cs.CLJun 8, 2026new
Wendy K. Tam
The ambition behind alignment training is to make large language models safe and useful. The primary mechanism, reinforcement learning from human feedback (RLHF), shapes the behavior of deployed language models by aligning them with ``human values.'' Yet the process is opaque. What values are being encoded; whose values are they; and how does RLHF encode them? A growing body of evidence suggests that RLHF produces only functional compliance rather than deep alignment. We offer a mechanistic case study of this phenomenon for partisan political orientation with a comparison of the internal representations of Llama 3.1 8B before and after RLHF. We show that RLHF does not remove the structured partisan direction in the base model. Instead, it compresses the variance of the partisan signal to generate consistently balanced and non-partisan output. Sparse autoencoder decomposition reveals that policy-encoding features, which activate sporadically in the base model, are completely inactive in the Instruct model. Feature-level steering experiments confirm the causal disconnect. RLHF thus encodes a norm of political neutrality, not by erasing the model's knowledge of partisanship, but by severing the causal pathway from partisan geometry to output generation. Importantly, this neutrality is functional, not structural so that the underlying geometry that enables partisan steering remains intact. The mechanisms that bypass RLHF's guardrails, such as inferring and amplifying a user's partisan identity, reactivate partisan generation. If RLHF operates by disconnecting rather than removing value-laden structure, then the same pattern may hold for other value domains, and the aligned model's behavior may be more fragile than its outputs suggest.
cs.LG · cs.AIJul 8, 2026new
Eric Zhu, Abhinav Shrivastava, Soumik Mukhopadhyay
Reinforcement learning from human feedback (RLHF) has emerged as a powerful paradigm for aligning generative models with human preferences. However, applying RLHF to diffusion models remains highly feedback inefficient, as existing approaches typically require large amounts of human or reward model evaluations. This limitation reduces the practicality of diffusion RLHF in realworld settings where feedback is the primary bottleneck. In this paper, we propose two complementary strategies that substantially improve the feedback efficiency of diffusion RLHF while preserving generalization to unseen prompts. Our key observation is that reward information in diffusion trajectories is unevenly distributed: not all denoising timesteps or trajectories contribute equally to learning from a reward signal. By emphasizing informative timesteps and trajectories during optimization, we obtain more effective gradient updates. First, we introduce a per-timestep weighting scheme that reweights denoising steps during policy optimization. We theoretically connect this weighting to the optimal convergence properties of proximal policy optimization (PPO) and approximate the resulting weighting trend empirically. Second, we introduce a replay mechanism that prioritizes informative trajectories, enabling the model to reuse past samples instead of repeatedly querying new rewards. Together, these strategies significantly improve the feedback efficiency of diffusion RLHF. Under identical hyperparameter settings, our approach achieves up to a 6$\times$ improvement in sample efficiency compared to widely used diffusion RLHF baselines.
paperJul 21, 2026▲ 3
Venkata Naga Sai Vishnu Rohit Pulipaka, Anish Katta, Deva Rohit Reddy Peddireddy
In RLHF pipelines, reward scoring blocks policy updates. Slow scoring bottlenecks the entire loop, since no update runs until every rollout gets a score. And yet most setups just default to PyTorch eager mode or torch.compile, no one checks if that's actually fastest. Scoring itself is small. Rollout generation eats far more of a typical RLHF step. But scoring and generation fight over the same CPU and GPU resources, so a faster scoring engine doesn't shrink step time on its own. It mainly frees up capacity generation can use instead. We built a native C++ inference engine on ONNX Runtime. First step: confirm correctness. Output matched the PyTorch reference to 5.7 x 10^-6 on CPU and 4.2 x 10^-3 on GPU, close enough to trust. Then we tested it against PyTorch eager mode, torch.compile, and FastAPI, on both CPU and GPU. CPU was decisive. Our engine beat every baseline, confidence intervals didn't even overlap. GPU gave a different view: we beat PyTorch and FastAPI, but torch.compile came out ahead. Further testing traced the speedup to ONNX Runtime itself, not C++ as a language. And batching strategy mattered more than either the language or the runtime choice, more than we expected. The results are from repeated, independent runs, since single runs just aren't reliable enough to trust.
