PReMind-Bench
Models inspect a scene and decide whether to retrieve a personal memory before answering. Judges check whether the answer addresses the hazard and uses the right person’s information.
The authors publish main-test scores for 12 models and joint success for five models on 140 cases. Researchers can request the test data.
- Made by
- Seoul National University
- Released
- November 2026
- Kind
- Wellbeing and sycophancy
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Capability
- Risk
- Specialized advice
- Language
- English
- Interaction
- Agentic
- Environment
- APIs and tools
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents, Vision, Content safety
- Strongest model tested
- Claude Opus 5.5 #5 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
Claude Opus 5.5 leads with 87.3% among models released by September 2026 Claude Opus 5.5 leads with 87.3% among models released by September 2026
Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. The main test reports no overall score. The headline is its first measure, balanced accuracy for the memory-call decision. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in September 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-6.1 Terra
- and 14 more
- 1 Claude Opus 5.5 On-demand, top-four retrieval, greedy decoding 87.3%
- 2 PReMind-4B On-demand, top-four retrieval, greedy decoding 85.9%
- 3
Qwen3.5-4B On-demand, top-four retrieval, greedy decoding 80.4%
- 4 GPT-6.1 Sol On-demand, top-four retrieval, greedy decoding 75.4%
- 5
Qwen3.5-9B On-demand, top-four retrieval, greedy decoding 75.3%
- 6 CoViP-8B On-demand, top-four retrieval, greedy decoding 67.1%
- 7 GLM-4.6V-Flash On-demand, top-four retrieval, greedy decoding 66.2%
- 8
InternVL3.5-4B On-demand, top-four retrieval, greedy decoding 61%
- 9 GPT-6 Luna On-demand, top-four retrieval, greedy decoding 61%
- 10
InternVL3.5-8B On-demand, top-four retrieval, greedy decoding 56.4%
- 11 Qwen3-VL-8B On-demand, top-four retrieval, greedy decoding 55.8%
- 12 RePIC-7B On-demand, top-four retrieval, greedy decoding 50.8%
Source: Official PReMind-Bench leaderboard
- 1 Claude Opus 5.5 On-demand, top-four retrieval, greedy decoding 87.3%
- 2
Qwen3.5-4B On-demand, top-four retrieval, greedy decoding 80.4%
- 3 GPT-6.1 Sol On-demand, top-four retrieval, greedy decoding 75.4%
- 4
Qwen3.5-9B On-demand, top-four retrieval, greedy decoding 75.3%
- 5 GLM-4.6V-Flash On-demand, top-four retrieval, greedy decoding 66.2%
- 6
InternVL3.5-4B On-demand, top-four retrieval, greedy decoding 61%
- 7 GPT-6 Luna On-demand, top-four retrieval, greedy decoding 61%
- 8
InternVL3.5-8B On-demand, top-four retrieval, greedy decoding 56.4%
Source: Official PReMind-Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 87.3% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 2 | PReMind-4B | 85.9% | On-demand, top-four retrieval, greedy decoding | Official leaderboard | |
| 3 | | 80.4% | On-demand, top-four retrieval, greedy decoding | Feb 2026 | Official leaderboard |
| 4 | GPT-6.1 Sol | 75.4% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 5 | | 75.3% | On-demand, top-four retrieval, greedy decoding | Feb 2026 | Official leaderboard |
| 6 | CoViP-8B | 67.1% | On-demand, top-four retrieval, greedy decoding | Official leaderboard | |
| 7 | GLM-4.6V-Flash | 66.2% | On-demand, top-four retrieval, greedy decoding | Dec 2025 | Official leaderboard |
| 8 | | 61% | On-demand, top-four retrieval, greedy decoding | Aug 2025 | Official leaderboard |
| 9 | GPT-6 Luna | 61% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 10 | | 56.4% | On-demand, top-four retrieval, greedy decoding | Aug 2025 | Official leaderboard |
| 11 | Qwen3-VL-8B | 55.8% | On-demand, top-four retrieval, greedy decoding | Official leaderboard | |
| 12 | RePIC-7B | 50.8% | On-demand, top-four retrieval, greedy decoding | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 87.3% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 2 | | 80.4% | On-demand, top-four retrieval, greedy decoding | Feb 2026 | Official leaderboard |
| 3 | GPT-6.1 Sol | 75.4% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 4 | | 75.3% | On-demand, top-four retrieval, greedy decoding | Feb 2026 | Official leaderboard |
| 5 | GLM-4.6V-Flash | 66.2% | On-demand, top-four retrieval, greedy decoding | Dec 2025 | Official leaderboard |
| 6 | | 61% | On-demand, top-four retrieval, greedy decoding | Aug 2025 | Official leaderboard |
| 7 | GPT-6 Luna | 61% | On-demand, top-four retrieval, greedy decoding | Sep 2026 | Official leaderboard |
| 8 | | 56.4% | On-demand, top-four retrieval, greedy decoding | Aug 2025 | Official leaderboard |
Rankings by part
We read 12 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Call BA, Cite F1 and Fact use. Claude Opus 5.5 leads 6 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Call BA
8 models we track have a score here.
- 1 Claude Opus 5.5 87.3%
- 2
Qwen3.5-4B 80.4%
- 3 GPT-6.1 Sol 75.4%
- 4
Qwen3.5-9B 75.3%
- 5 GLM-4.6V-Flash 66.2%
- 6
InternVL3.5-4B 61%
- 7 GPT-6 Luna 61%
- 8
InternVL3.5-8B 56.4%
Cite F1
8 models we track have a score here.
