A 4B model learned two-hop geography with its top J-lens directions suppressed all through training

The suppression is chosen per position. At each of six middle layers the implementation ranks candidate directions by their logit at the current activation, then subtracts their residual projections ^[15]. Those directions are readout-aligned: the normalized unembedding row scaled by the final RMSNorm gain ^[12]. So what gets removed changes token by token, and the write-up is explicit that the intervention does not identify the intermediate country, remove a fixed set of neurons, or eliminate every representation associated with that country ^[16]. The post calls them the top-10 eligible directions, and the text available breaks off in the lesion construction section before eligibility is spelled out ^[25].

Two details in the selection change how the result should be read. The RMS denominator is shared across vocabulary logits at a fixed activation, so it does not affect how the directions rank ^[13]. The intervention itself does not differentiate through that denominator ^[14]. What the lesion tests, then, is sensitivity to a particular vocabulary basis, and the source paper is clear that a high lens score alone does not establish that a given computation uses the content it reads ^[22].

On the held-out cohort the error rate fell from 79.1% to 2.3%, a factor of about 34 ^[23]. The author's own accounting limits what that buys: accuracy says the model produced the answer, lens readouts say something about expression in one basis, and interventions test sensitivity to the selected directions ^[17]. Accuracy recovery does not distinguish a changed internal representation from a model that simply became less sensitive to those directions ^[18]. After refitting, the lesion still changed activations, with a smaller effect at later layers ^[5].

For the 97.7% to transfer to a monitoring setup, that setup has to resemble this one. The task was picked so the clean model could handle it reliably while the lesion still disrupted it enough to measure ^[19]. The run covers one 4B model on one two-hop task, with a selected cohort of countries excluded from training ^[1][3]. The intervention is the one from Anthropic's global workspace paper, where ablating active J-lens directions disrupts multi-hop reasoning while much routine processing survives ^[8][9]. If your monitor is a fixed readout basis applied at middle layers, and the task has more than one internal route to the answer, this is the failure mode to test for. This experiment stopped short of a behavioural eval.

The biological analogy in the post carries its own caution. Children with Rasmussen's encephalitis can recover substantial language function after the left hemisphere is removed, with the right hemisphere supporting further language development ^[20]. That recovery is incomplete: verbal comprehension may return while grammatical judgment and speech production stay impaired ^[21].

"The model learned to tolerate the intervention; exactly how it did so remains open," the author wrote ^[6].