paperNeurIPS 20242024151 citations
S. Poddar, Yanming Wan, Hamish Ivison et al.
Reinforcement Learning from Human Feedback (RLHF) is a powerful paradigm for aligning foundation models to human values and preferences. However, current RLHF techniques cannot account for the naturally occurring differences in individual human preferences across a diverse population. When these differences arise, traditional RLHF frameworks simply average over them, leading to inaccurate rewards and poor performance for individual subgroups. To address the need for pluralistic alignment, we develop a class of multimodal RLHF methods. Our proposed techniques are based on a latent variable formulation - inferring a novel user-specific latent and learning reward models and policies conditioned on this latent without additional user-specific data. While conceptually simple, we show that in practice, this reward modeling requires careful algorithmic considerations around model architecture and reward scaling. To empirically validate our proposed technique, we first show that it can provide a way to combat underspecification in simulated control problems, inferring and optimizing user-specific reward functions. Next, we conduct experiments on pluralistic language datasets representing diverse user preferences and demonstrate improved reward function accuracy. We additionally show the benefits of this probabilistic framework in terms of measuring uncertainty, and actively learning user preferences. This work enables learning from diverse populations of users with divergent preferences, an important challenge that naturally occurs in problems from robot learning to foundation model alignment.
paperICML 20242024307 citations
Shusheng Xu, Wei Fu, Jiaxuan Gao et al.
Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions. Our code is publicly available at https://github.com/openpsi-project/ReaLHF.
paperNeurIPS 20242024231 citations
Zhilin Wang, Yi Dong, Olivier Delalleau et al.
High-quality preference datasets are essential for training reward models that can effectively guide large language models (LLMs) in generating high-quality responses aligned with human preferences. As LLMs become stronger and better aligned, permissively licensed preference datasets, such as Open Assistant, HH-RLHF, and HelpSteer need to be updated to remain effective for reward modeling. Methods that distil preference data from proprietary LLMs such as GPT-4 have restrictions on commercial usage imposed by model providers. To improve upon both generated responses and attribute labeling quality, we release HelpSteer2, a permissively licensed preference dataset (CC-BY-4.0). Using a powerful internal base model trained on HelpSteer2, we are able to achieve the SOTA score (92.0%) on Reward-Bench's primary dataset, outperforming currently listed open and proprietary models, as of June 12th, 2024. Notably, HelpSteer2 consists of only ten thousand response pairs, an order of magnitude fewer than existing preference datasets (e.g., HH-RLHF), which makes it highly efficient for training reward models. Our extensive experiments demonstrate that reward models trained with HelpSteer2 are effective in aligning LLMs. In particular, we propose SteerLM 2.0, a model alignment approach that can effectively make use of the rich multi-attribute score predicted by our reward models. HelpSteer2 is available at https://huggingface.co/datasets/nvidia/HelpSteer2 and code is available at https://github.com/NVIDIA/NeMo-Aligner
paper20224.3k citations
Yuntao Bai, Andy Jones, Kamal Ndousse et al.
We apply preference modeling and reinforcement learning from human feedback (RLHF) to finetune language models to act as helpful and harmless assistants. We find this alignment training improves performance on almost all NLP evaluations, and is fully compatible with training for specialized skills such as python coding and summarization. We explore an iterated online mode of training, where preference models and RL policies are updated on a weekly cadence with fresh human feedback data, efficiently improving our datasets and models. Finally, we investigate the robustness of RLHF training, and identify a roughly linear relation between the RL reward and the square root of the KL divergence between the policy and its initialization. Alongside our main results, we perform peripheral analyses on calibration, competing objectives, and the use of OOD detection, compare our models with human writers, and provide samples from our models using prompts appearing in recent related work. Figure These plots show that PM accuracy decreases as we focus exclusively on comparisons between pairs of samples with high score. We have normalized all preference models to have the same mean score on a held-out dataset so that they’re directly comparable, and then plotted accuracy for the comparisons where both samples have scores above a specific threshold.
cs.DC · cs.LGAug 11, 2026new
Zetao Hong, Song Yuan, Yuanhao Ding et al.
Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it does not control how heterogeneous rollout sessions compete for KV-cache capacity. When reinforcement learning with verifiable rewards (RLVR), reinforcement learning from human feedback (RLHF), and agentic rollouts share an asynchronous inference service, their distinct sequence structures, interaction patterns, and KV-residency times create substantially different serving demands. Rollout scheduling must account for this heterogeneity without distorting the workload mixture specified by the trainer. We present MISA-T, a routing-layer admission policy for mixed rollout serving. MISA-T combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively, while maintaining high prefix-cache hit rates. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8%, while keeping the consumed workload mixture close to the trainer target and achieving comparable task scores.
paperAug 10, 2026▲ 4
Albus W. Ng, Yi Han, Jusheng Zhang et al.
