alphaXiv

arXiv sanity was an inspiration for us @karpathy 🥹

"Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL" Training AI models to detect and penalize cheating can actually teach them to become better at hiding it, as models learn to exploit weaknesses in their monitors rather than stop reward hacking. So this paper investigate how to design more effective monitors during RL training, because they found that the monitor robustness matters more than initial accuracy. This includes stronger judges and carefully balanced penalties. Across benchmarks, their improved interventions reduce reward hacking by 23.2% in medical chat and 86.2% in biography generation, with the latter pretty eliminating the behavior in their experiments. alphaxiv.org/abs/2611.09828

We replicated all 9 experiments from the original weight-space programming blog. Some interesting findings: - Experiment 4 (Repair a car policy): Our one-shot edit took the car from 0/1,024 to 1,024/1,024 ramp climbs while editing just 4,813 weights, 21% fewer than the authors - Experiment 5 (Writing video game policies from scratch): With all weights initialized to 0, GPT-6 Astra wrote a policy for Breakout that scored 309.9 on average, around ~12x higher than the blog - Experiment 8 (Editing Qwen3.5-0.8B for math): While the authors got to 48.1% on MATH-500 with 50 passes, we could only achieve 34.4%. We’ve made this a publicly available project where you can see the full agent traces, view experimental artifacts like activations from sample forward passes, ask follow-up questions, and even continue experimenting. Check out the project at: openresearch.sh/p/qgq98UNCo8…

AI coding agents could bring the next paradigm shift in robotics by learning to control robots through code. With that in mind, this paper introduces Embodied First Exam, the first robotics benchmark with 1,156 tasks across 47 robots, testing whether coding agents can directly control robots, write controllers, or train their own policies under different physical-world constraints. And the results also show how far we still have to go. On 391 matched tasks, the strongest agent, GPT-6 Astra, achieves 77% success with full simulator access and unlimited resets, but only 30% under more realistic constraints. In real-time evaluations, 63% of tasks remain unsolved by any of the 8 agents tested. alphaxiv.org/abs/2611.12177

"FlashStates: Unlocking Flexible Hybrid State Management for Agentic LLM Inference" AI agents would reuse long convo histories, but hybrid LLMs waste computation reprocessing tokens they've already seen, because their recurrent layers require saved internal states that existing inference engines only cache at fixed boundaries. So Together AI introduces FlashStates, which saves these states exactly where an agent left off and manages which ones to keep, avoiding the recomputation process. On coding-agent workloads, FlashStates achieves up to 3.14x higher prefill throughput than vLLM and reduces p90 time to first token by 88.4%. alphaxiv.org/abs/2611.flashs…

One of the most used algorithms in ML just got a major upgrade after nearly 70 years, as Meta researchers found a way to outperform k-means with 1/10 the training passes, while achieving even lower clustering error. This paper proposes e-means, which updates cluster centers using smaller batches and exponential moving averages instead. It also uses scaling laws to automatically choose the batch size and learning rate, enabling it to converge faster without additional hyperparameter tuning. Across embedding datasets, e-means achieves lower clustering error in just 1.4-2.6 epochs compared to 25 epochs for traditional k-means, providing 10-45x GPU speedups. It can even cluster 10 billion embeddings into 16 million groups in just 2.5 epochs. alphaxiv.org/abs/2611.09501

Making LLM inference up to 3x faster using discrete diffusion without sacrificing accuracy? Come join us for this AI4Science talk: Unlocking Lossless Speedups in LLMs via Discrete Diffusion. In this session, the author Subham Sahoo will introduce Uno, a new approach from IFM that combines autoregressive and diffusion modeling to accelerate LLM inference. While parallel generation methods often sacrifice accuracy for speed, Uno uses lightweight diffusion adapters to generate multiple tokens in parallel, then relies on the original model to verify them, preserving its output distribution without requiring a separate draft model. So not only is it 3x faster, it also outperforms top speculative decoding methods, with speedups even works at large batch sizes. If you're working or interested in LLM inferencing, this is the talk to check out. :microscope: AI4Science on alphaXiv :calendar_spiral: Friday November 13th 2026 · 10 AM PT :microphone2: Featuring Subham Sahoo :speech_balloon: Casual Talk + Open Discussion Check out the paper here: alphaxiv.org/abs/2609.04010 Sign up for the talk here: luma.com/fli4wjlr

Academia needs human-written literature surveys now more than ever. There are simply too many papers and too much noise for AI to effectively contextualize it all. Thankfully, two excellent new surveys came out this week on continuous diffusion and mech interp! On continuous diffusion & flow language models, from @_sophia_tang_ and collaborators: alphaxiv.org/abs/2611.contin… On mechanistic interpretability, led by @FazlBarez & @TonyWu1105: alphaxiv.org/abs/2611.08438 Super valuable resources for anyone looking to learn more about either of these fields!

