KAN Author Liu Ziming Reveals LLM 'Forking' Overfitting

Episode 1 · KAN author Ziming Liu uncovers "forking" overfitting and warns against AutoResearch (2026-10-26, 7 posts)

Ziming Liu, MIT researcher and author of KAN, shared a research anecdote: a student came to him with a strange fork-shaped training curve, and after digging in, it turned out to come from output produced by an AutoResearch agent running automated experiments — training loss kept dropping while validation loss rose. Normally such "garbage output" from agents gets discarded, but this time the team decided to investigate the underlying mechanism, ultimately turning the failure mode into a new paper, "Forking: Sudden Overfitting Under Rep…," presented on day 2 of the MetaCircle AI Research Exposition.

Confirmed

• The forking phenomenon was triggered by erroneous generalization of a training trick found by an AutoResearch agent's automated search, occurring under settings such as over-encoding
• The team characterized it as: training loss declining while validation loss suddenly reverses and rises, forming a fork-shaped curve
• Ziming Liu presented the new forking paper on day 2 of the MetaCircle AI Research Exposition

Why it matters

• Ziming Liu uses this to critique the limits of AutoResearch (automated research pipelines): tricks found via search fail when experimental conditions shift slightly, even causing the overfitting catastrophe he calls forking; without mechanistic understanding, he argues, automated research is just blind-worship parameter search
• The case also demonstrates an alternative way to treat anomalous agent output: digging into "failed results" as research material can yield new discoveries

• Researchers identify "Forking": train loss falls while validation loss turns up — ZimingLiu11 · 2026-10-26
• MIT's Ziming Liu: AutoResearch doesn't generalize — blind training tricks cause 'forking' disasters — ZimingLiu11 · 2026-10-26
• Forking: AutoResearch's training tricks trigger sudden overfitting, new paper — ZimingLiu11 · 2026-10-27
• AutoResearch agent's weird training curve gets dug into instead of discarded — ZimingLiu11 · 2026-10-27
• AutoResearch agent produces weird 'forking' loss curves, and the team dug into the science — ZimingLiu11 · 2026-10-27
• Ziming Liu's "forking" phenomenon: over-encoding designs trigger sudden phase-change overfitting — ZimingLiu11 · 2026-10-27
• KAN author Ziming Liu warns AutoResearch generates chaos, not insight — ZimingLiu11 · 2026-10-27

Episode 2 · Explaining LLM Forking: Mechanism and Mitigation (2026-10-27, 2 posts)

Liu Ziming explains the LLM 'forking' mechanism, where overcoding lets an n-gram table hijack context signals and supply about three-quarters of the final residual stream norm, and notes it also emerges during post-training replay, mitigable by freezing the table.

• How forking happens: n-gram tables split logits into non-interacting subspaces — ZimingLiu11 · 2026-10-27
• Forking returns in post-training replay; freezing the n-gram table helps — ZimingLiu11 · 2026-10-27