CORTEXA
← Browse
arxivcs.CLcs.LG2026-07-20

Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift

Zitong Huang, Gustavo Lucas Carvalho, Deqing Fu, Robin Jia

We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task. Our key intuition is that by training the model to distinguish good and bad tokens in a response, we naturally guide the model towards generating good tokens, while avoiding the pitfalls that come with directly training the model to generate off-policy tokens. Experiments on document summarization tasks show that TOPL achieves strong out-of-distribution generalization across 11 datasets against a diverse set of sequence-level and token-level baselines. We further demonstrate that TOPL transfers effectively to machine translation, suggesting that its benefits generalize across different faithful generation tasks. Through ablation studies, we confirm that our token-level learning signal is critical to good performance; sequence-level analogues do not confer similar benefits. Finally, we show that TOPL induces interpretable model updates: the LoRA adapters learned through TOPL function as linear classification heads and steering vectors.

View free PDFSource page

Related papers

arxivcs.CLcs.AIcs.LG2026-07-06

Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment

Yu Li, Xiuyu Li, Mingyang Yi, Jiaxing Wang, zhangliangxu, Zhaolong Xing, et al.

Reinforcement learning (RL) post-training for large language models (LLMs) follows a efficient paradigm of "rollout then update", which inevitably results in off-policy training data. To resolve this, Importance sampling (IS) is proposed, while the token-level ratios compound ove…

View free PDFSource page
arxivcs.LGcs.AIcs.CL2026-07-11

ARMOR: Stabilizing On-Policy LLM RL with Off-Policy Anchor Samples

Kexin Huang, Junkang Wu, Jinda Lu, Shuo Yang, Chiyu Ma, Jiancan Wu, et al.

Reinforcement learning (RL) has significantly enhanced the reasoning capabilities of large language models (LLMs), yet the training process remains notoriously fragile. In this work, we investigate a critical source of this instability: over-optimization, where models exploit tra…

View free PDFSource page
arxivcs.LGcs.CL2026-07-04

Rethinking AI-Generated Text Detection: A Strong Baseline and the Distribution-Shift Problem That Remains

Zhuoer Shen, Mingyi Wang, Shaofeng Zou, Yuheng Bu

Recent AI-generated text detection work often introduces a new benchmark together with a specialized detector tailored to it. We revisit this practice from a baseline-first perspective. Across several benchmarks, we show that a plain, fully fine-tuned RoBERTa matches or exceeds t…

View free PDFSource page
arxivcs.LGcs.AIcs.CL2026-07-16

Beyond Entropy: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization

Weiwen Xu, Jia Liu, Hou Pong Chan, Long Li, Deng Cai, Min Chen, et al.

Reinforcement learning with verifiable rewards (RLVR) commonly uses entropy for advantage shaping. However, entropy cannot distinguish useful uncertainty from detrimental confusion, limiting its effectiveness as a correctness signal. We propose Contrastive Policy Optimization (CP…

View free PDFSource page
arxivcs.CRcs.CLcs.LG2026-07-11

One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions

Tomas Bruckner

Large language models (LLMs) are increasingly consumed through opaque serving chains - API aggregators, resellers, and inference providers - in which the client has no technical means to confirm that the model answering is the model advertised, and recent audits show that a subst…

View free PDFSource page
arxivcs.CLcs.LG2026-07-08

Diagnosing and Calibrating Tool-Call Boundary Drift in Multi-Teacher On-Policy Distillation

Jiabin Shen, Guang Chen, Chengjun Mao

Agentic language models must learn when to call tools, when to consume tool responses, and when to answer directly. This makes multi-teacher on-policy distillation a natural training strategy: one teacher can specialize in tool calls, another in direct responses, and the student…

View free PDFSource page