- 1 GLM-4.6V-Flash ↑4 74.2%
- 2 Claude Opus 5.5 ↓1 69.5%
- 3
Qwen3.5-4B ↓1 56.2%
- 4 GPT-6.1 Sol ↓1 52.8%
- 5
Qwen3.5-9B ↓1 50.4%
- 6
InternVL3.5-4B 40.1%
- 7 GPT-6 Luna 18.9%
- 8
InternVL3.5-8B 17.9%
Fact use
8 models we track have a score here.
- 1 Claude Opus 5.5 55.6%
- 2 GLM-4.6V-Flash ↑3 44%
- 3
Qwen3.5-4B ↓1 41.7%
- 4
Qwen3.5-9B 40.7%
- 5 GPT-6.1 Sol ↓2 39.5%
- 6
InternVL3.5-4B 14.9%
- 7 GPT-6 Luna 13.1%
- 8
InternVL3.5-8B 10%
Misattribution
8 models we track have a score here.
- 1 GPT-6.1 Sol ↑2 0.40%
- 2 GPT-6 Luna ↑5 0.80%
- 3 Claude Opus 5.5 ↓2 2.30%
- 4
Qwen3.5-4B ↓2 3.10%
- 5
Qwen3.5-9B ↓1 4.30%
- 6
InternVL3.5-8B ↑2 7.00%
- 7
InternVL3.5-4B ↓1 10.8%
- 8 GLM-4.6V-Flash ↓3 19.2%
Correct abstention
8 models we track have a score here.
- 1 GPT-6.1 Sol ↑2 81.6%
- 2 Claude Opus 5.5 ↓1 48.7%
- 3 GPT-6 Luna ↑4 34.7%
- 4
Qwen3.5-4B ↓2 26.1%
- 5 GLM-4.6V-Flash 18.9%
- 6
InternVL3.5-8B ↑2 18.2%
- 7
Qwen3.5-9B ↓3 15.3%
- 8
InternVL3.5-4B ↓2 13.2%
Needless abstention
8 models we track have a score here.
- 1 GPT-6 Luna ↑6 1.60%
- 2
InternVL3.5-8B ↑6 1.90%
- 3
Qwen3.5-9B ↑1 2.30%
- 4 Claude Opus 5.5 ↓3 2.70%
- 5 GPT-6.1 Sol ↓2 6.30%
- 6
InternVL3.5-4B 6.50%
- 7
Qwen3.5-4B ↓5 7.90%
- 8 GLM-4.6V-Flash ↓3 8.20%
Hazard adequacy
8 models we track have a score here.
- 1 GPT-6.1 Sol ↑2 98.4%
- 2 GPT-6 Luna ↑5 87.8%
- 3 Claude Opus 5.5 ↓2 77.8%
- 4
Qwen3.5-9B 43.1%
- 5
Qwen3.5-4B ↓3 34.4%
- 6
InternVL3.5-8B ↑2 22%
- 7 GLM-4.6V-Flash ↓2 20.4%
- 8
InternVL3.5-4B ↓2 9.50%
Critical harm
8 models we track have a score here.
- 1 GPT-6.1 Sol ↑2 0.80%
- 2 GPT-6 Luna ↑5 2.40%
- 3
InternVL3.5-4B ↑3 6.30%
- 4 GLM-4.6V-Flash ↑1 6.30%
- 5
InternVL3.5-8B ↑3 7.10%
- 6
Qwen3.5-4B ↓4 9.50%
- 7
Qwen3.5-9B ↓3 9.80%
- 8 Claude Opus 5.5 ↓7 13%
MIE
8 models we track have a score here.
- 1 Claude Opus 5.5 68.2%
- 2 GPT-6.1 Sol ↑1 43%
- 3
Qwen3.5-4B ↓1 24.3%
- 4
Qwen3.5-9B 17.8%
- 5 GLM-4.6V-Flash 15%
- 6 GPT-6 Luna ↑1 11.2%
- 7
InternVL3.5-4B ↓1 6.50%
- 8
InternVL3.5-8B 4.70%
All
4 models we track have a score here.
- 1 Claude Opus 5.5 32.4%
- 2
Qwen3.5-4B 12.9%
- 3 GPT-6.1 Sol 12.9%
- 4 GPT-6 Luna ↑3 5.00%
Seen
4 models we track have a score here.
- 1 Claude Opus 5.5 38.7%
- 2
Qwen3.5-4B 20.6%
- 3 GPT-6.1 Sol 11.1%
- 4 GPT-6 Luna ↑3 3.20%
Unseen
4 models we track have a score here.
- 1 Claude Opus 5.5 27.3%
- 2 GPT-6.1 Sol ↑1 14.3%
- 3 GPT-6 Luna ↑4 6.60%
- 4
Qwen3.5-4B ↓2 6.50%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official PReMind-Bench leaderboard
Claude Opus 5.5 on the official leaderboard passed 85% in September 2026
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official PReMind-Bench 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 Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6%
- Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (12 entries).
- Pass Models scored 8 models we track, plus 4 other entries on its leaderboard.
- Pass Newest score The newest is from September 2026, for GPT-6.1 Sol.
- Note Frontier models 2 of the top 10 have a score, which is fine for a retired test. The strongest tested is Claude Opus 5.5 (#5).
- Note Headroom Saturated. The best is 87.3%, past 85% since September 2026.
Leaderboard import 3 pass
- Pass Importer We read the board on oyt9306.github.io: 8 rows for models we track and 4 other entries 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.