
In RL, learner and sampler must agree on token log-probs—those numbers drive the importance ratio and the clip.
Native NVFP4 often breaks that: same weights, same "FP4" name, but different scale domains / kernels / reduction orders across engines → probabilities drift, then late-training collapse.
miniTE (w/ NVIDIA): diagnose the gap, align the numerical path (strict ~1e-8), keep W4A4 forward + BF16 backward.
~1.37× BF16 learner · 1.64× serving
infix.io/research/miniTE
🤔 How to solve train–inference mismatch in FP4 RL?
💡 InfiX with NVIDIA found the answer:
miniTE turns native NVFP4 into a practical low-precision RL training loop.
⚡ Up to 1.37× BF16 RL loop speedup
⚡ 1.64× BF16 serving throughput
👉 Read the report: infix.io/research/miniTE
3
219
Same image. Same image tokens. Same 2D positions.
Only the order the LLM reads them changes.
Qwen3.5-9B: 4 squares → 8 squares.
New blog: Qwen VLMs count by token order, not 2D position 🧵
ludicrous-apology-7fc.notion…
4
9
730
This may explain why earlier shuffle tests looked harmless: they used small models or caption-level checks.
If your method prunes, merges, or reorders image tokens, test counting separately, and on large models.
1
16
Ruled out: linear attention (Qwen2.5-VL drops too), split objects, per-token double counting, background leakage, DeepStack.
Part II: where is the order read?
Part III: can we make the model use its 2D positions?
Blog: ludicrous-apology-7fc.notion…
25
Looped Transformers should get better the more they loop. In practice, more loops often make them worse. We call this latent overthinking.
We found why, and how to fix it. Introducing InfiLoop 🧵
📄 arxiv.org/abs/2610.11570
Our fix: loop-native residual connections.
Instead of only the latest state, InfiLoop keeps a running summary of past states, weighted by content + learned decay—downweighting noisy updates while memory stays constant as loops grow.
With just 7M params:
• 97.9% on Sudoku-Extreme
• 13.6% pass@2 on ARC-AGI-2
• Still improving beyond 20,000 test-time loops
More loops → better reasoning.
💻 github.com/pixeli99/InfiLoop
w/ @pumpkinnnnne @luu_yinn @Shiwei_Liu66



