- Rating
- 2
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies whether interpretability interventions (circuit ablation, activation patching, steering vectors have been mentioned in the paper) discovered on a development model remain valid when reused unchanged on a weight-quantised version of that model. It distinguishes two targets: source fidelity, phi = u/d, the transported intervention's effect on the quantised model relative to its original effect; and native parity, rho = u/v, the transported effect relative to that of an intervention re-discovered on the quantised model, which the authors call "interventional commutativity". Both are measured only on inputs where the two models' unintervened outputs agree. To diagnose disagreements, the paper proposes crossed execution: running the re-discovered intervention back on the development model, which splits the gap into a part that already exists before quantisation and a part caused by the two interventions responding differently to quantisation. On the theory side, a round-to-nearest construction shows that exactly equivalent behaviour can coexist with complete loss of a transported intervention's effect, and an effect-mass bound separates preservation of the mean effect's sign from preservation of its magnitude.
Validation centres on an FP16/INT8 audit that crosses seven small base models (up to 1.5B parameters), four tasks, three intervention types and three seeds, using effect floors calibrated on FP16 controls and prompt-clustered uncertainty. Native parity is close to one in the vast majority of cells, with a small number of severe gaps, mostly in an assistant-prefix ablation stress task, while source fidelity remains largely preserved. A held-out study on one Qwen configuration uses crossed execution to show that nearly all of its native gap already exists before quantisation. Further extensions cover additional quantisation and pruning schemes, instruction-tuned models, official GPTQ/AWQ checkpoints, sequence-level refusal checks, disjoint discovery/evaluation splits on GPT-2, and a small gate-based repair pilot.
Critical strengths
The main strength is the distinction between keeping an intervention's original effect and matching what re-discovery finds on the quantised model. The two sound similar but they are not the same, and crossed execution (running the re-discovered intervention back on the original model) is a simple, clean way to tell them apart. I think the lesson reaches beyond quantisation: whenever people compare interventions across model versions, a large gap to a re-discovered intervention does not necessarily mean the original one broke.
The paper's evaluation framework is also rigorous. The authors calibrate against FP16 controls, report uncertainty at the prompt level, re-ran the whole main audit after finding a bug, and released the raw records and code needed to regenerate the main results. Because of this I trust the numbers. My concerns are about what the numbers mean, not whether they are right.
Critical weaknesses
As mentioned in the Summary part, the main weakness is that the theory, the experiments and the conclusions point in different directions. Proposition 1 and the title are about quantisation breaking interventions, but the INT8 experiments show the opposite: the original effects are mostly preserved. The handful of cases with very low rho happen because the re-discovered intervention is very different from the original one, usually much stronger, not because the original effect was lost. In the case the authors look at closely, their own crossed execution shows that the gap was already there before quantisation. So what the paper really finds is that the discovery procedure is unstable. Switching the compute backend, or fixing a single implementation detail, is enough to pick a quite different intervention. That is interesting in its own right, but the paper does not try to explain it. In the end, the experiments do not answer the question the paper asks, and the question they do raise is left open. This is the main reason for my recommendation.
A second weakness is that the evidence is thinnest exactly where problems are most likely. INT8 with group-wise rounding is close to lossless, so it is not surprising that the original effects survive. The settings practitioners actually worry about are 4-bit quantisation and larger models, and there the paper has only a few results, some of them from before the bug fix. Even accepting the paper's framing, it cannot tell a practitioner whether their interventions will still work in the settings they actually deploy.
Questions for authors
1, How do the experiments answer the question the paper poses? As I mentioned that the biggest problem is the misalignment between the theory, experiments, and conclusions. The experiments mostly show that the original effects are preserved, and in the one configuration analysed with crossed execution, the gap to the re-discovered intervention already exists before quantisation. Could the authors state explicitly what rho is meant to measure: the effect of quantisation on an intervention, or the stability of the discovery procedure? If it is the former, could they apply crossed execution to all of the severe cases, not only the one examined in detail, and report how much of each gap is caused by quantisation rather than already present on the development model? If a substantial part of these gaps turns out to come from quantisation, the paper's framing would be much better supported and I would reconsider my assessment. If not, the paper's main message is about discovery stability rather than quantisation, and the framing should change accordingly.