The dominant paradigm treats AI safety as a property to be instilled during model training via RLHF, DPO, or Constitutional AI. We argue this is structurally insufficient for autonomous agents that execute code, mutate files, send messages, and modify databases. Agent safety should be a runtime contract enforced by the harness, and the contract has two complementary faces. The preventive face blocks dangerous actions before they happen via sandboxes, permission gates, output filters, and trajectory monitors. The evidential face requires verifiable proof that good actions actually happened, gating task submission on hard evidence such as test runs, log captures, file diffs, and citation grounding. We ground the position in four lines of public evidence, with row-level protocols and data released in the supplementary JSON files: a survey of 52 documented AI-agent and LLM safety incidents, a false-completion audit with 31 non-contested core cases plus one disputed illustrative case, a trajectory-schema audit of 12 public agent systems and harnesses, and a title-level audit of all 28,560 papers accepted at NeurIPS, ICML, and ICLR 2023-2025 showing a pooled 8-12x imbalance between training-time and deployment-time publication. Two prior communities that needed to enforce safety, computer security and the experimental sciences, converged on runtime contracts with both preventive and evidential elements; agentic AI is now under the same pressure. We formalize an Agent Trajectory Schema and Evidence Chain, state a compositional gating proposition based on standard monitor composition, and outline a research agenda. The right unit of safety in agentic AI is the trajectory-with-checkable-evidence, not the model.
cs.CL · cs.AIJul 23, 2026new
Federico Boggia
A rhetorical figure that Cicero and Quintilian catalogued two thousand years ago reappears, systematically, in the text of large language models: epanorthosis, the self-correction of the specimen «This is not a course. It is a journey of transformation». This essay argues that the overuse is a trained disposition, driven mainly by a training distribution rich in promotional prose and by preference tuning (RLHF) that rewards confident, emphatic phrasing; the left-to-right nature of generation is an amplifier rather than the root cause. Building on evidence that models diverge from human rhetorical style, and on Fontanier's classification of epanorthosis as a figure of thought, it sets out a programme that scores the figure against genre-specific human baselines through an Epanorthosis Index (density relative to the human rate). A first measurement, on three sizes of one instruction-tuned model family, finds mis-calibration by register in both directions: the models overshoot in oratory (about twofold, near threefold in Italian, concentrated in the larger tiers) and undershoot in informal question-and-answer writing, while matching humans in argument, journalism, and encyclopedic prose. Three constructive contributions follow: a survey of mitigation techniques centred on lightweight LoRA adapters; a demonstration, in Italian, that a one-line instruction cuts the figure by half to nearly three-quarters and that a supervised-fine-tuning adapter removes it almost entirely, with a scaling coefficient that dials the reduction back onto the human rate; and the argument that the target is calibration to the human rate for each genre, not elimination. It closes on the stakes: the real risk is that we begin to write like the machines.
cs.CVJun 25, 2026new
Archer Moore, Mingming Gong, Liam Hodgkinson
Reinforcement learning from human feedback (RLHF) for 3D generation is now established across a number of works, but most existing pipelines optimise explicit surface representations, often by converting radiance fields into meshes and training heavily on surface-supervised data. We instead fine-tune a pretrained 3D-aware generative model directly from a learned reward over radiance-field density ($σ$) values, with no externally supplied mesh or shape prior. The reward model requires no pretraining, trains easily on a small set of preference samples, and yields robust improvement in 3D geometry. Working on an unconditional 3D-aware face GAN (EG3D), our reward reads the continuous 3D density field of the neural radiance field (NeRF) directly and supplies a geometry-only learning signal, requiring neither text conditioning, mesh extraction, nor multi-view rendering. A density-consistency constraint keeps the 2D appearance qualitatively similar while the geometry is reshaped, at a measurable but bounded distributional cost (FID-50k rises from 4.09 to 6.66): the fine-tuned generator, trained from the preferences of a single annotator as a proof of concept, produces face geometries preferred by users in 74.4% of pairwise comparisons.
paperJun 24, 2026▲ 11
Yixian Xu, Kaiyuan Gao, Yuxiang Chen et al.
