AI Paper Atlas

Research map / Benchmarks / Arena-Hard

Arena-Hard leaderboard

32 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on Arena-Hard. Best win rate (%): 87 (RLOO w/ English RM, NeurIPS 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.

Arena-Hard · win rate (%)

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

#win rate (%) ↑MethodPaperTopic
187RLOO w/ English RM
Llama-3.3-70B-Instruct
HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages
NeurIPS 2025 · Table, Performance of Aligned
Alignment and preferences
276.71MC-PO
Llama-3.1-8B-Instruct
Preference Optimization via Contrastive Divergence: Your Policy Is Secretly an NLL Estimator
AAAI 2026 · Table 1
RL from feedback
362.4RLOO-eva (3x)
Gemma-2-9B-IT
Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs
ICML 2025 · Table 1
Alignment and preferences
461.5GAPO
Gemma-2-Instruct (9B)
Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment
NeurIPS 2026 · Table 1
Alignment and preferences
560.9SimPO +ours-10k
Gemma-2-9b-it
Larger or Smaller Reward Margins to Select Preferences for LLM Alignment?
ICML 2025 · Table 1
Alignment and preferences
660.1RePO
Qwen3-4B-Base
A Regret Minimization Framework on Preference Learning in Large Language Models
ICML 2026 · Table 1
Alignment and preferences
758.4MetaAPO
Qwen2.5-7B
Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization
ICLR 2026 · Table 1
Alignment and preferences
857.2Llama3.1-8B-Instruct WSPO
Llama3.1-8B-Instruct
Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model
ICLR 2025 · Llama results table
Alignment and preferences
956.5OnlineRubrics-pi_old
Qwen2.5-7B-Instruct
Online Rubrics Elicitation from Pairwise Comparisons
ICML 2026 · Table 2
Alignment and preferences
1055.6Gemma-2-9B-it-MoAA-DPO
Gemma-2-9B-it
Improving Model Alignment Through Collective Intelligence of Open-Source Models
ICML 2025 · Table 1
Alignment and preferences
8 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

Arena-Hard · score

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

#score ↑MethodPaperTopic
188.44JiSi (Ours)Beyond Gemini-3-Pro: Revisiting LLM Routing and Aggregation at Scale
ICML 2026 · Table 1
Agents, code and math
285REINFORCE
Llama-3.1-70B-Instruct
HelpSteer2-Preference: Complementing Ratings with Preferences
ICLR 2025 · Aligned Models results table
Alignment and preferences
384.92Ultra-253B-Base
Llama-3.1-405B-Instruct
FFN Fusion: Rethinking Sequential Computation in Large Language Models
NeurIPS 2025 · Table: Comparison of Ultra-253
Attention and transformers
482.9RIP
Llama 3.3-70B-Instruct
R.I.P.: Better Models by Survival of the Fittest Prompts
ICML 2025 · Table: RIP on WildChat with Ll
Alignment and preferences
567.57GPT-Conversation*
GPT-Conversation
CoLM: Collaborative Large Models via a Client-Server Paradigm
AAAI 2026 · Table 2
Personalized and heterogeneous FL
663COMAL
Qwen2.5-7B-SFT
COMAL: A Convergent Meta-Algorithm for Aligning LLMs with General Preferences
ICLR 2026 · Table 3
RL from feedback
761.3AdvGame-DPO-MD
Qwen2.5-7B
Safety Alignment of LMs via Non-cooperative Games
ICML 2026 · Table 1
Attacks on text and graphs
859.5SSR-D
Qwen3-8B-Think
Measuring and Mitigating Post-Hoc Rationalization in Reverse Chain-of-Thought Generation
ICML 2026 · Table 3
Reasoning and chain of thought
942.5Iterative DPO w. DA
Llama 3 70B
HelpSteer 2: Open-source dataset for training top-performing reward models
NeurIPS 2024 · Aligned Models results table
Alignment and preferences
1032.6MAmmoTH2-Plus-8x7B
Mixtral-8x7B
MAmmoTH2: Scaling Instructions from the Web
NeurIPS 2024 · Additional benchmarks table
Instruction and fine-tuning
4 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.