Knocking Out Shortcuts That Have Backups

When a neural network learns a "shortcut", it latches onto a surface cue that happens to predict the label in training data, such as the word "not" in an argument-reasoning benchmark. Recent work has found the internal parts that carry such cues, such as specific attention heads or neurons, and switched them off. This proposal asks whether that fix holds. The concern comes from biology. Genomes routinely keep backup copies of important controls, including duplicated genes, "shadow" enhancers (extra DNA switches for the same gene), and parallel feedback loops. Knocking out one copy often shows no effect because another takes over. We hypothesize that trained networks store shortcuts the same way. If so, disabling the one component that carries a shortcut would let a dormant backup take over, and the shortcut would quietly return once training continues.

The proposal borrows one experimental habit from genetics, the conditional knockout. Biologists do not simply delete a gene everywhere. They delete it only in chosen cell types and only during chosen time windows, and they delete several redundant copies together when a single deletion shows nothing. We translate each of these moves to networks. A unit is disabled only on the training examples where it is actively pushing the shortcut, only during a chosen window of training, and together with whatever backup units start carrying the shortcut once it is gone. The core claim is testable. Shortcut reliance in text models is buffered, meaning single-component ablation underestimates it and allows it to recover. A knockout that is targeted by example, limited in time, and extended to backups removes shortcut reliance more durably than always-on suppression, and costs less in-distribution accuracy.

What the literature establishes, and what it leaves open

The biology supplies the mechanism we test for. Dasmahapatra models duplicated genes that repress each other, so the second copy stays silent until the first is lost and then switches on. Zhang et al. find that when primary hindlimb enhancers mutate, backup enhancers compensate, so a trait decays only when primary and backup elements both degrade. Wen et al. describe promoters contacted by so many enhancers that losing any one changes little. Zeevi et al. show yeast promoters keeping expression stable after losing their main binding motifs, because many weak sites are spread across the promoter. The research file names the practical consequence as an open problem: single-factor knockout screens miss buffered components unless several are perturbed together.

We use these papers as motivation and as a source of hypotheses, not as evidence about networks. They predict two things we can measure. First, backups can stay dormant while the primary component is present, as in the duplicate-gene model. Second, a signal can be carried by many weak components rather than a few strong ones, as in the yeast promoters.

The machine-learning side supplies the phenomenon and the tools. Shortcut learning in deep neural networks frames shortcuts as a general consequence of minimizing average training loss. Niven et al. show that BERT's success on the argument task depends on cue words like "not", and that performance falls to chance on a version balanced to remove the cue. The 3-SAT studies (art_254ad3e6, art_442e2d73; Can Large Language Models Reason? A Characterization via 3-SAT) show language models predicting satisfiability from problem length. ReClor separates easy (biased) from hard (unbiased) test subsets.

Three methods constrain or supply our approach. Short-circuiting Shortcuts uses path patching to locate shortcuts in specific later-layer attention heads that decide before reading the full context. We reuse this localization method as our way of finding the component to knock out. Its natural fix, ablating those heads, is the "single knockout" whose durability we question. Alignment-Gated Suppression, abbreviated AGS, decays neurons whose class-conditional alignment is unusually high, gated by a batch percentile, throughout training. In genetic terms this is a global, always-on knockout, and its authors note it can wrongly suppress genuine features when batches are small or classes are imbalanced. Localized Shortcut Removal edits shortcut regions out of the input under a fixed budget. It fails on diffuse cues, and problem length in 3-SAT is exactly such a cue.

The research file's section on work that already combines these directions was cut off. I found no paper in the file that measures whether a network compensates after shortcut components are ablated, or that limits ablation by example or by time window. The closest work, the head-localization paper and AGS, either ablates once or suppresses everywhere. Neither searches for backups that activate after the primary component is removed.

Core Claim

Shortcut reliance is buffered. Two predictions follow. First, after the localized shortcut heads are ablated, continued training on the original biased data will restore much of the shortcut advantage, and new units will newly show shortcut alignment. Second, a knockout that is conditional by example and by time window, and that is extended to each newly recruited backup, will leave lower residual reliance than either a single ablation or always-on AGS, at equal or smaller in-distribution cost.

We make "buffered" measurable with a compensation index,

$$C = \frac{S_{\text{after}} - S_{\text{ablate}}}{S_{\text{orig}} - S_{\text{ablate}}},$$

where $S$ is shortcut reliance. Reliance is defined as the accuracy gap between the easy (cue-bearing) and hard (cue-balanced) test subsets. $S_{\text{orig}}$ is measured before ablation, $S_{\text{ablate}}$ immediately after ablation, and $S_{\text{after}}$ after $k$ further training steps with the ablation kept in place. A value of $C$ near 0 means the ablation fixed the problem. A value near 1 means a backup fully took over.

