CSE Roadmap

Common tier · about 3 hours · 12 parts

Part 4 of 12 · 1 min read

Facts in weights go stale; retrieval keeps up

Baking a changing fact into the model's weights means it's wrong again the moment reality moves on, with no way to patch just that one fact.

Think of printing a phone directory versus keeping a searchable database: the directory is accurate the day it's printed and increasingly wrong every day after; the database is updated in place.

In plain words: when the underlying facts change over time, a model fine-tuned on a snapshot of them gets steadily more wrong unless it's retrained on a schedule. A retrieval system that looks facts up fresh each time doesn't degrade at all.

A worked example. With facts changing at 4 percent a month and no retraining, a tuned model's accuracy on those facts falls from 95 percent at month 0 to 74 percent at month 6, 56 percent at month 12, and 33 percent at month 24. Retrieval, pulling from a knowledge base that's kept current, holds steady at 86 percent the entire time. It was never ahead at month 0, but it never falls behind either.

The key lines. From code/mini-finetune/staleness.py:

# from code/mini-finetune/staleness.py
for m in range(months + 1):
    if m > 0:
        for i in range(n_facts):
            if rng.random() < change_per_month: version[i] += 1
    if retrain_every and m > 0 and m % retrain_every == 0:
        known = list(version)
    ft = sum(ft_acc for i in range(n_facts) if known[i] == version[i]) / n_facts

Read it as: the tuned model's accuracy on a fact is counted only when known[i] == version[i]. The version it learned still matches reality. Once a fact changes, the model's accuracy on it doesn't degrade gracefully, it drops to wrong outright, which is why the curve falls as steadily as it does.