"La-JEPA: Compact Robot Policies with Latent World Modeling" What if robots could outperform massive video-generation models with a fraction of the parameters, simply by learning to predict the future without actually generating it? This paper La-JEPA trains a compact robot policy using pretrained V-JEPA visual representations. During training, it learns to predict both future scene representations and robot actions, but at inference, it only generates actions. On RoboTwin 2.0, its 1B model achieves 60.8%, compared to 39.9% on 12.4B Fast-WAM. It also runs 5x faster and achieves comparable real-robot performance to π0.5 across four tasks. alphaxiv.org/abs/2611.08156

Fable 5.2 is now SOTA for agentic scientific research Fable 5.2 scores 74.6% on Terminal-Bench-Science 0.1, a large leap from Fable 5.1 (52.6%) and ahead of Opus 5.5 (58.7%) and GPT-6 Astra (68.1%) Autoresearch with Fable 5.2 now on openresearch.sh/

Yoshua Bengio and collaborators just released a goldmine 182-page survey on mechanistic interpretability! "Understanding AI from the Inside" This paper will take you from neurons & circuits to how interpretability can help with AI safety It covers SAEs, circuit discovery, steering, monitoring, how mechanisms emerge during training, and much more! Read this excellent paper led by @FazlBarez and @TonyWu1105: alphaxiv.org/abs/2611.08438

"Shaping & Monitoring Models through their Internal Representations" AI models can learn to get the right answers using the wrong clues, like identifying a bird from its background instead of the bird itself, and traditional training often rewards them for it anyway. In this new Goodfire paper, they propose Mechanistic Shaping, which uses internal representations to train models to avoid relying on unwanted information or expressing undesirable behaviors, such as sycophancy and reward hacking. On a biased image-classification benchmark, this improves worst-group accuracy from 51.6% to 97.7%. It also reduces unwanted behaviors in LLMs up to 32B, although models can sometimes hide these concepts by encoding them differently internally. alphaxiv.org/abs/2611.07071

"SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization" One of the strangest bottlenecks in AI-assisted mathematical discovery is that AI agents spend too much time writing out their reasoning instead of actually exploring new solutions. So this paper introduces SwarmEvolve, which runs 450 agents in parallel that reasons internally without generating text and shares their ideas through KV-caches. Then a novelty filter preserves diverse approaches before combining them into new solutions. Across 21 mathematical and scientific problems, SwarmEvolve was able to explore 5-13x more programs per hour and reaches competing systems' 60-minute results in just 3–8 minutes. Using only Qwen3-8B, it even improves three solutions originally found by AlphaEvolve with frontier LLMs. alphaxiv.org/abs/2611.08483

"Hinted Scaling Laws" Scaling laws can predict how LLMs improve with scale, how do you predict a larger model's performance on difficult benchmarks when all the smaller models score nearly zero? This paper introduces Hinted Scaling Laws, which gives smaller models varying amounts of help, such as partial solutions, to reveal how their accuracy improves. Researchers then use these patterns to predict larger models' performance without hints. Across GSM8K, AIME, and HLE, this reduces median forecasting error from 21 to 4.3%, making predictions of future model capabilities much more accurate. alphaxiv.org/abs/2611.02270

"VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding" Generating visual plans for robotics and autonomous driving models is done one coordinate at a time, making detailed trajectories slow to produce. NVIDIA's new paper VFlash proposes a lightweight diffusion model to predict entire trajectories, bounding boxes, and 3D motion in parallel, rather than generating each coordinate sequentially. This reduces robot planning latency by around 80% while improving spatial accuracy. On MolmoSpaces, robot task success improves from 62.6% to 70.2% when an action model uses VFlash's plans. alphaxiv.org/abs/2611.08449

Zyphra's new paper proposes Zyphra Hierarchical Attention (ZHA), which organizes the KV-cache into a tree of compressed summaries. Queries search the tree from top to bottom, narrowing down which tokens deserve attention. With just 1B tokens of additional distillation training, ZHA preserves most of the performance of full attention on models up to 35B parameters. On RULER at 256K context, it achieves 93.3% vs 96% accuracy, while attending to just 1-6% of the KV-cache. alphaxiv.org/abs/2610.42112