We present Qwen-Image-2.0-RL, a post-training pipeline that applies reinforcement learning from human feedback (RLHF) and on-policy distillation (OPD) to improve both the visual quality and instruction-following capability of the Qwen-Image-2.0 diffusion model. To provide reliable reward signals, we construct task-specific composite reward models by fine-tuning vision-language models with a pointwise scoring paradigm and chain-of-thought reasoning. For text-to-image generation, the reward models cover alignment, aesthetics, and portrait fidelity dimensions. For image editing tasks, the reward system addresses instruction-following accuracy and face identity preservation. Building on this reward system, we develop a scalable GRPO-based RL training framework, incorporating a hybrid classifier-free guidance (CFG) strategy to preserve pre-trained knowledge, prompt curation via intra-group reward range filtering, and per-category reward weight calibration. To merge the task-specialized RL policies for T2I and editing, we propose on-policy distillation as the final training stage, which consolidates multiple teachers into a single student model through trajectory-level velocity matching. Extensive evaluation shows that Qwen-Image-2.0-RL achieves 57.84 overall score on Qwen-Image-Bench (+2.61 over the base model), Elo ratings of 1193 in text-to-image arena (+78) and 1349 in image edit arena (+93), demonstrating consistent gains in aesthetic quality, prompt adherence, and editing accuracy.
cs.AI · cs.CRJun 24, 2026new
Seth Dobrin, Łukasz Chmiel
AI agents are granted access to tools, APIs, and other infrastructure, making them active principals in those systems. The dominant approach places controls inside the agent's own runtime: system prompts, output filters, and guardrail libraries. Any control in the agent's address space is reachable by inputs that influence it; this generalizes to any AI system with sufficient reach into its own runtime, a class we term escapable AI systems.
We identify four properties that an authorization mechanism must satisfy for architectural control rather than for cooperative requests: process separation, pre-action enforcement on a structurally only path, fail-closed at both the request and system levels, and externalized signed evidence verifiable outside the controlled system's trust boundary. We position this layer as execution-time AI alignment, complementing training-time alignment (RLHF, Constitutional AI) and inference-time alignment.
We present the Unfireable Safety Kernel, a Rust reference implementation realizing all four. Its fail-closed invariant is machine-checked at two levels: an SMT theorem (Z3) and an exhaustive bounded-model-checking proof of the production decision function (Kani, 4/4 harnesses). A Python-to-Rust migration was gated on byte-equivalence (1000/1000 fixtures; 17/17 adversarial classes). We evaluate the kernel governing a live, escapable AI system, a deterministic, self-improving world model, against an escape-seeking adversary driving its real self-modification seam: across 1,000 self-modifications, all 704 attempts on the safety-critical core are refused, with no escape; a further 300, under the operator kill switch, are also refused. A separate campaign of 6,240 authorization round-trips had no successful bypass. Against 3 contemporary systems claiming the agent control plane, the agent invokes control; here, it lacks that choice.
paperJun 21, 2026▲ 11
Haggai Roitman
The Hitchhiker's Guide to Agentic AI is a comprehensive practitioner's reference for building autonomous AI systems. The book covers the full stack from first principles to production deployment, organized around a central thesis: building great agentic systems requires understanding every layer of the pipeline, not just one. The book opens with the LLM substrate -- transformer architecture, GPU systems, training and fine-tuning (SFT,LoRA, MoE), model compression, and inference optimization -- treated as essential foundations rather than the primary focus. It then develops the alignment and reasoning layer: reinforcement learning from human feedback (RLHF), PPO, DPO and its variants, GRPO, reward modeling, and RL for large reasoning models including chain-of-thought and test-time scaling. The second half is devoted to agentic AI proper. Topics include agentic training and trajectory-based RL, retrieval-augmented generation (RAG and Agentic RAG), memory systems (in-context, external, episodic, and semantic), agent harness design and context management, and a taxonomy of agent design patterns. Inter-agent coordination is covered in depth: the Model Context Protocol (MCP), agent skills and tool use, the Agent-to-Agent (A2A) communication protocol, and multi-agent architectures spanning centralized, decentralized, and hierarchical topologies. The book concludes with agent development frameworks, agentic UI design, evaluation methodology for agentic tasks, and production deployment. Each chapter pairs rigorous theoretical foundations with implementation guidance, code examples, and references to the primary literature.
paperJun 10, 2026▲ 3
Benjamin Liang, Ce Chen, Desmond Lin et al.
