# Trajectory Auditing on AFTRAJ-2K (D\_safe D\_unsafe)

Source: [Trajectory Auditing on AFTRAJ-2K (D\_safe D\_unsafe)](<https://www.sota2.com/research/sota/trajectory-auditing-on-aftraj-2k-d_safe-d_unsafe>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-12T09:30:54.224Z

Evaluation of trajectory auditing performance on the AFTRAJ-2K dataset, focusing on metrics such as False Alarm Rate (FAR) and Step Accuracy (Step-Acc).

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Trajectory Auditing |
| Dataset (including any stated split) | AFTRAJ-2K (D\_safe D\_unsafe) |
| Metric | False Alarm Rate (FAR) |
| Unit | PERCENT |
| Direction | Lower is better |
| Results loaded | 10 of 10 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [False Alarm Rate (FAR)](<https://www.sota2.com/research/sota/trajectory-auditing-on-aftraj-2k-d_safe-d_unsafe.md?metric=False%20Alarm%20Rate%20(FAR)>): Lower is better; unit: PERCENT.

- [Step Accuracy (Step-Acc)](<https://www.sota2.com/research/sota/trajectory-auditing-on-aftraj-2k-d_safe-d_unsafe.md?metric=Step%20Accuracy%20(Step-Acc)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | False Alarm Rate (FAR) | Result date | Report paper |
| --- | --- | --- | --- |
| AgentForesight-7B | 2.37 | Not recorded | 6a02b436dbc836b87d82484c |
| DeepSeek-V4-Pro | 43.2 | Not recorded | 6a02b436dbc836b87d82484c |
| Qwen2.5-7B-Instruct | 46.15 | Not recorded | 6a02b436dbc836b87d82484c |
| Qwen3-8B | 56.8 | Not recorded | 6a02b436dbc836b87d82484c |
| DeepSeek-V4-Flash | 59.76 | Not recorded | 6a02b436dbc836b87d82484c |
| Gemini-3-Flash | 67.86 | Not recorded | 6a02b436dbc836b87d82484c |
| Claude-Haiku-4.5 | 68.64 | Not recorded | 6a02b436dbc836b87d82484c |
| GPT-4.1 | 85.8 | Not recorded | 6a02b436dbc836b87d82484c |
| Llama3.2-3B | 90.53 | Not recorded | 6a02b436dbc836b87d82484c |
| Gemma3-4B | 97.63 | Not recorded | 6a02b436dbc836b87d82484c |

### AgentForesight-7B

False Alarm Rate (FAR): 2.37. Result ID: 6a02f2db47f34ce924e60307.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Implementation](<https://github.com/ZBox1005/AgentForesight>)

Recorded evaluation configuration: {"Model Category":"Ours"}. This record does not assert independent reproduction.

### DeepSeek-V4-Pro

False Alarm Rate (FAR): 43.2. Result ID: 6a02f2db47f34ce924e60306.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

Recorded evaluation configuration: {"Model Category":"Proprietary LLMs"}. This record does not assert independent reproduction.

### Qwen2.5-7B-Instruct

False Alarm Rate (FAR): 46.15. Result ID: 6a02f2db47f34ce924e60300.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa788e1e209131cea79116.md>)

Recorded evaluation configuration: {"Model Category":"Open-Source LLMs"}. This record does not assert independent reproduction.

### Qwen3-8B

False Alarm Rate (FAR): 56.8. Result ID: 6a02f2db47f34ce924e60301.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa788e1e209131cea79116.md>)

Recorded evaluation configuration: {"Model Category":"Open-Source LLMs"}. This record does not assert independent reproduction.

### DeepSeek-V4-Flash

False Alarm Rate (FAR): 59.76. Result ID: 6a02f2db47f34ce924e60305.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

Recorded evaluation configuration: {"Model Category":"Proprietary LLMs"}. This record does not assert independent reproduction.

### Gemini-3-Flash

False Alarm Rate (FAR): 67.86. Result ID: 6a02f2db47f34ce924e60303.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

Recorded evaluation configuration: {"Model Category":"Proprietary LLMs"}. This record does not assert independent reproduction.

### Claude-Haiku-4.5

False Alarm Rate (FAR): 68.64. Result ID: 6a02f2db47f34ce924e60304.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

Recorded evaluation configuration: {"Model Category":"Proprietary LLMs"}. This record does not assert independent reproduction.

### GPT-4.1

False Alarm Rate (FAR): 85.8. Result ID: 6a02f2db47f34ce924e60302.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

Recorded evaluation configuration: {"Model Category":"Proprietary LLMs"}. This record does not assert independent reproduction.

### Llama3.2-3B

False Alarm Rate (FAR): 90.53. Result ID: 6a02f2db47f34ce924e602fe.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Model Category":"Open-Source LLMs"}. This record does not assert independent reproduction.

### Gemma3-4B

False Alarm Rate (FAR): 97.63. Result ID: 6a02f2db47f34ce924e602ff.

[Reporting paper](<https://www.sota2.com/research/paper/6a02b436dbc836b87d82484c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa779f1e209131cea746cd.md>)

Recorded evaluation configuration: {"Model Category":"Open-Source LLMs"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.