arXiv:2610.17093v1 Announce Type: new Abstract: A model diff credits post-training with whatever separates a base checkpoint from one post-trained descendant, presuming a rerun would land close by. O...
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.
arXiv:2609.25052v1 Announce Type: new Abstract: "An agent that writes its conclusions into a store it later retrieves from closes a loop usually reported as one-way contamination. Taking the loop to...
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. 03263v1 Announce Type: cross Abstract: We test whether the "compositional ignition" reported in latent-reasoning models is real computation, an instrument artifact, or inherited from verbal training data.
The paper introduces a Causal Market Simulator using a continuous‑time causal prior‑data fitted network (PFN) to estimate market impact, addressing bias from informational flow that traditional models overlook. On a synthetic market prior with adjustable confounding, the PFN matches an unbiased oracle while a naive estimator overestimates by up to 68%. In zero‑shot tests on real limit‑order books and equity data, the model’s intervention responses align with ground‑truth scales and improve correlation with measured impact when a recurrent state‑space variant is used.
arXiv:2609.01048v1 Announce Type: cross Abstract: Across the full Pythia suite (160M-12B, eight checkpoints, four task families), a linear probe can read a target variable from the residual stream as...
The paper presents a framework for causal attribution in agentic AI systems, outlining estimators and conditions where they fail. It distinguishes between marginal total effects and common‑random‑number total effects, introduces a natural direct effect under pinned downstreams, and derives a coupling method to keep direct effects estimable. The authors also propose a traceability specification to meet upcoming regulatory requirements for high‑risk AI systems.
The paper investigates the safety of Code World Models, where a language model generates executable world models that a planner uses. It shows that accepting a model based on sampled transitions only guarantees sample consistency, not full safety, because the probability of missing critical events decays as (1‑r)^N. Experiments on hybrid instruments reveal that omitted mode‑boundaries can severely limit planner performance, and that even sophisticated LLMs (GPT‑5.x) struggle to repair such omissions in higher‑dimensional settings.
arXiv:2606. 15621v1 Announce Type: new Abstract: Per-token counterfactual credit estimation asks which token in a language-model rollout caused the final answer to be right or wrong: cut the transcript at a pivot, substitute an alternative token, replay continuations, and compare outcomes.
arXiv:2610.09541v1 Announce Type: cross Abstract: Signals that predict whether a chain-of-thought (CoT) trace is correct are compared by AUC, but deploying one requires a threshold with a guarantee....
arXiv:2608.21262v1 Announce Type: cross Abstract: Many machine-learning systems set a threshold at a quantile of a calibration set: conformal predictors that promise 90% coverage by drawing their cut...