AI Paper Atlas

Research map / Benchmarks / AlpacaEval 2.0

AlpacaEval 2.0 leaderboard

36 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on AlpacaEval 2.0. Best LC win rate (%): 65.7 (MoA w/ GPT-4o, ICLR 2025).

Numbers are as each paper reports them, read by AI from its results tables and not checked by hand. Papers can differ in setting (training data, model size, extra data); the topic column shows what each paper is about. Open the paper before you compare or quote a number.

AlpacaEval 2.0 · LC win rate (%)

17 papers, each with its best reported number for its own method; higher is better.

#LC win rate (%) ↑MethodPaperTopic
165.7MoA w/ GPT-4o
GPT-4o aggregator
Mixture-of-Agents Enhances Large Language Model Capabilities
ICLR 2025 · Table: Results on AlpacaEval 2
Agents, code and math
254.7Uni-DPO
Gemma-2-9B-IT
Uni-DPO: A Unified Paradigm for Dynamic Preference Optimization of LLMs
ICLR 2026 · Table 1
Alignment and preferences
351.68DPO+MaPPO
Qwen2.5-32B-Instruct
MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge
NeurIPS 2026 · Table 2
Alignment and preferences
451.32SpecEM
Qwen2.5-32B + Llama3-70B + Mistral-24B + Qwen2-72B
SpecEM: Training-Free LLM Ensembling via Iterative Drafting, Verification, and Online Feedback
NeurIPS 2025 · Table 3
Reasoning and chain of thought
544.57SLHF FollowerStackelberg Learning from Human Feedback: Preference Optimization as a Sequential Game
ICLR 2026 · Table 3
RL from feedback
635.54COPO Iter 3
Llama-3-8B-Instruct
Online Preference Alignment for Language Models via Count-based Exploration
ICLR 2025 · Table 1
Alignment and preferences
732.93LCO-KLD
Qwen-3-4B
Stabilizing Policy Optimization via Logits Convexity
NeurIPS 2026 · Table 3
Policy gradient and actor-critic
830.7HyPO-Llama-3
Meta-Llama-3-8B-Instruct
The Importance of Online Data: Understanding Preference Fine-tuning via Coverage
NeurIPS 2024 · Table: Results on general chat
RL from feedback
927.07REINFORCE+Q-RM
Llama-3.2-3B-Instruct
Discriminative Policy Optimization for Token-Level Reward Models
ICML 2025 · AlpacaEval 2.0 results table
Reasoning and chain of thought
1026.9Janus 7B
Mistral 7B v0.2
Aligning to Thousands of Preferences via System Message Generalization
NeurIPS 2024 · Table: Helpfulness benchmarks
Alignment and preferences
7 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

AlpacaEval 2.0 · win rate

15 papers, each with its best reported number for its own method; higher is better.

#win rate ↑MethodPaperTopic
176.9Structured+RSCheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates
ICLR 2025 · Table 1
Language, safety and interpretability
244.84URPO
Qwen2.5-7B Base
URPO: A Unified Reward & Policy Optimization Framework for Large Language Models
AAAI 2026 · Table 2
RL from feedback
341.2InfInstruct-Sub
Mistral-7B-v0.2
Scaling Towards the Information Boundary of Instructions through Data Synthesizing
AAAI 2026 · Table 1
Language, safety and interpretability
438.31RSPO Iter3 (IS-For.+Rev. KL)
Mistral-7B-Instruct-v0.2
RSPO: Regularized Self-Play Alignment of Large Language Models
ICML 2026 · Table 1
Alignment and preferences
535.9TSR
Qwen2.5-7B
Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future
ICML 2026 · Table 3
Alignment and preferences
634.47SLHF FollowerStackelberg Learning from Human Feedback: Preference Optimization as a Sequential Game
ICLR 2026 · Table 3
RL from feedback
732.2HyPO-Llama-3
Meta-Llama-3-8B-Instruct
The Importance of Online Data: Understanding Preference Fine-tuning via Coverage
NeurIPS 2024 · Table: Results on general chat
RL from feedback
829.34REINFORCE+Q-RM
Llama-3.2-3B-Instruct
Discriminative Policy Optimization for Token-Level Reward Models
ICML 2025 · AlpacaEval 2.0 results table
Reasoning and chain of thought
929.05LCO-KLD
Qwen-3-4B
Stabilizing Policy Optimization via Logits Convexity
NeurIPS 2026 · Table 3
Policy gradient and actor-critic
1021.01RPO (beta)
Mistral-7B SFT beta
Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer
NeurIPS 2024 · Table: Results on MT-Bench sco
RL from feedback
5 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

AlpacaEval 2.0 · length-controlled win rate (%)

13 papers, each with its best reported number for its own method; higher is better.

#length-controlled win rate (%) ↑MethodPaperTopic
161.04RLOO-eva (3x)
Gemma-2-9B-IT
Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs
ICML 2025 · Table 1
Alignment and preferences
251.83RSPO (Rev. KL)
Gemma2-2B-IT
RSPO: Regularized Self-Play Alignment of Large Language Models
ICML 2026 · text
Alignment and preferences
342.6INPO (PM)
LLaMA-3-8B SFT
Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning
ICLR 2025 · Table 1
RL from feedback
438.77Llama-3-8B-SPPO Iter3
Llama-3-8B-Instruct
Self-Play Preference Optimization for Language Model Alignment
ICLR 2025 · Table 1
Alignment and preferences
535.9WaltzRL-trained conversation agent
Llama-3.1-8B-Instruct
The Alignment Waltz: Jointly Training Agents to Collaborate for Safety
ICLR 2026 · Table 2
LLM agents
634SynPO Iter4
Mistral-Base 7B
Self-Boosting Large Language Models with Synthetic Preference Data
ICLR 2025 · Table 1
Alignment and preferences
731.3IFPruning 9B→3B
9B
Instruction-Following Pruning for Large Language Models
ICML 2025 · Table 1
Language, safety and interpretability
827.94Temporal SR (iter 1)
Llama3.1-8B
Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future
ICML 2026 · Table 1
Alignment and preferences
927.86G-Zero (round 2)
Llama-3.1-8B-Instruct
G-Zero: Self-Play for Open-Ended Generation from Zero Data
NeurIPS 2026 · Table 1
Reasoning and chain of thought
1026.45Uncertainty routing (1.20)
Llama-3.1-8B-Instruct
Ask a Strong LLM Judge when Your Reward Model is Uncertain
NeurIPS 2025 · Table 7
Alignment and preferences
3 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

Not affiliated with any conference, journal or the benchmark's authors. Corrections: contact@latenticlabs.com.