Can a Language Model Learn Facts Continually in Its Weights?

Why this sits on harnesses

The paper's conclusion is a harness conclusion: if weights are an unreliable place to put new facts and context is the reliable one, then memory is the harness's job, not the model's. It is the empirical backing for "file system as persistent memory" as a design trend rather than a convenience.

⚠️ Correction to the referring source

This entered the KB via a bookmark quoting tweet 13/ of the author's thread, which said:

"all the training to create the model is engineered around crafting the best possible ICL mechanism. Further training degrades this mechanism, at least for knowledge acquisition"

The paper's own wording is weaker and partly opposite:

"The same writes barely degrade the model's use of facts in context."

The two can be reconciled (degradation for knowledge acquisition vs preserved in-context use), but they are not the same claim. Cite the paper, not the thread. This is the second case in this batch where a thread overstated its own paper.

Abstract

Can a language model accumulate facts in its weights through continual learning after training? The finding: facts can be stored in weights, but they remain question-keyed — later writes redirect queries to the newer facts without preserving the earlier ones. Reliable retention across sequential updates requires context, not weight modification.

Setup

• Models: Qwen3 variants
• Regime: 20–100 sequential weight writes
• Arms compared:
  • bare-statement training (a fact presented once)
  • diverse-restatement training (many phrasings of the same fact)
  • reference baseline: the original model with facts supplied in the prompt
• Evaluation: five types of held-out questions, testing both recitation and behavioural use

Results

Finding	Number
Recitation-to-use gap, bare statement → diverse restatements	27.4 → 5.4 points
Retention after 20 sequential writes — bare-statement facts	1%
Retention after 20 sequential writes — facts from broad study data	46%
Wrong answers that contain the most recently written fact	70%
A forgotten study fact re-supplied in the prompt	recovers to 77–80%

Two of these carry most of the weight:

1% vs 46% after twenty writes. How a fact is written matters more than that it was written. Bare statements are effectively erased by subsequent updates.

70% of wrong answers contain the most recent fact. This is the mechanism, not just the symptom: the model is not forgetting into noise, it is being redirected. Later writes capture the query path. That is why the authors describe stored facts as "question-keyed."

77–80% recovery from the prompt is the harness-relevant number: the knowledge was not destroyed, only made unreachable, and context restores access.

Stated limitations

• "No intervention we tested, including those built on accurate local measurements of each write, keeps earlier facts reachable." — a notably strong negative result, and the authors state it plainly rather than burying it.
• Weight-based knowledge is not robust to interference from later training.
• Forgotten facts retain log-probability but become inaccessible to normal querying.

Not reported

Despite the referring thread's emphasis on multi-hop failure — the "should I hire a Waterloo or an MIT intern" example — the paper reports no separate multi-hop or second-order metric. Anyone citing multi-hop degradation from this source is citing the thread, not the research.

Evidence grade

Moderate-to-strong, first-party, single model family. Vendor-published (Baseten) and self-measured, on Qwen3 variants only — scale and family generality are untested here. The negative result is stated cleanly and the numbers are specific, which is more than most vendor research offers. Treat the 1%/46% split and the 77–80% context recovery as the citable figures; treat anything about multi-hop as unsupported by this document.