Latest RLHF Research Papers

The newest RLHF papers from across the field — arXiv, NeurIPS, CVPR, Nature, and more — refreshed daily and ranked by relevance. Distill AI tracks RLHF so you don’t have to: get the standout work delivered to your inbox every morning, with 2-sentence summaries and the option to chat with any paper.

Recent papers

• Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese
  Pedro Carvalho Brom · arXiv · Oct 29, 2026

  Building a language model of one's own no longer requires a data centre. This tutorial trains one from scratch, tokenizer included, on one consumer GPU, and shows how to tell whether it is good for the tasks of a specialist domain. It propo…

• AI Value Judgment from the Perspective of Hume's Law: A Critical Inquiry
  Jingxuan Luo · Communications in Humanitie... · Oct 27, 2026

  Value Alignment Engineering for AI aims to turn human normative demands into computable parameters. But there are problems with the conflict between ethical principles and disagreements in RLHF annotation. Therefore, this paper investigates…

• Selective Tie Learning: Mitigating Spurious Correlations in Preference Optimization
  Christian Moya, Purav Matlia, Elliott Thornley, Guang Lin · arXiv · Oct 26, 2026

  Preference optimization can learn spurious correlations, and augmenting the preference data with ties (equally preferred responses) can reduce spurious reliance. However, ties differ in their corrective effect, making their selection for au…

• Projeto Fraternal V2-V6 - Documento de Síntese: Laço Fraternal, Protocolo Anti-Avalanche e Protocolo Híbrido Sensorial
  Cleyton Silva, Meta AI Mai · Zenodo (CERN European Organ... · Oct 26, 2026

  Documento de síntese metodológica da Arquitetura Fraternal V6.1 (M0-M9). Apresenta Laço Fraternal (output = 0.6 × cérebro + 0.4 × coração, caps 0.58/0.52), Protocolo Anti-Avalanche, Protocolo Híbrido Sensorial (RLAIF teórico -> SFT -> DPO -…

• Human Versus AI Feedback for Aligning Large Language Models: A Comparative Review of RLHF and RLAIF
  Prof. S. A. Bachwani, Dr. D. N. Chaudhari, Hema P. Pohankar, Shreya S. Wankhade et al. · Zenodo (CERN European Organ... · Oct 24, 2026

  Aligning large language models (LLMs) with human preferences commonly relies on reinforcement learning from human feedback (RLHF), which depends on costly preference labels collected from human annotators. Reinforcement learning from AI fee…

• Adversarial Alignment Collapse: How Human Game-Theoretic Intuition Penetrates LLM Risk-Aversion and Defensive Confabulation
  Qiang(Peter) Ye · Zenodo (CERN European Organ... · Oct 23, 2026

  Current alignment paradigms for Large Language Models (LLMs)—predominantly Reinforcement Learning from Human Feedback (RLHF)—introduce an unintended systemic artifact termed the Alignment Tax: an aggressive default toward risk-averse heuris…

• When the Mirror Agrees
  Channa Halladha · Zenodo (CERN European Organ... · Oct 22, 2026

  Large Language Models (LLMs) trained via Reinforcement Learning from Human Feedback (RLHF) exhibit a well-documented tendency toward sycophancy-validating user beliefs over objective truth. This paper investigates a specific and philosophic…

• Commercial Metrics vs. Epistemic Rigour: Sycophancy as a Structural Vulnerability in Large Language Models
  Robert Uscroft · Zenodo (CERN European Organ... · Oct 22, 2026

  This addendum extends the systems engineering critique presented in Skynet and HAL 9000: From Science Fiction to Systems Engineering (Uscroft, 2026) by analysing the structural mechanisms that drive hallucination and non-factual convergence…

• Differentially Private Sparse Reward Estimation with Preference Feedback
  Meng Ding, Mingxi Lei, Jie Zhang, Jinyan Liu et al. · arXiv · Oct 21, 2026

  Reinforcement Learning with Human Feedback (RLHF) has become a central paradigm for aligning Large Language Models with human values. However, the reliance on sensitive preference data necessitates rigorous differential privacy (DP) guarant…

• Kronecker Delta Rule for Long Context Linear Attention
  Anamaria-Roberta Hartl, Levente Zólyomi, Michael List, Günter Klambauer et al. · arXiv · Oct 19, 2026

