WhatWorkedBench is a benchmark designed to evaluate how well AI agents understand the impact of experimental changes. It measures predictions about component changes after budgeted experimentation, using a response surface that predicts scores for different configuration settings. The benchmark includes data from 36 tasks, 30 data sources, and 8 workflow types, with 1,248 configuration records. Evaluation involves 4,206 numerical-control records and 108 agent episodes, showing improvements in effect recovery when using a Gaussian process (GP) model.
The chart will appear after repeat observations. The current metric comes from the source.
6 source votesReal observations only. History before source connection is not reconstructed.
This benchmark helps researchers evaluate how well AI agents can predict the outcomes of experimental changes, improving their ability to design and interpret experiments effectively.
IntBMoE introduces a block-conditioned Mixture-of-Experts (MoE) framework that decouples participation, execution, and materialization by combining dense expert composition with sparse block execution. It uses a learned codebook for blocks and a hypernetwork to merge experts, achieving full participation while keeping execution and memory costs low. Experiments show improvements in image classification, language modeling, and sequential recommendation, with real-world deployment in AMap's recommendation system.
6 days agoDaily PapersThe paper introduces Taste-Bench, a benchmark for evaluating an agent's 'taste' in making long-horizon decisions. It measures the ability to choose the better path in decision forks without knowing future outcomes, revealing that current models perform poorly and that taste can be improved through distillation.
2 days agoDaily PapersThe paper introduces Document Retrieval-Aware Chunking (D-RAC), a method for efficiently processing enterprise documents by normalizing them into PDF and converting them into retrieval-optimized Markdown using a multimodal LLM. This approach preserves document structure, reduces token costs, and improves chunking efficiency compared to traditional methods.
3 days agoDaily PapersThis paper introduces an all-in-one multilingual scene text recognition system that uses a script-aware Mixture-of-Experts (MoE) architecture. The system, called ScriptMoE, outperforms existing methods on multilingual text recognition tasks by leveraging a large synthetic dataset and a lightweight architecture that shares a visual encoder while using script-specific experts for decoding.
3 days agoDaily PapersCodeMidas is an agentic pipeline that converts implemented functionality in open-source codebases into executable reinforcement learning (RL) environments using only source code as input. It generates 5,545 training tasks from 3,185 codebases across 23 languages and 15 domains, improving performance on multiple coding benchmarks when training MiMo-V2.5 with GRPO.
6 days agoLessWrongThe post explores the author's confusion about how neural networks, particularly large language models (LLMs), perform and learn computations. It challenges the conventional 'mechinterp' (mechanistic interpretation) approach that relies on circuit-based explanations, arguing that 'circuits' and 'computation' may not be the most suitable framework for understanding LLMs. The essay discusses representational drift as a key obstacle to weight-based circuit analysis and suggests a 'co-selectionist' view of circuits as emergent units. The author also reflects on how the concept of 'universality' should influence explanations of LLM function.
2 days ago