Results (Qwen3-1.7B):
✅ Math reasoning (GSM8K): 81.2% → 86.7% with ProbL2 (+5.5 points)
✅ Code generation (MBPP pass@1): up to 60.8% with neural mirrors
✅ Zero-shot transfer (HumanEval): 53.5% → 62.1% pass@1 (+8.6 points) with ProbL2, NM-GRPO-ES reaches 60.7% using ~30% fewer tokens
✅ Less verbosity: 24–36% shorter generations (big deal for deployment)
✅ More reliable training: variance drops from ±0.7 → ±0.4 (math) and down to ±0.2 with ES (code)
Why this matters:
If you’re running post-training in production, you don’t just want “best run” performance, you want:
➡️ Predictable training
➡️ Better reasoning
➡️ Lower inference cost (fewer tokens)
➡️ And a knob that adapts to task structure (math ≠ code ≠ dialogue)
Conclusion:
Better reasoning doesn’t always need more reward tricks. Sometimes it just needs the right geometry.
Kudos to the team who worked on this - @RuiYuan11926485@McHandoga@SVinayKumar12
#PolicyOptimization #RL #LLMResoning #LexsiLabsParis #LexsiLabs
@realTrumpNewsX Ukrainian people survives thanks to the military help from the US. One can't destroy tanks and enemy soldiers who come to take his land with just hopes and prayers.
Elon Musk amplifies Russian propaganda about Crimea, Ukraine.
Here's how these lies sound to Ukrainians: "Alaska is seen as a core part of Russia by Russians. Alaska is of critical national security importance to Russia. It's only in the USA because of Tsar Alexander's mistake."
@elonmusk It should be decided by the entire country. You can't cut it into pieces - it is in the Constitution.
And who are the "people who live"? Those who had to flee or so were killed during the occupation?
In 1930s, the Soviets allowed Wehrmacht to use training centers in Kazan (tanks) and Lipetsk (Luftwaffe). This was a prelude to the WW2. It looks like since 2014, German company (and the state?) allowed Russians to use German technology to prepare for what can turn into the WW3.
"I held his hand until I realized it was over and he was starting to get cold." Careful reconstruction of murders and rapes committed by Russian soldiers in Bohdanivka, Ukraine.
https://t.co/3cttoYPCfZ