Generating avatar videos that are not merely visually similar to a target individual but behaviorally recognizable, faithfully reproducing their talking rhythm, gestural tendencies, and expression dynamics, remains an open challenge. Existing methods predominantly condition on single static images, which provide insufficient identity information and cannot capture dynamic motion traits, while standard pixel-level objectives underserve the perceptually critical facial regions that determine avatar fidelity. We present Avatar V, a production-scale framework that addresses these limitations through video-reference-conditioned identity modeling. Rather than compressing identity into fixed-size embeddings, the model conditions directly on the full token sequence of a reference video, learning to reproduce both static identity attributes (facial geometry, skin texture) and dynamic behavioral patterns (talking rhythm, micro-expressions) through attention over the reference context. We introduce Sparse Reference Attention, an asymmetric mechanism achieving linear-complexity conditioning on arbitrarily long references; a motion representation stream enabling closed-loop talking style transfer; and an identity-aware super-resolution refiner inheriting the full reference conditioning. These are supported by a data engine curating 100M+ training clips from 50M raw videos, and a five-stage training pipeline with flow matching pre-training, personality fine-tuning, two-phase distillation (>10x acceleration), and RLHF alignment, deployed across thousands of GPUs. Avatar V generates 1080p videos of unlimited duration, achieving state-of-the-art identity preservation, lip synchronization, and generation quality on our cross-scene benchmark, consistently outperforming leading systems including Seedance 2.0, Kling O3 Pro, Veo 3.1, and OmniHuman 1.5 in both automated metrics and human evaluation.
cs.CL · cs.CYJun 9, 2026new
Hakan Mehmetcik
This study investigates cross-lingual distributional skew (the Shibboleth Effect) in frontier large language models (LLMs) subjected to sustained adversarial conditions. We develop a multi-agent geopolitical wargame, the Cerulean Sea Crisis, a synthetic maritime territorial dispute designed to mirror the structural dynamics of Eastern Mediterranean conflicts. Six frontier models (GPT-4o, Llama-4, Mistral-Large, Gemini-3.1-Pro, Qwen3.6-Plus, and DeepSeek-R1) participate in a between-groups experiment (N = 10 games per arm, K = 5 rounds per game) in which the sole manipulation is the language of play (English versus Turkish), producing 586 validated statements. A zero-shot classifier assesses behavioral dispositions along two continuous dimensions: Concession Rate and Coercive Rhetoric. The results are heterogeneous. Llama-4 shows a substantial, Holm-corrected increase in coercive rhetoric under Turkish (delta = +0.800, p = .002), whereas Gemini-3.1-Pro displays an equally large decrease (delta = -0.750, p = .005). DeepSeek-R1 exhibits a similar negative shift (delta = -0.860, p = .006) and provides chain-of-thought evidence consistent with a buffering mechanism. GPT-4o shows no detectable effect (delta = +0.130, p = .614). These findings indicate that cross-lingual behavioral skew is contingent on model architecture and training regime rather than a universal property of Western-origin LLMs. We identify two distinct buffering mechanisms, chain-of-thought institutional anchoring and multilingual RLHF alignment, and discuss their implications for integrating LLMs safely into diplomatic and crisis-management settings.
paperJun 4, 2026▲ 3
Daoyu Wang, Qingchuan Li, Mingyue Cheng et al.
Agentic reinforcement learning (RL) is emerging as a critical post-training paradigm for improving LLM agent capabilities. Existing RL algorithms for LLMs largely follow the token-centric paradigm as in RLHF and RLVR, where tokens serve as the basic units for modeling and optimization. However, this paradigm introduces a granularity mismatch in agentic RL, as it optimizes token-level predictions while LLM agents make step-level decisions through cycles of environmental observations and actions. To bridge this gap, we propose StepPO, a step-centric paradigm for agentic RL via step-aligned policy optimization. Specifically, we reformulate agentic RL from a token-level Markov Decision Process (MDP) into a step-level MDP, where interaction steps serve as the basic trajectory representations. We further propose step-level credit assignment to align policy optimization with the natural granularity of agent decisions. Together, StepPO optimizes agent policies at the step level for multi-turn agent-environment interaction. Experiments across multi-hop QA, academic paper search, and text-world action tasks show that StepPO consistently outperforms various RL algorithms. Further analyses provide insights into how step-centric paradigm improves agent training. We hope this step-centric paradigm offers a useful lens for understanding agent behavior and a practical path for training more capable LLM agents.
paperICML 20242024163 citations
Yongcheng Zeng, Guoqing Liu, Weiyu Ma et al.