Research Program

The first stage is a measurement study. We fine-tune BERT-family models and small transformers trained from scratch on three tasks with known cues. We use the argument task with its cue-balanced version, 3-SAT with length decorrelated from satisfiability in the test set, and ReClor's easy and hard subsets. We localize shortcut heads with the head-localization method, ablate them, keep training, and track $C$ over time. We also re-run localization on the ablated model to see whether new components appear. This separates two cases that the biology distinguishes. In the "dormant backup" case, a few units that were previously silent on the cue become strongly aligned with it. In the "distributed" case, many units each carry a small share of the cue.

The second stage builds the conditional knockout, which has three parts, each borrowed from genetics. The first part is example-conditional masking. A unit is zeroed only on training examples where its activation is aligned with the cue-predicted label. This is the analogue of a driver active only in one cell type, and it leaves the unit free to carry genuine features elsewhere. The difference from AGS is that AGS applies a batch-level decay to every example, whereas we decide per example. The second part is windowed timing. The knockout runs only during a chosen training window, early, middle, or late, so we can test whether shortcut units play a necessary early role in feature formation, as essential genes do in embryos. The third part is iterative combinatorial knockout. After each window we re-localize, add any newly recruited backups to the knockout set, and repeat until $C$ stays low or in-distribution accuracy falls past a preset limit.

Baselines

The naive reference is standard fine-tuning on the biased data with no intervention. It defines $S_{\text{orig}}$ and the hard-subset accuracy floor. No reported result found on Lacuna for the numeric accuracy of this reference on the argument task, 3-SAT, or ReClor. The file reports only qualitatively that the argument task falls to chance under balancing, so we reproduce every number ourselves.

The single-knockout baseline ablates the heads found by Short-circuiting Shortcuts, both permanently and followed by continued training. It is the most direct competitor because it applies the same localization tool without conditioning or iteration.

The strongest existing method that needs no group labels is AGS, run always-on. We tune its percentile and decay rate over the same number of runs we spend tuning our threshold and window. Both use identical batch sizes, and we report results at small batch sizes as well, since the AGS authors identify that as its weak point.

Localized Shortcut Removal serves as an input-side baseline. We expect it to work on the "not" cue and to fail on the length cue.

Every method gets the same data, the same base checkpoints, the same compute, and five random seeds. Beating the baselines means lower $S_{\text{after}}$ and higher hard-subset accuracy than both single-head ablation and tuned AGS, at no more than a two-point in-distribution loss, with the gap exceeding seed variance. Two ablations isolate our ingredients. We remove the iterative step, giving an example-conditional single-round knockout, and we remove per-example conditioning, giving a windowed global knockout.

Why not just…?

Why not just ablate the localized heads and stop? If $C$ turns out near 0, this simple fix is enough and our method is unnecessary. The first stage is designed to settle exactly this question, and a near-zero $C$ would be a publishable negative result that simplifies the field. We predict otherwise because the biological analogues predict dormant backups, and because the heads the localization paper finds sit in later layers, downstream of representations that still contain the cue.

Why not just run AGS, which already suppresses shortcut neurons without group labels? AGS is the serious threat. It may already suppress backups as they appear, since it is always on. Our comparison tests two things. We check whether AGS reaches low $C$ without hurting the hard subset, and we run the windowed global-knockout ablation to see whether any advantage we find comes from timing or from conditioning. If tuned AGS matches us, we will report that its always-on design already achieves the effect of combinatorial knockout.

Why not just rebalance the data? Rebalancing requires knowing the cue. Niven et al. also report that removing the cue from the argument task drives BERT to chance, which suggests rebalancing removes the signal without teaching the task. We include cue-balanced fine-tuning as a reference that uses privileged knowledge of the cue, as an upper bound on what cue knowledge can buy rather than as a competitor.

Decisive Checks

Three results would decide the proposal. The first is whether $C$ after single ablation is reliably above zero across seeds on at least two of the three tasks, and whether re-localization finds new cue-aligned units. If neither holds, the redundancy hypothesis fails for these settings. The second is whether iterative conditional knockout gives lower $S_{\text{after}}$ than tuned AGS on the 3-SAT length cue, the diffuse case where input editing cannot help. The third is whether the knockout window matters. If early-window knockout harms hard-subset accuracy while a late window does not, shortcut units play a necessary early role in feature formation. That would be a concrete, testable carry-over of the essential-gene logic that makes conditional knockout necessary in biology.