- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies how punctuation and chat-turn boundaries affect retrieval from context. In random-token sequences, boundaries change the probability of retrieving a previously observed continuation, with effects depending on their placement relative to the evidence and query. Path patching in GPT-2 small attributes much of the effect to key/value pathways at the continuation token. The paper also tests 14 natural-language domains and finds that separating two facts with an empty assistant turn generally shifts predictions toward the second fact.
Critical strengths
The synthetic experiments are well controlled. Replacing filler tokens preserves sequence length and the positions of the cue and target. Sweeping the boundary across the sequence is particularly useful. It separates an effect near the stored association from an effect near the retrieval query. This provides more information than a comparison between two prompt templates.
The GPT-2 patching results go beyond an attention visualization. Jointly patching keys and values at the cue and continuation positions restores about 85% of the effect, while the reverse intervention removes about 84%. The strong interaction between key and value channels is also interesting. These results give a reasonably concrete account of where the effect is carried in this setting.
Critical weaknesses
I am not sure the GPT-2 experiments explain what happens in the natural-language tasks. Putting the second fact in a new user turn adds tokens and changes how the conversation reads. The model might interpret it as a correction or update to the first fact. The random-token experiments do not rule out this explanation. Although the authors acknowledge that the circuit analysis is limited to GPT-2, the discussion sometimes treats it as an explanation for both sets of results. A control with the same token length would help, as would testing whether intervening on the proposed mechanism changes the natural-language effect in an instruction-tuned model.
The natural-language results suggest a real shift toward the second fact, but some claims about its size seem too strong. In Figure 6, dividing by a very small positive single-fact gap can make the normalized effect look large. The ratio in Appendix A.7 has a similar problem when the single-turn order effect is near zero. It also uses absolute values, so it does not distinguish a stronger preference for the first fact from a stronger preference for the second. Finally, a normalized swing of one does not necessarily mean the model answers as though it saw only the second fact. To support that claim, the authors should compare the actual predictions with the second-fact-only predictions. Showing the raw changes alongside the normalized results would make the “tenfold” claim easier to assess.
Questions for authors
Does the natural-language effect remain when the single-turn control contains an equal-length separator, or when the intervening assistant turn contains a neutral acknowledgment? These comparisons would help distinguish a local boundary effect from conversational update behavior.
Can you report the raw single- and split-turn swings, the single-fact gaps, and uncertainty across fact-pairs? In particular, does Figure 6 retain its pattern after excluding small positive denominators, and do the individual split-turn predictions actually approach their corresponding single-fact predictions?
Which experimental setup produced Figure 2: the one described in Section 4.1 or Appendix A.2?
2 comments
Thanks the authors for taking efforts to address most of my concerns. I am willing to raise my score to 3.
Response to Reviewer wZjj
Thank you for taking the time to craft such a thoughtful review. We enjoyed reading your suggestions and worked hard to incorporate them into our paper.
New Experiments: We conducted a series of new experiments that more granularly defines and tests boundaries, including your suggestions of an equal-length separator and a neutral acknowledgment from the assistant turn. There are four single-turn variants: the two facts are separated by (1) a single space, (2) the word 'and', (3) a new sentence and (4) a new line. There are three split-turn variants: the assistant turn is (1) blank, (2) replies 'Okay.' and (3) acknowledges the first fact. Across all four single-turn variants, the amplification of the more recent fact was significantly smaller than across the three split-turn variants. These results help to confirm that the natural language results hold even to perturbations of boundary types within a single turn, when compared to different assistant turns across two different turns. We ran this on three representative configurations spanning three families and both training types (Qwen 2.5 1.5B-Instruct, Gemma 4 E2B-it, and OLMo 2 1B), across all 14 domains, and included it in our main manuscript with figures that portray the conditions and results.
New Dependent Variable: We have changed the way we measure amplification to address your concerns around dividing by small, positive gaps. Instead, we now normalize E(LH) - E(HL) by the 'bucket range', or the difference in the maximum and minimum answers for a specific domain. This allows for a dependent variable that is both interpretable (what percent of the feasible range is shifted by switching fact order?) and robust to skewness (it is a constant fixed for each domain and doesn't vary by model). The core of our results (split-turns increase swing from switching fact order) holds under this new framing, as is best depicted in Figure 6. However, we also included the single- and split-turn swings and single-fact gaps in the Appendix per your request (even though the main manuscript now uses the 'bucket range'). The vast majority of rows see a larger 'Split swing' than a 'Single swing'.
Experimental Setup for original Figure 2: We thank the reviewer for this catch; the settings used for Fig. 2 were what was described in Appx. A.2 (|F1|=32, |F2|=128, |K|=1), not what we had written in Sec. 4.1 (|F1|=|F2|=32, |K|=3). We have corrected the text.