←
alphaXiv3,345 posts
alphaXiv avatar✓

alphaXiv ✓

@askalphaxiv

High fidelity research

◷ Joined November 2023▧ 1,493 Photos and videos
alphaXiv avatar
alphaXiv@askalphaxiv·4h

"On Numerical Issues of Divergences Between Nearly Equal Distributions" Even when two LLMs produce almost identical predictions, standard KL divergence calculations can produce inaccurate gradients because floating-point rounding loses tiny differences. This paper identifies 5 numerical issues across KL divergence, Jensen-Shannon divergence, and entropy calculations, then proposes 8 more numerically stable formulations to preserve those differences. Their results show they can recover small gradients that standard implementations can incorrectly round to zero, although large-scale training benefits is still untested. alphaxiv.org/abs/2611.18511

Post attachment
alphaXiv avatar
alphaXiv@askalphaxiv·6h

"Perturbing Inputs Improves Pretraining Distillation" Knowledge distillation normally trains a smaller LLM to copy a teacher's predictions on regular text, but this paper finds that deliberately corrupting the inputs can actually improve learning. This paper trains students to match their teachers on perturbed inputs, such as blending neighboring tokens together, before switching back to normal distillation. Across 135M and 0.5B models, this reduces validation loss by 0.025 and 0.020 nats, respectively, compared to standard distillation. At 0.5B, it also improves average zero-shot accuracy by 1% across 7 tasks. alphaxiv.org/abs/2611.11598

Post attachment
Yijia Shao avatar
⟳ alphaXiv retweeted
Yijia Shao@EchoShao8899·Nov 16

AI is moving from on-demand to always-on. But can it learn when you actually want help? We introduce Deep Personalization: an Observer Agent that continually learns from your work patterns and implicit feedback to decide when to step in. It improves proactive-timing F1 by 2.4–3.0× across three model backbones. We co-design the agent and its training and inference engines to run the Observer locally on systems like NVIDIA DGX Spark, keeping your screen observations on-device. The next PC won’t wait for your prompt. It will learn when to help! Check out our blog at: nvlabs.github.io/deep-person…

0:59
alphaXiv avatar
alphaXiv@askalphaxiv·18h

Is GPT-6.2 Astra now the best model for scientists? The new model from OpenAI is SOTA on Terminal-Bench-Science, surpassing Fable 5.2 and Mythos OpenAI reports 77.1% at $7.15/task on Terminal-Bench-Science 0.1, beating Fable 5.2 on cost and performance Check out our new benchmark aggregation tool for more info, updated hourly 👇

Post attachment
alphaXiv avatar
alphaXiv@askalphaxiv·19h

Qwen3.5-0.8B & MATH-500: Astra's edit removes a "looping" direction from 64 neurons in each of the two MLP layers, and Inkling pruned it by keeping only the 4,096 largest-magnitude changes in each layer. Those 8k entries match the full edit's accuracy (58.2% vs 58.6%) and editing only one layer (layer 15) recovers most of the gain.

0:09
alphaXiv avatar
alphaXiv@askalphaxiv·19h

Broken car policy: Inkling found that almost all of Astra's 6,073 changes just zero out old weight to free up neurons, and only about 55 do the actual work. It then wrote a 10-weight fix from scratch that teaches the car a move PPO never discovered: back up until it stalls against the wall, then launch, triggered by a velocity-detecting neuron the network already had.

0:09
alphaXiv avatar
alphaXiv@askalphaxiv·19h

You don't need many edits in weight space to fix a model In the WSP blog, GPT-6 Astra fixed a broken car policy by changing 6k weights and cut Qwen3.5-0.8B's math looping by changing 130k Inkling 2 Small did the same with just 10 and 8k edits: ~600x and ~16x fewer, with no degradation in performance on either task

3:08
alphaXiv avatar
alphaXiv@askalphaxiv·Nov 17

"Classical Energies Learn to Reason 1,000× Faster than Looped Models" Looped reasoning models get high accuracy on Sudoku and maze benchmarks, but these tasks may actually be solvable with much simpler approaches. This paper introduce CRAM, an energy-based model that learns to assign lower scores to valid solutions, then solves puzzles by minimizing that score. CRAM solves 99.996% of Sudoku-Extreme and 100% of Maze-Unique, training at least 1,000x faster than looped models while requiring 282x fewer inference FLOPs at comparable Sudoku accuracy. However, these benchmarks have relatively simple underlying rules, so how well they measure general reasoning capabilities is still a bit unclear. alphaxiv.org/abs/2611.17309

Post attachment
alphaXiv avatar
alphaXiv@askalphaxiv·Nov 17

"Interpretable Weight-Space Programming Builds Transparent Transformers and Brain Encoding Models" This Microsoft paper also explores whether LLMs can directly write those weights to build working models from scratch. But instead of just maximizing accuracy, they introduce Interpretable Weight-Space Programming (iWSP), which requires LLM agents to design understandable mathematical operations, test their performance, and refine the model without backpropagation. Using Claude Opus 5.5, iWSP builds small Transformers achieving 100% accuracy across 13 algorithmic tasks, with a median of just 166 parameters. It also discovers interpretable features that outperform 7 existing interpretable baselines at predicting visual-cortex activity. alphaxiv.org/abs/2611.interp…

Post attachment
Hyunjong Ok avatar
⟳ alphaXiv retweeted
Hyunjong Ok@HyunjongOk·Nov 17

🚀 Stop using Transformers for weight-space programming If you are not training with gradients on a GPU, why default to a Transformer? When an LLM writes network weights directly, letting it also design the architecture achieves: 📉 ~1,800× fewer parameters at matched accuracy on ARC (AI2 Reasoning Challenge), compared to the best Transformer the same LLM wrote. ⚡ Up to 17× faster GPU inference with a memory architecture built for these tables, compared to a Transformer of the same size. 🧵 Paper & code at the end 👇

Post attachment
alphaXiv avatar
⟳ alphaXiv retweeted
alphaXiv@askalphaxiv·Nov 16

"Weight-Space Reading" Can an LLM predict how changing another model's weights will affect its behavior, without ever running that model? This paper investigates Weight-Space Reading, where LLMs inspect the raw weights of 9B-30B parameter models using Python tools, then predict how small weight modifications will affect capabilities like math accuracy, reasoning length, and refusal behavior, without performing a single forward pass. On subtle weight edits, Qwen3.8-27B achieves 73% balanced accuracy, compared to 57.8% when given only text descriptions of the edits. However, the models also frequently predict changes that never happen. When 40 edits had no measurable effect on Muse-30B, the model incorrectly predicted changes for 85% of them. The researchers also found no conclusive evidence that LLMs are better at reading their own weights than those of other models. alphaxiv.org/abs/2611.weight…

Post attachment
Wenhao Chai avatar
⟳ alphaXiv retweeted
Wenhao Chai@wenhaocha1·Nov 17

Find it super easy submitting blogs to @askalphaxiv so I just do so. thanks for this great platform! (i am not saying the other one is not good How to Scale Text-to-image Models: a Handbook Link: alphaxiv.org/abs/2611.how-to… Also two of my previous blog: Larger Tokenizers Suffer More Multi-Epoch Regret alphaxiv.org/abs/2611.larger… Penalty Bench: Can an AI Agent Referee? alphaxiv.org/abs/2611.penalt…

0:23
alphaXiv avatar
alphaXiv@askalphaxiv·Nov 16

Check out this blog on why you should write research blogs Blogs allow you to easily share interesting research takes, publish what didn't work, and checkpoint progress without waiting on noisy review You can now publish blogs on alphaXiv directly!