIBE
:third_place_medal: :repeat: Multimodal Recurrent Ensembles: https://github.com/erensemih/Algonauts2025_ModalityRNN
:sports_medal: :feather: Multimodal LLMs and a Lightweight Encoder: https://github.com/MedARC-AI/algonauts2025
:sports_medal: :question: CVIU-UARK: no info, very sad
:sports_medal: :seal: Multimodal Seq2Seq Transformer: https://github.com/Angelneer926/Algonauts_challenge
## What do I look for in a code base without running it
1. Set up instructions
2. Documentation on workflow
3. Clear versioning / release (rare)
4. **Can I identify key ideas / important features in the paper in the code?**
### Set up instructions
:slightly_smiling_face: - :moyai::vibration_mode::repeat::feather::seal:
All projects provide some environmental set up instructions. TRIBE model :moyai: has the least standard setup. Most projects have a simple one liner installation.
### Documentation on workflow
Everyone wrote a paper so I guess everyone pass the lowest bar, but I do like a pretty readme page. This also provide context for mapping the code base into the science :scientist:
:slightly_smiling_face: - clear explaination on each key steps :vibration_mode: :feather: :seal:
:slightly_frowning_face: - no instruction on feature extraction :moyai:
:no_mouth: - you only need to run `main.py` for real? :repeat:
### Clear versioning
:crown: - thank you for using the release feature :vibration_mode:
:slightly_smiling_face: - at least the version control is there :feather: :seal:
:no_mouth: - Very few commits :moyai: :repeat:
### Special mention: Pretrained model download
:moyai: built into feature extraction code
:vibration_mode: built into feature extraction code
:repeat: built into feature extraction code
:feather: Should pre downloaded - no explicit instruction, but the procedure is standard
:seal: built into feature extraction code
### Discussions: Key mentions from the Insight paper
ðŋ Modality dropout during training
ðŋ Validation performance per model per brain parcel as softmax weights
ðģ Trained the model without a causal mask
ðģ Explicitly modeling the HRF
ð Extracted features on a strict TR-by-TR basis
ð A brain-inspired curriculum, early sensory regions -> higher-order association areas
ðŠķ A shared group head plus subjectspecific residual heads
ð Contrastive learning training objectives
### License
:seal: No license!!!!!