NLA Maintains Accuracy but Confabulates More
Turn_Trout · x · 2026-07-13
The author proposes an explanation: Claude's guesses might leak some "real stuff," such as the identity of the last token; these rare clues are particularly critical for reconstruction tasks.
The thread also mentions a key finding: an NLA initialized with confabulation achieves almost the same reconstruction accuracy as the control group, but massively increases the confabulation rate, reaching 99.3%. In other words, training can reduce confabulation to some extent, but it's not enough to offset the bias introduced by this initialization; meanwhile, the control group actually learns to confabulate more.
More from Research
- rauchg reproduces OpenAI's covert CPU-channel finding: chat across Firecracker microVMs with no network — ccerrato147 · 2026-10-27
- Microsoft Research: ForkPilot cuts long-horizon agent compute by up to 59% — EliasEskin · 2026-10-27
- Microsoft's TeleTune evolves agent skills from offline user telemetry — EliasEskin · 2026-10-27
- FlexGlue: sparse-to-dense matching that beats dense accuracy at a fraction of the cost — zhenjun_zhao · 2026-10-27
- Mundo turns a few room photos into an explorable, simulatable 3D world — at 1.5 FPS — mark_k · 2026-10-27
- Mathematics as an experimental science: Marc Lelarge to talk LLM-driven math experiments at ENS — marc_lelarge · 2026-10-27