finding

active

finding:filler-gap-mechanism-in-pythia-1b-crosses-over-several-different-positions-before-arriving-at-output-position

Filler-gap mechanism in pythia-1b crosses over several different positions before arriving at output position

Mechanistic finding from CausalGym case study showing complex multi-step movement for filler-gap

Source paper

extracted_from

CausalGym: Benchmarking causal interpretability methods on linguistic tasks

(2024) · Aryaman Arora · Dan Jurafsky · Christopher Potts

Neighborhood — ranked by edge-count

Claims (1)

claim

For both NPI and filler-gap tasks, the model initially learns to move information directly from alternating token to output; intermediate steps are added later in training
supports
Mechanistic interpretation of training dynamics in case studies

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Filler-gap dependency mechanism in pythia-1b emerges in two discrete stages (steps 2000 and 10K) not graduallyfinding0.878
Training dynamics finding showing filler-gap takes longer to learn than NPI licensing
NPI mechanism in pythia-1b moves negation feature through complementiser 'that', auxiliary verb, and main verb across layers before predicting NPI 'any'finding0.762
Mechanistic finding from CausalGym case study showing multi-step information movement in NPI mechanism
NPI licensing mechanism in pythia-1b emerges in discrete stages (steps 1000, 2000, 3000) not graduallyfinding0.753
Training dynamics finding showing abrupt rather than gradual emergence of NPI mechanism
The mechanisms implementing NPI licensing and filler-gap dependencies are learned in discrete stages, not graduallyclaim0.739
Main mechanistic finding from case studies; evidence from training checkpoint analysis of pythia-1b
Gap junction blockade stochastically induces heads of other planarian species in genetically wild-type worms.finding0.731
Temporary disruption of gap junctions causes planaria to reconstruct heads appropriate to other species, revealing latent morphospace attractors.
Smaller fully trained Pythia models (31M, 70M) show slightly reduced alignment accuracy compared to larger models despite non-linear mapsfinding0.727
Attributed to model anisotropy from saturation making hidden states harder to access
Across 5 Pythia seeds, one seed fails to learn IOI task and another fails alignment despite learning the task; all other seeds achieve perfect alignment with ϕ_nonlinfinding0.722
Robustness check across seeds showing occasional failures of alignment map training
Filler-gap dependencyconcept0.721
Linguistic phenomenon where interrogatives extracted from a clause leave behind an empty gap; studied as case study in CausalGym