Linear probes can decode safety‑relevant concepts such as truthfulness from language‑model activations, but probe accuracy may reflect only decodability, not causal influence on model behavior. The authors show that probe weight geometry alone cannot identify the features the model actually uses, because geometrically aligned features need not be causally relevant. They introduce a sparse‑autoencoder (SAE) decomposition that ranks features by probe alignment and gradient sensitivity, and demonstrate that ablating shared, probe‑only, and random feature sets reveals a sharp dissociation: shared features drive model output changes far more than probe‑only or random features, confirming that causal relevance requires intervention beyond weight geometry.
By Devesh Tiwari, Camille Davis, Shivank Sinha, Talia Weaver, Aditya Shah, Maheep Chaudhary
arXiv:2608. 12652v1 Announce Type: cross Abstract: Benchmark contamination is diagnosed today with n-gram overlap, with likelihood-based membership inference, or with canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at dataset release.
By Florian Braun
The paper investigates in-context binding errors in language models, showing that a linear probe can recover correct entity bindings from frozen hidden states even when the model outputs incorrect bindings. Across 16 checkpoints, the probe’s accuracy on failure cases surpasses a baseline by about 0.196, and a probe‑based score improves failure detection over the model’s confidence by 0.079 AUROC. Steering the residual stream toward the probe‑decoded binding further boosts accuracy by an average of 0.168 across eight models.
By Manas Venkata Sai Ravulapalli, Samrath Singh Chadha, Abhinav M. Hari
The study measured the impact of a single training example on a GPT‑2 model by running 24 counterfactual experiments. 32 models were trained from scratch on OpenWebText, and at a specific training step a single batch row was replaced with a 194‑token passage under three conditions (fluent prose, fabricated subject, random characters) or left unchanged. Results showed that the passage was learned from one exposure and decayed, with measurable differences in cross‑entropy up to 50 steps after injection but no lasting effect at the final step.
By Zachary Speck, Asa Shepard
arXiv:2608. 01676v1 Announce Type: cross Abstract: Sparse attention is widely deployed in long-context serving stacks, yet no framework audits how discarding blocks changes the influence of specific content on model output.
By Xingyu Ren, Youran Sun, Chugang Yi, Haizhao Yang
arXiv:2608. 11797v1 Announce Type: new Abstract: Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap.
By Chencheng Zhu
arXiv:2607. 16821v1 Announce Type: cross Abstract: Task arithmetic, sequential fine-tuning, activation steering, and first-order random search all operate through relatively small perturbations around an already trained checkpoint, and they rely on different local approximations: individual perturbations should be first-order predictable, task updates should compose with controlled interference, useful tangent structure should be stable and possible to estimate, and weight edits should have counterparts in representation space.
By Irina Piontkovskaia, Sergey Nikolenko
The paper investigates how a probe can decode in‑context bindings on model errors and how probe‑guided steering can repair them. It tracks probe accuracy, model output, and steering response across public pretraining and post‑training checkpoints, noting that probe accuracy improves during Pythia pretraining and that steering benefits grow with model size. The study also shows that decoders trained on final state or candidate logits do not outperform each other on late‑checkpoint errors, and presents an information‑theoretic counterexample explaining why decodability on errors alone cannot prove discarded output information.
By Manas Venkata Sai Ravulapalli, Samrath Singh Chadha
arXiv:2608. 08029v1 Announce Type: cross Abstract: Khatri et al.
By Alizishaan Khatri, Dun Li Chan
The paper investigates whether hidden-state probes can identify the type of reasoning failure and the specific step that causes it in large language models. By comparing probes trained on final-answer correctness with those trained using Monte Carlo rollouts, the authors find that standard probes miss a large proportion of recoverable failures and that probe performance depends heavily on training design. Across 4B and 8B models, no probe outperforms a simple rule based on trace length for locating the failing step.
By Shubham Subhnil, Yannis Belkhiter, Erik Miehling, Adam Dahlgren Lindstr\"om, Elizabeth M. Daly
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.
By Han-yu Wang
arXiv:2607.06640v2 Announce Type: replace-cross
Abstract: World models learn task-relevant information through many routes: observation reconstruction, recurrent state, temporal filtering, and explic...
By Donna Vakalis