2, This one should be considered as a kind of suggestion for future work. The results suggest that the instability of the discovery procedure may be a more fundamental finding than the quantisation question itself. Have the authors considered developing it as a study in its own right? Possible directions include: treating quantisation as one of several small perturbations (compute backend, numerical precision, batching, small weight noise) and measuring how often the selected sites change under each; relating this instability to how close the competing scores are, and to whether exact or approximate scoring is used; explaining why swapping a single site can change the joint effect so much, for example through redundancy or non-additive interactions between components; and testing whether discovery procedures that aggregate over such perturbations yield more reliable interventions. I would be interested to hear whether the authors plan to pursue this direction, either by reframing the current paper around it or as separate work.
2 comments
Thank you for the additional experiments and clarifications. But I regret to say that I will maintain my original assessment. The additional evidence further reinforces my concern that the paper's experimental findings and contributions do not adequately support its original claims of significance. In my view, the current work is better considered as an exploratory or diagnostic study, rather than a paper claiming substantial methodological or theoretical contributions. Therefore I will hold the same judgement, and I believe this work might be more suitable for another venue with a different scope and expectations.
Best and good luck.
We agree with your central distinction: a low ratio to re-discovery is not itself evidence that quantisation broke the original intervention. Rho measures mean-effect parity between the reused and native interventions on the compressed model; it is neither a pure quantisation-effect measure nor a calibrated measure of discovery instability. Source fidelity measures the original-effect comparison, and crossed execution separates reference selection from execution changes (Section 3, pp. 3–4; Equation 4 (p. 3)).
1. Alignment between the question, experiments and conclusion. We follow the branch you specify: “If not, the paper's main message is about discovery stability rather than quantisation, and the framing should change accordingly.” The revised framing is an audit of effect preservation and reference selection, not an attribution of the eight low-native-parity cells to quantisation-induced source loss.
The additional T4 crossing covers retained site choices in seven of the eight historical severe seed keys, across three of four configurations. It refits each intervention's means on its own model and recomputes matching before fixing both operators. The signed decomposition is:
All terms are in logits. These original-model terms account for 99.6–101.1% of the signed reference gaps; values above 100% reflect cancellation by the execution interaction. We also retain the individual changes: in Qwen-1.5B/prefix they are +0.516 for the source intervention and +0.469 for the native intervention, despite an interaction of -0.047. A small interaction therefore does not show that either effect was unchanged. Source means in two T4 configurations fall below the historical magnitude guard; the Llama steering configuration has no completed additional crossing. These are crossings of retained sites with refitted payloads, not replays of all historical operators (Table 4 (p. 7); Table 42 (p. 41); Appendix N.1 (p. 40)).
The paired NF4-grid capture answers a different part of the question: legacy scoring gives a resolved negative source mean and a resolved positive signed shift toward zero. Equal-example scoring changes the selected interventions, sharply narrows their reference gap and leaves the new source denominator unresolved. The paper reports both findings, rather than treating the low legacy point or high equal-example point as protocol-independent preservation evidence (Section 7.1 (p. 9); Table 45 (p. 44); Table 46 (p. 45)). The valid dense-7B comparison and matched-protocol discovery-null calibration remain unmeasured.
2. Discovery stability as a research direction. The revision brings the existing controlled diagnostics into the argument. It documents exact ties at the selection boundary, distinguishes exact single-site scoring from first-order scoring, and tests the joint interventions rather than assuming additivity. In the targeted held-out Qwen experiment, replacing site 22 with 21 changes the joint effect from 0.426 to 4.388 logits; the two-swap interaction is 0.787 [0.770, 0.802] logits. The equal-example follow-ups isolate partial-minibatch gradient weighting from effect averaging. These observations identify concrete selection sensitivities (Table 25 (p. 27); Appendix L.2 (p. 35); Table 38 (p. 37); Table 45 (p. 44)).
They do not yet establish a perturbation-aggregated discovery method, a factor-separated null for all flagged configurations or a broad repair advantage. The gate pilot retains its negative source-target result and its dominant mean-offset component. We treat systematic perturbation-robust discovery as the future direction you suggest, while the present contribution is the explicit effect reference, four-execution diagnosis and evidence-bound audit protocol (Table 35 (p. 36); Table 37 (p. 37); Section 8, p. 10).