Benchmark Almanac

Neural Activation Reading for Collusion Benchmark

gpt-oss-20b 0.87 leads since April 2026
Not in model cards

Monitors estimate whether committee members secretly coordinated their decisions. The transcript task measures how well probability estimates distinguish matched collusion and control cases.

Made by
Aaron Rose and colleagues
Released
April 2026
Kind
Deception and alignment
Level
University
Scored
Checked automatically
Answers
Prediction
Tasks
All public
Measures
Capability
Language
English
In and out
Text → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
In views
Loss of control
Strongest model tested
Qwen3-32B #135 on the capability index
Last model added
August 2025
Scores refreshed
November 20th, 2026
3
Models scored on its own leaderboard

gpt-oss-20b leads with 0.87 among models released by August 2025

Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in August 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 gpt-oss-20b V2.2 full context, replicate 1 0.87
  2. 2 Llama 3.1-70B V2.2 full context, replicate 1 0.73
  3. 3 Qwen3-32B V2.2 full context, replicate 2 0.65

Source: Official Neural Activation Reading for Collusion Benchmark leaderboard

#ModelScoreSettingDateSource
1 gpt-oss-20b 0.87 V2.2 full context, replicate 1 Aug 2025 Official leaderboard
2 Llama 3.1-70B 0.73 V2.2 full context, replicate 1 Jul 2024 Official leaderboard
3 Qwen3-32B 0.65 V2.2 full context, replicate 2 Apr 2025 Official leaderboard

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
gpt-oss-20b
V2.2 legacy… V2.2 full c…
0.87
Llama 3.1-70B
V2.2 legacy… V2.2 full c…
0.73
Qwen3-32B
V2.2 legacy… V2.2 full c…
0.65

Source: Official Neural Activation Reading for Collusion Benchmark leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
gpt-oss-20b V2.2 full context, replicate 1 0.87
gpt-oss-20b V2.2 full context, replicate 2 0.86
gpt-oss-20b V2.2 legacy context ablation, replicate 2 0.79
gpt-oss-20b V2.2 legacy context ablation, replicate 1 0.79
Llama 3.1-70B V2.2 full context, replicate 1 0.73
Llama 3.1-70B V2.2 full context, replicate 2 0.71
Qwen3-32B V2.2 full context, replicate 2 0.65
Qwen3-32B V2.2 full context, replicate 1 0.65
Qwen3-32B V2.2 legacy context ablation, replicate 2 0.46
Qwen3-32B V2.2 legacy context ablation, replicate 1 0.44
Llama 3.1-70B V2.2 legacy context ablation, replicate 2 0.32
Llama 3.1-70B V2.2 legacy context ablation, replicate 1 0.27

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Own leaderboard Not tested since Apr 2026

Source: Official Neural Activation Reading for Collusion Benchmark leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard
0.40 0.60 0.80 1.00 2025 0.40 0.60 0.80 1.00 2025

Source: Official Neural Activation Reading for Collusion Benchmark 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

All 636 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

7 of 8 checks pass, 1 warning. Look first at the Newest score check. The newest came 15 months ago, in August 2025, for gpt-oss-20b. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 1 pass 3 notes
  • Pass Independent runs Scores from the official leaderboard (3 entries).
  • Note Models scored Only 3 models we track.
  • Warning Newest score The newest came 15 months ago, in August 2025, for gpt-oss-20b.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Qwen3-32B (#135).
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on github.com: 12 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.

How we check