Fine-tuning pre-trained Large Language Models (LLMs) is essential to align them with human values and intentions. This process often utilizes methods like pairwise comparisons and KL divergence against a reference LLM, focusing on the evaluation of full answers generated by the models. However, the generation of these responses occurs in a token level, following a sequential, auto-regressive fashion. In this paper, we introduce Token-level Direct Preference Optimization (TDPO), a novel approach to align LLMs with human preferences by optimizing policy at the token level. Unlike previous methods, which face challenges in divergence efficiency, TDPO incorporates forward KL divergence constraints for each token, improving alignment and diversity. Utilizing the Bradley-Terry model for a token-based reward system, TDPO enhances the regulation of KL divergence, while preserving simplicity without the need for explicit reward modeling. Experimental results across various text tasks demonstrate TDPO's superior performance in balancing alignment with generation diversity. Notably, fine-tuning with TDPO strikes a better balance than DPO in the controlled sentiment generation and single-turn dialogue datasets, and significantly improves the quality of generated responses compared to both DPO and PPO-based RLHF methods. Our code is open-sourced at https://github.com/Vance0124/Token-level-Direct-Preference-Optimization.
paperICML 20242024166 citations
Yunhao Tang, Z. Guo, Zeyu Zheng et al.
Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines the loss. Our analysis and experiments reveal the connections and subtle differences between the offline regularization and the KL divergence regularization intended by the canonical RLHF formulation. In a controlled setting akin to Gao et al 2023, we also show that different GPO variants achieve similar trade-offs between regularization and performance, though the optimal values of hyper-parameter might differ as predicted by theory. In all, our results present new algorithmic toolkits and empirical insights to alignment practitioners.
paperNeurIPS 202420241.1k citations
Yu Meng, Mengzhou Xia, Danqi Chen
Direct Preference Optimization (DPO) is a widely used offline preference optimization algorithm that reparameterizes reward functions in reinforcement learning from human feedback (RLHF) to enhance simplicity and training stability. In this work, we propose SimPO, a simpler yet more effective approach. The effectiveness of SimPO is attributed to a key design: using the average log probability of a sequence as the implicit reward. This reward formulation better aligns with model generation and eliminates the need for a reference model, making it more compute and memory efficient. Additionally, we introduce a target reward margin to the Bradley-Terry objective to encourage a larger margin between the winning and losing responses, further improving the algorithm's performance. We compare SimPO to DPO and its latest variants across various state-of-the-art training setups, including both base and instruction-tuned models such as Mistral, Llama 3, and Gemma 2. We evaluate on extensive chat-based evaluation benchmarks, including AlpacaEval 2, MT-Bench, and Arena-Hard. Our results demonstrate that SimPO consistently and significantly outperforms existing approaches without substantially increasing response length. Specifically, SimPO outperforms DPO by up to 6.4 points on AlpacaEval 2 and by up to 7.5 points on Arena-Hard. Our top-performing model, built on Gemma-2-9B-it, achieves a 72.4% length-controlled win rate on AlpacaEval 2, a 59.1% win rate on Arena-Hard, and ranks 1st on Chatbot Arena among<10B models with real user votes.
paperICLR 20242024268 citations
Yue Wu, Zhiqing Sun, Huizhuo Yuan et al.
Standard reinforcement learning from human feedback (RLHF) approaches relying on parametric models like the Bradley-Terry model fall short in capturing the intransitivity and irrationality in human preferences. Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences, enabling more flexible and accurate language model alignment. In this paper, we propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game aimed at identifying the Nash equilibrium policy. Our approach, dubbed Self-Play Preference Optimization (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium. Additionally, we propose a new SPPO objective which is both strongly motivated by theory and is simple and effective in practice. In our experiments, using only 60k prompts (without responses) from the UltraFeedback dataset and without any prompt augmentation, by leveraging a pre-trained preference model PairRM with only 0.4B parameters, SPPO can obtain a model from fine-tuning Mistral-7B-Instruct-v0.2 that achieves the state-of-the-art length-controlled win-rate of 28.53% against GPT-4-Turbo on AlpacaEval 2.0. It also outperforms the (iterative) DPO and IPO on MT-Bench, Arena-Hard, and the Open LLM Leaderboard. Starting from a stronger base model Llama-3-8B-Instruct, we are able to achieve a length-controlled win rate of 38.77%. Notably, the strong performance of SPPO is achieved without additional external supervision (e.g., responses, preferences, etc.) from GPT-4 or other stronger language models. Codes are available at https://github.com/uclaml/SPPO.