When a context asserts two values for one fact, a model commits to a cue -- recency, repetition, position -- but natural data rarely makes these disagree, so behavior cannot reveal which. We train 26M-parameter transformers on a synthetic language where recency and rarity are exactly coextensive, and separate them with a minimal causal edit that inverts one cue while holding the truth, token count and answer position fixed. All 75 runs reach accuracy >= 0.999, including where the trivial heuristic fails, so no held-in evaluation distinguishes them. Under intervention the per-cell readout does not replicate: 13 of 25 cells differ by more than 0.3 in sign fraction across three seeds, the largest by 0.879 against a standard error of 0.025. The construction predicts this -- coextensive rules leave the objective indifferent between them -- and the variance is ordered by how much of the optimization each comparison releases. What replicates is timing: escape from a positional shortcut with a closed-form ceiling, monotone in redundancy. Probed before that escape, attribution reverses sign in 32 of 75 runs at unchanged accuracy, and gating on circuit formation is necessary but not sufficient. The corpus fixes when a mechanism appears, not which one -- a criterion for when mechanistic attribution to data is available at all, and our construction makes the unavailable case exact.
Behavioral accuracy, linear decodability, and successful activation interventions do not by themselves show that a model carries an operation-level structure from one symbolic domain to another. We ask a narrower question in finite isomorphic state spaces: if the hidden-state difference between two operations is estimated separately for each source input, does adding that difference to a mapped recipient input move the model toward the corresponding recipient answer? The design compares this input-specific intervention with wrong-operation, norm-matched random, and no-op controls, and separates candidate construction from an independently isolated confirmation split. On a frozen Qwen2.5-7B-Instruct model at layers 20--21, one route--domain--operation candidate from a family pre-specified and frozen before confirmation access, transparent | integer_mod16--letters16 | successor->predecessor, passed both PyVene splits; its confirmation intersection--union p-value was 0.000198 and its 36-family Holm-adjusted p-value was 0.006943. A subsequent NNsight 0.7.0 experiment, pre-specified and frozen before its confirmation access, tested only this selected prompt route, without candidate or layer reselection. It reproduced all 12 confirmation effect estimates, confidence intervals, and exact sign-flip p-values numerically; its 36-family Holm-adjusted p-value was 0.007141. The result is therefore limited to one prompt route and one candidate, replicated across two intervention implementations on one model revision and one layer interval. It does not establish cross-model generalization, full-family backend independence, domain-general transfer, or algebraic invariance.
Arjun Pillai, Christian Hoang, Anjelo Jann Larozacs.CL cs.AI
Why do multilingual language models sometimes generate in the wrong language, and why is this so hard to fix? We introduce Language Identity Head Ablation (LIHA), a causal intervention that zeros each attention head individually and measures the resulting language switch rate across a parallel dataset of 2,700 prompt-language pairs spanning seven languages. Applied to GPT-2, LIHA identifies a small set of first-token broadcaster heads - led by L6H1 (switch rate 0.32, 3.23 $σ$ above the population mean) - that attend persistently to the first prompt token, propagating its language signal throughout generation. Compensatory redistribution when heads are ablated is statistically significant (p < $10^{-5}$) and follows a directional, hierarchical pattern: compensation always recruits heads in layers above the ablated head, suggesting a feedforward cascade rather than global diffusion. To probe how training regime shapes these circuits, we apply LIHA to a controlled pair - Qwen2.5-1.5B-Base and Qwen2.5-1.5B-Instruct - identical in architecture and size, differing only in training. The base model is nearly flat (max SR=0.016, 200/336 heads at SR=0.0); the instruct model concentrates causal influence sharply at layer 0, led by L0H5 (SR=0.224, 8.93 $σ$ above mean), with all other layers near zero. This controlled comparison provides direct causal evidence that instruction tuning reorganizes language identity circuits toward early-layer localization. Extended experiments with Chinese and Russian confirm that first-token broadcasting is script-specific in GPT-2, with non-Latin languages handled at layer 0 - the same locus as the instruction-tuned model. Code and data will be released upon publication.
Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood. In this work, we ask whether language identity is merely linearly decodable from hidden states, or if it can be causally controlled by a compact activation direction. We conduct an exhaustive causal intervention analysis across multiple model families, including Qwen 3.5-2B and Llama-3.2-1B-Instruct, isolating PCA-derived "language axes" to perform steering and ablation experiments across 1.26 million generations on the FLORES-200 dataset. Steering along these geometric directions reliably forces language switching in both cross-script (English to Chinese) and same-script (English to Spanish) settings, whereas equal-magnitude random perturbations yield virtually no effect. Our layerwise analysis reveals that language commitment is highly localized and explicitly language-pair-dependent. While English to Chinese switching resists early intervention and steers easily in the later layers, the English-Spanish transition shifts earlier, displaying a distinct, bimodal sensitivity. Furthermore, targeted ablation uncovers a fundamental reversion to English: once the language signal is removed, the model falls back to English regardless of the input prompt. Ultimately, these findings demonstrate that language decision boundaries function during inference as causally active features that are direction-dependent and layer-specific.