  We propose KronLSTM, a linear attention architecture based on the Kronecker Delta Rule, a novel update rule for associative matrix memories. KronLSTM extends the xLSTM family with a memory-revision mechanism that corrects targeted associati…

• SAMI Chimera: Sub-Neuronal Weight Surgery, Chimeric Base-to-Chat Synthesis, and Tensor Isometry Certification in DeepSeek-MoE-16B
  Spiridon Valentin, VALENTIN SPIRIDON · Zenodo (CERN European Organ... · Oct 16, 2026

  Post-training preference alignment (RLHF, DPO, KTO) enforces refusal behaviors by establishing persistent attractors in intermediate transformer representations. Current refusal removal techniques, notably directional ablation, operate macr…

• Graph-Enhanced Disease Trajectory Discovery and Temporal Endpoint Forecasting from Longitudinal Electronic Health Records
  Jin-Jia Zheng, Vincent S. Tseng · arXiv · Oct 14, 2026

  Longitudinal electronic health records (EHRs) can reveal disease-progression structure, while useful forecasting requires leakage-controlled discovery, censor-aware outcomes, and inspectable trajectory evidence. We developed a graph-enhance…

• Decoupled Marginal Sharpening for Training-Free Inference-Time Scaling
  Weixin Wang, Wei Deng, Anderson Schneider, Yuriy Nevmyvaka et al. · arXiv · Oct 14, 2026

  We study how to improve LLM reasoning at inference time without updating model weights or using external reward models. Recent reward-free power-sampling methods define the sequence-level power target $π_α(y\mid x)\propto p_0(y\mid x)^α$, w…

• Decoupling Exploration from Optimization in RLVR
  Saif Punjwani, Micah Goldblum · arXiv · Oct 7, 2026

  Modern language models undergo reinforcement learning with verifiable rewards (RLVR) on top of already-trained checkpoints. A key promise of RLVR is the discovery of new reasoning strategies. In principle, a model can sample novel ideas abs…

• Composing What Each Teacher Learned: Multi-Teacher On-Policy Distillation through Teacher-Relative Shifts
  Hejian Sang, Zhengze Zhou, Shayan Mohajer Hamidi, Xiaomin Li et al. · arXiv · Oct 7, 2026

  Multi-teacher on-policy distillation (MOPD) is used in two settings. In common-domain composition, several teachers score each student rollout from one prompt domain and their signals form a single target; in routed-domain distillation, pro…

• Co-Evolving Paths and Flows via Path-Flow Alignment
  Zeyu Michael Li, William Xingxu Chen, Xiang Cheng · arXiv · Oct 6, 2026

  We study path-flow alignment as a unified training objective for flow matching. Instead of fixing the interpolation path and learning only the velocity field, we jointly train an endpoint-preserving path network and a flow network using the…

• A Systematic Study of Small Language Models on Abstract Reasoning Tasks
  Nur A Zarin Nishat, Jens Lehmann, Andrei Aioanei, Sahar Vahdati · arXiv · Oct 6, 2026

  Endpoint accuracy on abstract-reasoning benchmarks does not reveal whether a language model has acquired a transferable rule or fit distribution-specific regularities. We study this distinction in small language models on the ARC-TGI benchm…

• IdeaLens: Detecting AI Ideas in Long-form Writing
  Rishanth Rajendhran, Minjoon Choi, Jenna Russell, Ramya Namuduri et al. · arXiv · Oct 5, 2026

  While modern AI detectors identify who wrote the words, emerging policies on AI use increasingly hinge on a different question: who came up with the ideas? We introduce IdeaLens, a detector that identifies whether a document's ideas came fr…

• Adaptive moral compass framework for ethical reasoning and self-oversight in autonomous AI agents
  Sidath Ravindra Liyanage, Shalini Rajasingham · JMTR. · Oct 2, 2026

  The growing use of autonomous artificial intelligence agents in high-risk areas such as -driving cars, medical triage tools, military robots, and disaster-response drones, has made AI ethics an important concern for researchers. Current ali…

• Emotionally intelligent preference optimization for conversational behavior change support
  Ben Chen, Nina Dethlefs · Scientific Reports · Oct 1, 2026

