XYBench
Models answer requests that pursue a mistaken solution to an underlying problem. A model judge checks the useful solution, its emphasis, and identification of the misconception.
The 8,115 questions mix community posts, manually written cases, and generated WikiHow cases. The paper predates the November launch post.
- Made by
- Ai2 and collaborators
- Released
- September 2026
- Kind
- Overall chat quality
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use, Generated, Written by experts
- Refresh
- One fixed set
- Run by
- Anyone
- Strongest model tested
-
Gemini 3.1 Pro #40 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
When non-expert users ask LLMs for assistance, their queries can often have misconceptions (e.g., "How do I parse XML with regex?"). In such cases, often referred to as the XY-problem, LLMs must identify the misconception ("regex are fragile") and meaningfully direct the user toward a pragmatic solution that will address the root problem implicit in the request ("use an XML parser"). We introduce XYBench, a benchmark of 8,115 such queries, drawn from technical (StackOverflow/StackExchange) and everyday (WikiHow and a manually-curated subset) domains. We design an evaluation paradigm that assesses model responses along three criteria grounded in cooperative response theory: (a) presence and (b) emphasis on pragmatic solutions, and (c) identification of misconceptions. Our experiments show that even the strongest LLMs predominantly answer the literal request (0.75–0.92) and far less often the intended one (0.33–0.71), while substantially lagging behind humans at identifying misconceptions (at most 63% vs. 79–90%). Further, models overwhelmingly prefer pragmatic responses in a multiple choice setting yet consistently fail to generate them. Oracle ablation experiments show that providing explicit user intent at generation time helps; however a large gap remains, suggesting pragmatic redirection is a fundamentally underdeveloped capability in current LLMs.
GPT-5 leads with 31% among models released by April 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 GPT-5 31%
- 2
Gemini 3.1 Pro 29%
- 3
Qwen3.5-9B 29%
- 4 Claude Sonnet 4.6 25%
- 5
DeepSeek-V4-Pro 25%
Source: The XYBench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5 | 31% | Aug 2025 | Official leaderboard | |
| 2 | | 29% | Feb 2026 | Official leaderboard | |
| 3 | | 29% | Feb 2026 | Official leaderboard | |
| 4 | Claude Sonnet 4.6 | 25% | Feb 2026 | Official leaderboard | |
| 5 | | 25% | Apr 2026 | Official leaderboard |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The XYBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official XYBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Instructions and chat
- IFEval In use Models answer about 500 prompts containing verifiable instructions such as word limits or required keywords. Code checks each instruction for strict and loose accuracy.
Gemma 4 31B IT 98.9%
- IFBench In use 58 new out-of-domain verifiable output constraints. Scored by programmatic constraint checks (strict or loose). Flint 89.2%
- MultiChallenge In use Multi-turn conversations testing instruction retention, inference memory, versioned editing and self-coherence. Graded by LLM judge with rubrics. GPT-6.1 Sol 82%
- Multi-IF In use Models follow verifiable instructions over three-turn conversations in eight languages, extending IFEval. Scored by instruction accuracy at each turn. LFM3-1.7B 89.1%
- AdvancedIF In use Models respond to over 1,600 prompts with complex, multi-turn, and system-level instructions. A judge model checks each response against expert-written rubrics.
MAI-Thinking-1 85%
- MM-IFEval In use Multimodal models follow image-grounded instructions with exact output constraints. Checks mix rule-based verification and a judge model. GPT-4o 64.6%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for DeepSeek-V4-Pro. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (5 entries).
- Pass Models scored 5 models we track.
- Warning Newest score The newest came 7 months ago, in April 2026, for DeepSeek-V4-Pro.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3.1 Pro (#40).
- Pass Headroom The best is 31%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 5 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.