AI TrackerFrontier desk · AEST
Back to news
normalaitech

WeightBench: Evaluating Language Models as Weight-Space Programmers

arXiv cs.AI · 14 Nov 2026, 04:58 AEST

Recent research indicates that large language models can manipulate the parameters of other neural networks without gradient descent. The introduction of WEIGHTBENCH provides a benchmark to evaluate these capabilities across various tasks.

Key facts

Read aloud

Recent research indicates that large language models can manipulate the parameters of other neural networks without gradient descent. The introduction of WEIGHTBENCH provides a benchmark to evaluate these capabilities across various tasks. Recent work shows that large language models L L M s can operate directly on the parameters of other neural networks. Exploring this new paradigm systematically requires a robust benchmark. We introduce WEIGHTBENCH, a benchmark that evaluates L L M s of increasing capability on forty two WSP tasks across five tracks: reading, fixing and writing weights, recognizing edits that are impossible, and data-free compression. Experimental benchmark results show that frontier agents can already read small networks and write working control policies without ever running them.

Sources in this cluster