Agentic Critical Training (ACT) is a reinforcement learning approach that trains language‑model agents to judge actions directly by pairing expert actions with alternatives and rewarding correct selection. Unlike imitation learning or self‑reflection, ACT does not require reference rationales and can be reused across model sizes. Experiments on Qwen3‑8B and Olmo‑3‑7B‑Instruct show that ACT improves performance on ALFWorld, WebShop, ScienceWorld, and outperforms supervised reflection baselines, with significant gains over IL and IL→RL baselines.
By Weize Liu, Minghui Liu, Sy-Tuyen Ho, Yongkyun Lee, Andrew Adams Schoen, Souradip Chakraborty, Xiyao Wang, Furong Huang
arXiv:2607. 29484v1 Announce Type: cross Abstract: Interventional data is widely regarded as the gold standard for teaching models causal reasoning.
By Xining Xun
The paper reviews On‑Policy Self‑Distillation (OPSD), a method where a language model learns from its own generations using privileged information such as reference solutions or plans, eliminating the need for a larger teacher model. It identifies a key failure mode—collapse, where the model’s reasoning paths narrow progressively—and analyzes it through three levers: signal application, privileged information, and teacher dynamics. The review focuses on mathematical reasoning, offering a unified vocabulary and distinguishing settled facts from ongoing debates.
By Justin Robert, Raheel Qader
arXiv:2606. 11709v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with the distribution it produces under privileged context, typically a verified solution.
By Leyi Pan, Shuchang Tao, Yunpeng Zhai, Lingzhe Zhang, Zhaoyang Liu, Bolin Ding, Aiwei Liu, Lijie Wen
The paper investigates whether large language models (LLMs) truly perform formal causal reasoning by testing them on datasets generated by the CLadder framework. It finds that LLMs and even simple models like random forests can achieve high accuracy on CLadder data, largely due to superficial, highly predictive patterns embedded by the generator. The authors argue that strong performance on CLadder likely reflects exploitation of these patterns rather than genuine causal reasoning, and that many commercial LLMs show brittle results when critical text is removed.
By Justin Clarke, Isha Karn, Patrick Murray, KJ Hill, David Jensen
The paper introduces a method for analyzing how interventions in a prompt affect a language model’s reasoning by using behavioral counterfactuals. It trains a module to edit the residual stream so that the next-token distributions for a fixed response match between a source and a counterfactual prompt, outperforming standard activation reconstruction. The resulting attribution maps and visualizations help identify causally important locations for counterfactual recovery and guide hypothesis testing with activation patching, as demonstrated in a case study on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B.
By Shawn Im, James Oldfield, Adel Bibi, Sharon Li
On-policy distillation trains a language model on its own generations while a teacher scores them token by token. It combines the dense supervision of imitation learning with the on-policy sampling of...
Faithfully understanding the internal process that occurs as models form their chain of thought is an important problem with the rise in model capabilities. However, this remains challenging due to th...
arXiv:2510. 27544v3 Announce Type: replace Abstract: Current training paradigms, optimized for long-horizon reasoning trace execution, have made Large Language Models (LLMs) excel at pattern matching and forward simulation of reasoning, but underperform at counterfactual causal understanding and reasoning.
By Nikolaus Holzer, William Fishell, Baishakhi Ray, Mark Santolucito
arXiv:2609.37041v1 Announce Type: cross
Abstract: Self-Distillation Fine-Tuning (SDFT) enables a language model to act as its own teacher: by conditioning on a demonstration, the model produces an im...
By Su Ee Tan, Xiaotong Ji, Rasul Tutunov, Haitham Bou-Ammar, Matthieu Zimmer
The paper investigates when in‑context learning (ICL) can succeed even if the instruction or target task is not covered by a model’s pretraining data. It introduces a theoretical framework that distinguishes two types of distribution shift—unsupported instructions and unsupported target tasks—and shows how demonstrations can mitigate these shifts. The authors further extend the analysis to long‑horizon tasks that can be decomposed into pretraining‑supported subtasks, providing error bounds that explain how ICL can handle tasks beyond the model’s original support.
By Xuhan Tong, Yuchen Zeng, Jiawei Zhang
The paper shows that controlled visual in‑context learning can help vision‑language models (VLMs) overcome errors caused by training priors. By providing demonstrations, the models focus more on visual evidence, yielding significant accuracy gains on counterfactual examples while leaving prior‑aligned performance largely unchanged. This effect is observed across five VLM families and five vision‑centric tasks, indicating that many VLMs possess latent visual grounding abilities that standard tests may miss.
By Kun Wang, Xindi Wu, Sanghyuk Chun, Olga Russakovsky, Esin Tureci