Cambrian-S is a valuable first step in defining what “supersensing” might mean for video models. Our results simply highlight how subtle benchmark design choices can be exploited — and how we can improve them together.

📄 arxiv.org/abs/2511.16655
🔗 github.com/bethgelab/s...

This indicates that the tailored Cambrian-S inference strategy may rely on benchmark-specific shortcuts (e.g. rooms are never revisited), rather than building a persistent, spatial world model over time.

For VSI-Super-Counting (VSC), we run a sanity check:

🔁 VSC-Repeat: we concatenate each video with itself 1-5×
✅ Unique object count stays the same
❌ Cambrian-S accuracy drops from 42% → 0%

A genuine supersensing system should be robust here.

We introduce a simple baseline called NoSense, an image-only (SigLIP) model that discards almost all temporal structure.

Surprisingly, it reaches 95% accuracy on VSI-Super-Recall (VSR), even on 4-hour videos.

This suggests VSR can be solved without true spatial supersensing.

🖐️

7/ Takeaway?

Many supposed gains don’t hold up under scrutiny.
Progress is possible—but let’s build on reproducible foundations.

🧠 Full paper: arxiv.org/abs/2504.07086

🧑‍🔬 By: @hrdkbhatnagar.bsky.social @vishaalurao.bsky.social @samuelalbanie.bsky.social @bayesiankitten.bsky.social @MatthiasBethge

6/ Our recommendations: – Evaluate with ≥10 seeds

– Tune decoding per model
– Use appropriate prompts/templates
– Standardize hardware/software (we use Docker)
– Open-source everything

📦 Code, prompts, outputs: github.com/bethgelab/so...

5/ What actually works?
🔹 RL methods over distillations? Often negligible gains, prone to overfitting.

🔹 Supervised finetuning (SFT) on reasoning traces? Stable & generalizable.

4/ Variance is everywhere:

– Random seed: swings Pass@1 by 5–15pp
– Temperature/top-p: another ±10pp
– Software & Hardware? Yes, even that changes scores

🎯 Single-seed results on small datasets are essentially noise.

3/ We re-evaluated recent 1.5B and 7B reasoning models on 6 benchmarks under controlled settings.

➡️ Performance dropped by up to 17%
➡️ Improvements fall within variance range of the base model
➡️ Some models don’t beat the baseline!

2/ Reasoning is the next frontier for LMs—but current evaluation practices often lack rigor.

We find that many celebrated gains from RL methods vanish once you:

✅ average over multiple seeds
✅ control decoding
✅ standardize prompt & infra

We are just getting started! We're building better filters, aggregating released benchmarks — datacomp style — and develop fast, accurate OpenThinking models. Stay tuned! w/
@hrdkbhatnagar.bsky.social, @vishaalurao.bsky.social, @bayesiankitten.bsky.social, Matthias Bethge [6/6]

These issues encourage shortcuts and flawed reasoning. If GRPO rewards bad logic, models reinforce errors instead of improving. Garbage In, Garbage Out 🚨 [5/6]

🔸 Some questions reference figures that aren't included! Text-only models can't infer missing visuals. [4/6]

🔸 Mathematical proofs are a challenge. There's no automated way to verify them, and answers often only show an initial equation, leading to unreliable training signals. [3/6]

Blog (For Updates): huggingface.co/datasets/bet...

🔸 Some questions contain subquestions, but only one answer is labeled. The model may get penalized for "wrong" but valid reasoning. [2/6]

This is joint work with @oripress.bsky.social, @vishaalurao.bsky.social, @bayesiankitten.bsky.social, @ofirpress.bsky.social and Matthias Bethge