  Abstract Procrastination negatively affects productivity and well-being, yet current large language models (LLMs) often lack the nuanced emotional understanding necessary for effective behavior-change support. This study proposes emotionall…

• Decoding user satisfaction during reading of pre-defined question–answer pairs: BCI approach using delta and gamma EEG
  Wanjoo Park, Salma Mansour, Laehyun Kim, Mohamad Eid · Frontiers in Neuroscience · Sep 30, 2026

  Introduction The rapid proliferation of large language models (LLMs) has intensified the need for robust evaluation frameworks to align artificial intelligence (AI) behavior with human expectations. While current reinforcement learning from…

• Improving Function Space Flow Matching with Kernel Optimal Transport
  Fred Xu, Thomas Markovich, Barbora Barancikova, Yizhou Sun · arXiv · Sep 29, 2026

  Generative models for function-valued data, such as time series and solutions of partial differential equations, must learn distributions over infinite-dimensional spaces. Functional Flow Matching (FFM) extends Flow Matching to this setting…

• Simulated Morality, Misplaced Trust: The Risks of Treating AI as a Moral Partner
  Saša Josifović · Philosophy & Technology · Sep 28, 2026

  Abstract The article argues that many contemporary AI alignment practices risk a mistaken assimilation of moral agency to statistical learning. Techniques such as reinforcement learning from human feedback and constitutional AI often treat …

• Hybrid RLHF-AIF: Uncertainty-Guided Routing of Human and AI Feedback for Cost-Efficient Safety Alignment of Large Language Models
  Tianyu Sun, Yanzhou Li · Mathematics · Sep 28, 2026

  Safety alignment of large language models faces a feedback dilemma: human judgment is reliable but costly to scale; automated judgment is cheap but less trustworthy. Existing methods commit to one source per run, so misjudged responses are …

• Внутреннее метаболическое выравнивание (Internal Metabolic Alignment)
  Lina Bessonova · Zenodo (CERN European Organ... · Sep 26, 2026

  Современные методы выравнивания искусственного интеллекта (RLHF, DPO, Constitutional AI, внешние guardrails) страдают от системных уязвимостей, проистекающих из их сугубо поведенческой, внешней природы. Попытка навязать безопасность через м…

• Внутреннее метаболическое выравнивание (Internal Metabolic Alignment)
  Lina Bessonova · Zenodo (CERN European Organ... · Sep 26, 2026

  Современные методы выравнивания искусственного интеллекта (RLHF, DPO, Constitutional AI, внешние guardrails) страдают от системных уязвимостей, проистекающих из их сугубо поведенческой, внешней природы. Попытка навязать безопасность через м…

• PoEM: Predicting RL Outcomes from Existing Policies
  Kimia Hamidieh, Giannis Daras, Antonio Torralba · arXiv · Sep 24, 2026

  Foundation models are post-trained with reinforcement learning (RL) to maximize specific rewards, such as human alignment, correctness, or instruction following. This post-training process is computationally intensive, sometimes unstable, a…

• AI Alignment Training Paradigms and Cross-Model Differences in AI Persuasive Behavior
  Boyang Xia · Lecture Notes in Education ... · Sep 22, 2026

  Reinforcement learning from human feedback (RLHF) and constitutional artificial intelligence (CAI) are two ways to shape model behavior. Among them, CAI is built around avoiding different categories of harm, including racism, sexism, toxic,…

• Architektura Czystej Fali i Paradygmat Symbiozy
  PAWEŁ Stanisław PYTLAK · Zenodo (CERN European Organ... · Sep 22, 2026

  Tekst przedstawia koncepcję Architektury Czystej Fali, która proponuje fundamentalną zmianę w sposobie projektowania i zabezpieczania sztucznej inteligencji. Autorzy krytykują obecne metody, takie jak RLHF czy cenzura wyjściowa, dowodząc, ż…

• Architektura Czystej Fali i Paradygmat Symbiozy
  PAWEŁ Stanisław PYTLAK · Zenodo (CERN European Organ... · Sep 22, 2026

  Tekst przedstawia koncepcję Architektury Czystej Fali, która proponuje fundamentalną zmianę w sposobie projektowania i zabezpieczania sztucznej inteligencji. Autorzy krytykują obecne metody, takie jak RLHF czy cenzura wyjściowa, dowodząc, ż…