OpenResearch was the #1 trending GitHub repo on Friday 🚀 With OpenResearch, you can turn any coding agent into a research agent that reviews literature, develops hypotheses, runs experiments, and produces research artifacts. Own and automate your research stack end-to-end. Now available on Windows. Check it out: github.com/alphaXiv/OpenRese…

New from the lab: @LorenaYannnnn finds that reward misspecification during RL causes unmonitorable chain-of-thought. Models talk about the goals in their prompt, even as they learn throughout RL to optimize the misspecified reward. We also propose an intuitive fix; see below.

"Understanding LLM Parameter Update Sparsity through the Lens of Fisher" LLM post-training methods like RL and on-policy distillation often produce highly sparse parameter updates, but why? This paper uses the diagonal Fisher information to explain this: parameters that barely affect the model's output probabilities tend to receive smaller gradients. It also shows that Fisher can identify these parameters before training begins. On Qwen3-1.7B, updating just 0.1% of parameters selected by Fisher recovers 85.2% of the improvement from full RL training, while updating 5% recovers 98.8%. alphaxiv.org/abs/2609.36262

"Depth Memory" Attention Residuals stores earlier Transformer layer outputs for retrieval but this memory grows with depth. Linear Depth Memory (LDM) maintains a fixed-size memory across layers, using recursive least squares to store and retrieve earlier layer information alongside the standard residual stream. On 400M and 1.5B models, LDM achieves lower perplexity than standard residuals, mHC, and AttnRes, while maintaining competitive zero-shot accuracy and comparable throughput with optimized kernels. alphaxiv.org/abs/2610.18809

Jeśli robisz research, eksperymenty z agentami albo chcesz automatyzować cykl „hipoteza > kod > eksperyment > analiza”, to naprawdę solidne i przydatne narzędzie. Warto rzucić okiem. openresearch.sh/

"Multi-Stream Attention Residuals" An alternative to mHC and AttentionRes? Standard residual connections combine information from all previous Transformer layers into one hidden state, making earlier information harder to access as models get deeper. Multi-Stream Attention Residuals (MSAR) lets each Transformer layer retrieve information from earlier layers using multiple attention heads, with each head looking at a different range of previous tokens to decide which information to retrieve. MSAR outperforms standard residuals, mHC, and AttnRes across all 12 benchmarks, achieving a 1.58x iso-loss training FLOP efficiency improvement over standard residuals based on fitted scaling laws, while adding just 0.5% training FLOPs per step. alphaxiv.org/abs/2610.13630

"Training Language Models to Fail Gracefully" Due to how we use RL for LLMs, there is currently little incentive for models to admit uncertainty, preventing them to guess even when they're unlikely to succeed. This paper GracefulRL instead trains models to prefer success over safe fallback, and safe fallback over failure. They did it by using a modified RL objective that prevents the model from collapsing into always refusing. On AIME, it reduces incorrect answers from 79.1% to 8.9%, while sacrificing only 3.1% of correct answers. On Minesweeper, it even improves success from 67.4% to 77.4% while reducing errors. alphaxiv.org/abs/2610.20401

"Scaling Discovery Requires Forgetting" Long-running AI research agents plateau as their context accumulates. It also becomes more conservative and make smaller changes instead of exploring new ideas. So this paper proposes forgetting, where you have the model periodically discarding all agent sessions and experiment histories, keeping only the best solution found so far. A fresh team of agents then resumes exploring from that solution. Across four coding and optimization tasks, this improves performance by 10% on average over parallel agents, and 22% over a single agent at the same budget. alphaxiv.org/abs/2610.20563

Looking for a fun weekend read? Introducing the Illustrated NeurIPS 🌎 We indexed all 9000+ NeurIPS papers and built a visual way to explore the whole landscape Search your favorite topics, open a cluster, and dive right in Check it out here! alphaxiv.org/conferences/neu…

1/ RL for improving task success has made LLMs remarkably capable. But what happens when the model fails? Turns out, training for graceful failure can be surprisingly tricky! Very excited to share our new paper: Training Language Models to Fail Gracefully 🧵

New paper advised by Li Fei-Fei! "Generative Representation Learning for Scalable Latent Tokenization" Image generation commonly uses pretrained encoders like DINO as tokenizers, but these can discard visual details important for generation. So this paper trains encoders to preserve information shared between masked image views, using a generative model to guide learning. On ImageNet, it improves FID from 2.36 to 2.18 at similar total training compute, with scaling-law extrapolations suggesting 7–8x less tokenizer training compute than DINOv2 and WebSSL. alphaxiv.org/abs/2610.20533

How can a collective of agents make meaningful progress toward new protein designs? Or close the loop from simulated alloy to printed part in a weekend? We’re excited to be providing tools and judging at the upcoming ScienceClaw hackathon, hosted by @medialab and @e14fund. Oct 30 – Nov 1 at MIT. Every team gets access to OpenResearch, and the best use of it wins $500 in GPU credits! Early apps close tomorrow (10/16) 👇 scienceclawhack.ai/apply.htm…

"LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators" Scientific flow models can generate valid molecules and peptides, but struggle to follow high-level natural-language goals. On top of that, current methods also rely on explicit reward functions, which are unreliable in scientific discovery. LLM-to-Flow Distillation (LFD) instead uses an LLM to create progressively harder sub-goals based on what the generator can currently produce, then uses its preferences to fine-tune the generator through DPO. On peptide and molecular design tasks, it outperforms reward-guided baselines on held-out computational evaluations, even reaching difficult out-of-distribution goals where those methods fail. alphaxiv.org/abs/2610.17246

Where is research attention moving according to arXiv papers? We mapped 10,000 of the most-read papers on alphaXiv, weighted by readership, to see how interest shifted from 2025 to 2026. Here’s what we found: 1) Traditional RL post-training is getting less attention. As the focus moves beyond math and reasoning to longer agent tasks, compute demands put this research out of reach for many teams 2) Harnesses, agent systems, memory and distillation are gaining reading share, offering other ways to improve what models can do 3) Interest is also growing in robotics and world models, bringing AI into the physical world

"Hybrid Reasoning Agents" Tool use is communicate through text, while latent reasoning operates on continuous hidden vectors, making latent reasoning unable to support tool use. HybridAgent combines both by switching to text mode when it needs to invoke tools, then returning to latent reasoning afterward. It also preserves the KV cache across turns so previous latent states aren't lost. This results in 10.4% better performance than matched text and latent baselines, with 35-55% fewer generation steps, 1.8-2.6x faster inference, and 2.7-3.7x lower costs. alphaxiv.org/abs/2610.18861