The Flow has not summarised this story yet — read it at arXiv Computation and Language.
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
arXiv:2606. 15733v1 Announce Type: cross Abstract: Instruction-tuned language models can answer the same causal-reasoning question differently after its English variable names are replaced by type-preserving placeholders, although the structural causal model and the gold answer are unchanged.
arXiv:2603.13911v2 Announce Type: replace Abstract: A language model can give the wrong answer even when the correct answer is decodable from its intermediate states. To study this gap between decoda...
arXiv:2608. 04021v1 Announce Type: cross Abstract: Cloze-style probes that vary how often a target token appears implicitly assume that more copies of a target affect prediction the same way regardless of where the readout slot sits.
arXiv:2607. 20524v1 Announce Type: new Abstract: Mean cross-positional attention degradation is widely reported in transformer interpretability, yet whether it causally limits contextual retrieval remains untested.
The paper investigates how visual interventions—training‑free edits applied to multimodal large language models—affect hallucinations and prediction quality. By applying a visual‑residual edit across successive decoder layers and matching its RMS strength, the authors identify an Evidence Anchoring Boundary (EAB) where the intervention’s influence drops sharply; effects before the EAB are 13.78–26.10× larger than after it. They further compare counterfactual subtraction across different question types and introduce calibrated selective routing, which only alters predictions when calibration supports an intervention, improving attribute accuracy by 1.56–1.81 percentage points while avoiding the harms of universal subtraction.