We provide an implementation to test spurious rewards out on Qwen-2.5 1.5B. Check it out here!
@askalphaxiv
-
Qwen excels at RL finetuning with random rewards
By
–
Qwen is hands-down the best model for RL finetuning Turns out you can even tune it with random rewards and it *still* improves. Paper shows it's not magic, pre-trained code reasoning patterns are uncovered with RLVR We provide hassle-free bash scripts to reproduce this below!
-
Alibaba’s Learnable RL Policy for LLM Context Management
By
–
This new paper from Alibaba Group makes context management into a learnable RL policy So the agent is capable of deciding when to store, update, or delete long-term info and when to retrieve, summarize, and filter short-term context bringing a fresh new way to tackle LLMs used
-
Imaginary RoPE Implementation Explored for Language Models
By
–
Check out the imaginary RoPE implementation here!
-
RoPE++ Improves Long-Range Dependencies in Language Models
By
–
The "imaginary" part of RoPE contains valuable information for long-range dependencies RoPE++ reintroduces it in heads with slower decay curves that better capture distant tokens We provide implementations with one-command scripts for both the original codebase and Nanochat!
-
NVIDIA’s Alpamayo-R1: Open Weights 4D Understanding Autopilot Model
By
–
As NVIDIA pushed its first ever open weights autopilot model alpamayo-R1 A key frontier is native 4D understanding: not just “describe the video” but reason about depth + motion + 3D interactions over time, even for a specific region. This paper introduces 4D-RGPT (also from
-
Evaluating LLM Legal Reasoning with LEXam Benchmark
By
–
How good are LLMs at legal reasoning? This Friday, join us for our AI for Law session led by @joelniklaus on LEXam — a new benchmark using real law school exams to evaluate long-form, structured legal reasoning. Link to register below
-
Causal prediction alone enables strong visual learning
By
–
A new mind-blowing finding:
causal prediction alone is sufficient for strong visual learning, no need for any fancy reconstruction, masking, or contrastive loss In this paper NEPA, instead of reconstructing pixels, they train a vision model to autoregressively predict the next -
Meta’s Query-Only Test-Time Training Improves Long Context Performance
By
–
This paper by Meta introduces Query-Only Test-Time Training for long contexts So by using a small amount of inference-time training that retunes how the model attends to the given input, it works far better than generating extra reasoning tokens
-
SonicMoE: Tri Dao Advances MoE Kernel Efficiency on H100s
By
–
The author of FlashAttention, Tri Dao, just dropped a new paper called SonicMoE With 1.86x higher MoE kernel throughput and 45% lower activation memory per layer on H100s, by introduceing tile-aware routing that cuts padding waste for sparse MoEs Trending on alphaXiv