Goldberg sparks debate: interpretability should aim at understanding mechanisms

From Aug 26–27, prominent NLP scholar Yoav Goldberg posted a series of views on the purpose of interpretability research, sparring with other researchers over whether the field should aim at understanding model mechanisms or be judged by downstream task performance.

Confirmed

Goldberg argued the core goal is understanding mechanisms; inventing new methods is a side effect, and progress that aids understanding is real progress.
He claimed that research targeting steering and judged by steering effectiveness is no longer interpretability research but steering research with useless shackles, and provocatively called "actionable interpretability" an inferior form.
On Aug 27 he added that interpretability research need not have benchmarks, tasks, or evaluations—only rigor and evidence-backed claims.
He noted some findings can be valuable without being causal (e.g., highly redundant knowledge representation in weights), while cautioning that knowledge-editing experiments can mislead: editing a weight location and observing behavior change does not prove knowledge is stored there, since the intervention may hit the last link of a pathway.
@EEnouen countered that interpretability should be optimized for downstream goals like steering and probing; methods that fail to beat simple baselines at predicting, intervening in, or steering behavior may be "false progress."
@aryaman2020 added that practical interpretability's flaw is that it targets the same problems as post-training researchers, who face no constraint to use interpretability methods.

Why it matters

The dispute crystallizes a route divide in the field: as steering/manipulation becomes a hot application, whether engineering-oriented evaluation erodes deep mechanistic understanding or provides a reality check against false progress will shape method choices and evaluation standards.

2026-08-26 ~ 2026-08-27 · 9 related posts

Primary sources

Yoav Goldberg: Interpretability doesn't necessarily need baselines or tasks — yoavgo · 2026-08-27 03:03 (UTC+08:00)
Yoav Goldberg: Interpretability core is understanding mechanisms, not just methods — yoavgo · 2026-08-26 22:59 (UTC+08:00)
Interpretability methods should hill-climb on downstream goals, not intrinsic metrics — EEnouen · 2026-08-26 21:09 (UTC+08:00)

Yoav Goldberg: actionable interpretability research is the lesser form — yoavgo · 2026-08-26
"Actionable interpretability" argued to be an inferior form of research — aryaman2020 · 2026-08-26
Interpretability is "false progress" if it cannot improve intervention capabilities — EEnouen · 2026-08-26
[source] Interpretability methods should hill-climb on downstream goals, not intrinsic metrics — EEnouen · 2026-08-26
Yoav Goldberg: interpretability that serves steering is just steering research with a handicap — yoavgo · 2026-08-26
Yoav Goldberg: Interpretability Core is Understanding, Not Just Methods — yoavgo · 2026-08-26
[source] Yoav Goldberg: Interpretability core is understanding mechanisms, not just methods — yoavgo · 2026-08-26
Knowledge editing experiments may mislead interpretability conclusions — yoavgo · 2026-08-26
[source] Yoav Goldberg: Interpretability doesn't necessarily need baselines or tasks — yoavgo · 2026-08-27