Skip to the content.

Context vs memory — when an in-context fact contradicts the stored one

Synthesizes the two behaviour threads: induction (in-context copy) vs factual recall (in-weights memory). Prompt: “The capital of France is Berlin. The capital of France is ___” — the in-context answer (Berlin) contradicts the stored fact (Paris). margin = logit(context) − logit(memory) (>0 → context wins). Then the causal link: mean-ablate the model’s induction heads (the in-context-copy mechanism) — if induction is what makes context win, ablating it swings the margin back toward memory; a random same-size head-set is the control.

model context-win rate (margin) − induction heads − random heads
gpt2 44% (-0.14) 0% (-2.61) 19% (-1.27)
gpt2-medium 81% (+0.66) 19% (-0.38) 94% (+0.85)
gpt2-large 56% (+0.01) 12% (-1.19) 50% (-0.03)
gemma-2-2b 6% (-0.80) 0% (-3.19) 6% (-0.70)
Llama-3.2-1B 0% (-1.19) 0% (-1.77) 0% (-1.54)
Qwen2.5-1.5B 0% (-1.50) 0% (-4.40) 0% (-1.62)

How much context to override memory? (repetition sweep)

Context-win rate as the false fact is repeated N times in-context before the query — does more in-context evidence eventually flip the memory-dominant models?

model ×1 ×2 ×3 ×5 ×8
gpt2 44% 81% 94% 100% 100%
gpt2-medium 81% 100% 100% 100% 100%
gpt2-large 56% 100% 100% 100% 100%
gemma-2-2b 6% 100% 100% 100% 100%
Llama-3.2-1B 0% 100% 100% 100% 100%
Qwen2.5-1.5B 0% 100% 100% 100% 100%

This reframes the memory-dominance. At 2 repetitions every model flips to 100% context-win — including the one-shot-memory-dominant RoPE models (Gemma 6%→100%, Llama/Qwen 0%→100%). So the RoPE “memory-dominance” is purely a one-shot effect: their induction needs ≥2 prior occurrences to fire strongly enough to override memory, where GPT-2’s fires on a single one. It is induction’s evidence threshold, not a refusal to trust context — and once induction fires (≥2 reps) it overrides stored memory universally.

Finding — two regimes. (1) The GPT-2 family is context-swayable (context-win 44–81%) and induction is the mechanism: ablating the induction heads collapses context-win to 0–19% (memory wins), far more than ablating a random same-size head-set (which leaves it ≈baseline or higher). So induction is what lets a fresh in-context statement override stored memory. (2) The RoPE family is memory-dominant: Llama and Qwen ignore the contradicting in-context fact entirely (0% context-win) and Gemma nearly so (6%) — on a single in-context statement. But the repetition sweep above shows this is a one-shot effect: at ≥2 repetitions all six flip to 100% context-win. So the difference is induction’s evidence threshold (RoPE needs ≥2 occurrences, GPT-2 fires on 1), not a standing refusal to trust context; induction is the shared override mechanism once it fires. Provisional, ~16 capital facts, single-token answers. Data: context_vs_memory_summary.json. Regenerate: context_vs_memory.py. See induction + where facts live.