arXiv:2609.08618v1 Announce Type: new Abstract: Benchmark scores describe what a checkpoint can do now, but they do not determine how it will respond to the next training episode. We measure this mis...
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.
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.
The paper introduces the concept of intervention fidelity in latent world models, measuring whether a model’s open‑loop transitions align with actual environment interventions. Experiments on TD‑MPC2, Cheetah, and DreamerV3 show that high reward fit does not guarantee fidelity, and that self‑supervised models can outperform task‑anchored ones in preserving intervention effects. The authors propose a capture‑gated audit to localize failures and argue that fidelity must be directly audited on the model’s native interface.
The paper investigates what a steerable valence direction in large language models (LLMs) actually represents, focusing on a good‑bad outcome direction in a maze task. By using controlled interventions that separate the realized outcome from the informational history that led to it, the authors find that directions trained on one explicit outcome encoding transfer well to another, suggesting the readout is not tied to surface form. However, when the same outcome is achieved through announced versus unannounced histories, transfer performance drops sharply, indicating that the post‑event readout remains strongly conditioned on the earlier announcement. In a matched maze‑reinforcement‑learning run, the post‑RL direction becomes more predictive of reference‑MDP return and the policy depends more on it, yet the history dependence persists. These findings support a functional, value‑related interpretation of the direction but argue against identifying it with a history‑invariant scalar valence state.
The study investigates how layer‑wise intervention responses in language models change over the course of pretraining, using single‑block identity bypass across multiple checkpoints and model‑domain combinations. It finds that while depth ordering of responses persists, their magnitudes shift, with nearby checkpoints showing stronger rank correspondence than distant ones and large changes occurring at positions that recur across samples and transfer across evaluation domains. Controlled experiments reveal that these longitudinal changes cannot be explained by a single downstream sensitivity and depend on perturbation strength and direction, indicating that layer sensitivity is structured but dynamic.
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.
arXiv:2607. 18553v1 Announce Type: cross Abstract: Can a language model read the quality of ongoing computation, and can an external intervention turn that readout into better outcomes?
arXiv:2608. 16085v1 Announce Type: cross Abstract: Capability development is routinely inferred from behavioural thresholds, from final checkpoints, or from what a decoder can read out of a hidden state.
The paper investigates whether large language models’ reasoning traces truly contain early, informative signals or merely reflect budget and difficulty confounds. Using a restart‑controlled truncation probe, the authors compare continuation success rates against from‑scratch restart curves across 178 problem‑model pairs, finding that only one case shows prefix‑limited success and that continuing a model’s own prefix generally outperforms restarting. A difficulty‑controlled test and two generation‑free analyses reveal that early internal signals do not carry outcome information beyond a problem‑difficulty baseline, underscoring the need for proper counterfactual controls.
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.
arXiv:2609.01108v1 Announce Type: new Abstract: TRACE (Math & Lienhart, arXiv:2602.01135) reads causal graphs over event types out of a pretrained autoregressive sequence model by thresholding a per-...