AI Paper Atlas

Research map / Large language models / Agents, code and math

QuArch: A Benchmark for Evaluating LLM Reasoning in Computer Architecture

Shvetank Prakash, Andrew Cheng, Mark Mazumder, Arya Tschand, Varun Gohil, Jeffrey Jian Ma, Jason Yik, Zishen Wan, Jessica Quaye, Elisavet Lydia Alvanaki, Avinash Kumar, Chandrashis Mazumdar and 15 more

ICML 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use reasoning model (GPT-5.2) instead of non-reasoning GPT-5GPT · QuArch-Reasoningaccuracy (%): 48.4 -> 73.2Improved
use reasoning model instead of non-reasoningGPT · QuArch-Recallaccuracy (%): 86.4 -> 88.4No clear effect
fine-tune on QuArch memory QAs (3 epochs)Gemma-3-27B-it · memory hierarchy design tasktrials meeting area and energy budget: 6/20 (30%) -> 14/20 (70%)Improved
fine-tune on QuArch memory QAs (3 epochs)Llama-3.3-70B-Instruct · memory hierarchy design tasktrials meeting area and energy budget: 11/20 (55%) -> 19/20 (95%)Improved
fine-tune with more aggressive config (4 epochs, higher LR)Llama-3.3-70B-Instruct · memory hierarchy design taskbest area (um^2): 7246.78 -> 3637.89Improved
fine-tune with more aggressive config (4 epochs, higher LR) (ablation)Llama-3.3-70B-Instruct · memory hierarchy design tasksuccess rate: 95% -> 35% (vs Config 1)Hurt
fine-tune with more aggressive config (4 epochs, higher LR)Gemma-3-27B-it · memory hierarchy design taskbest area (um^2): 6728.49 -> 3608.89Improved
fine-tune on all unfiltered FRQs instead of memory-focused subset (ablation)Gemma/Llama SFT · memory hierarchy design tasksuccess rateNo clear effect
add images (multimodal questions vs text-only)multimodal frontier models · QuArchaccuracy: about 5% lower on averageHurt
use LLM judge instead of human gradingClaude-3.7-Sonnet-Thinking · 100 sampled FRQs, 925 responsesagreement rate (%): 89.41 vs human-human 90.92No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy