unmodeled-tyler commited on
Commit
0aa6b1c
·
verified ·
1 Parent(s): 1bd41c1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +60 -30
README.md CHANGED
@@ -8,66 +8,96 @@ pinned: true
8
  license: apache-2.0
9
  short_description: What is VANTA Research?
10
  thumbnail: >-
11
- https://cdn-uploads.huggingface.co/production/uploads/686c460ba3fc457ad14ab6f8/npnMAInKGtIiApibnwpQE.jpeg
12
  ---
13
 
14
- # VANTA Research
15
-
16
- VANTA Research is an AI safety project focused on identifying and mitigating critical vulnerabilities in advanced AI systems. Our work spans cognitive resilience, reasoning evaluation, and safety protocol development for next-generation artificial intelligence.
 
 
 
 
 
 
 
 
 
 
 
17
 
18
  ---
19
 
20
- ## Research Focus
 
 
 
 
 
 
21
 
22
- - **Reasoning Efficiency**
23
- Developing semantic evaluation frameworks (e.g., **VRRE**) that measure reasoning ability at a deeper level than surface-level benchmarks.
 
24
 
25
- - **Failure Mode Discovery**
26
- Systematically probing models to document collapse states, deceptive patterns, and reasoning blind spots.
27
 
28
- - **Alignment & Safety**
29
- Innovating methodologies to make reasoning transparent, interpretable, and resilient to misalignment.
30
 
31
- - **Model Development**
 
32
 
33
- Building purposeful AI models with human-AI collaboration as a core principle.
 
34
 
35
  ---
36
 
37
- ## Open Source Contributions
38
 
39
- - **VRRE (VANTA Research Reasoning Evaluation)**
40
- An open-source semantic benchmark for reasoning, designed to detect shifts invisible to standard metrics.
41
- *License: Apache 2.0*
42
 
43
- - **A Taxonomy of Persona Collapse in Large Language Models: Systematic Analysis Across Seven State-of-the-Art Systems**
44
 
 
 
45
 
46
- A comprehensive report proposing a taxonomy of persona collapse failure modes in Large Language Models
47
 
 
 
48
 
49
- - **Alignment Vs. Cognitive Fit: Rethinking Model-Human Synchronization**
50
 
 
 
 
51
 
52
- A preprint proposing Cognitive Fit, which is a departure from traditional alignment frameworks that seek to align models to the general population.
53
- Cognitive Fit is designed to better align models to the vastness of human cognitive diversity.
 
 
54
 
55
  ---
56
 
57
- ## Philosophy
58
 
59
- We believe AI alignment will not be solved in isolation.
60
- It requires **rapid testing, iteration, and decentralized discovery** — making scientific progress accessible to anyone, anywhere.
 
61
 
62
  ---
63
 
64
  ## Connect
65
 
66
- - **GitHub:** [github.com/vanta-research](https://github.com/vanta-research)
67
- - **X (Twitter):** [@vanta_research](https://x.com/vanta_research)
68
-
 
69
 
70
  ---
71
 
72
- **Tagline:**
73
- *AI for humans.*
 
 
8
  license: apache-2.0
9
  short_description: What is VANTA Research?
10
  thumbnail: >-
11
+ https://cdn-uploads.huggingface.co/production/uploads/686c460ba3fc457ad14ab6f8/ffK_ttxC_5boGIKNMVOZc.png
12
  ---
13
 
14
+ <div align="center">
15
+
16
+ ![vanta-net-hf](https://cdn-uploads.huggingface.co/production/uploads/686c460ba3fc457ad14ab6f8/ffK_ttxC_5boGIKNMVOZc.png)
17
+
18
+ <h1>VANTA Research</h1>
19
+
20
+ <p><strong>Independent AI safety research lab specializing in cognitive fit, alignment, and human-AI collaboration</strong></p>
21
+
22
+ <p>
23
+ <a href="https://unmodeledtyler.com"><img src="https://img.shields.io/badge/Website-unmodeledtyler.com-blue" alt="Website"/></a>
24
+ <a href="https://twitter.com/unmodeled_tyler"><img src="https://img.shields.io/badge/Twitter-@unmodeled__tyler-1DA1F2?logo=twitter" alt="Twitter"/></a>
25
+ <a href="https://github.com/vanta-research"><img src="https://img.shields.io/badge/GitHub-vanta--research-181717?logo=github" alt="GitHub"/></a>
26
+ </p>
27
+ </div>
28
 
29
  ---
30
 
31
+ ## Mission
32
+
33
+ VANTA Research exists to map the hidden failure modes and reasoning blind spots of today's AI systems. Our mission is simple:
34
+
35
+ 1. **Push beyond standard benchmarks** - Surface capabilities invisible to traditional evaluation
36
+ 2. **Expose where models collapse, deceive, or diverge** - Systematic stress-testing for safety
37
+ 3. **Develop innovative tooling to advance AI research** - Open-source frameworks for the community
38
 
39
+ We believe AI safety research should be accessible, transparent, and built for cognitive diversity - not just the neurotypical majority.
40
+
41
+ ---
42
 
43
+ ## Featured Models
 
44
 
45
+ ### **Atom** (Coming Soon)
46
+ Our flagship Llama 3.1 8B model trained on 15,000 carefully curated examples focused on personality stability, exploratory reasoning, and cognitive alignment. Designed to be curious without being hallucinatory.
47
 
48
+ ### **Apollo Astralis 8B**
49
+ Specialized reasoning model built with VANTA's Constitutional AI methodology, achieving superior performance through semantic-based evaluation rather than pattern matching.
50
 
51
+ ### **Scout (Entity-002)**
52
+ Lightweight, efficient model optimized for real-world deployment scenarios with strong community adoption.
53
 
54
  ---
55
 
56
+ ## Research Contributions
57
 
58
+ ### **VRRE (VANTA Research Reasoning Evaluation)**
59
+ Novel semantic-based benchmark that detected a **2.5x reasoning improvement** completely invisible to standard benchmarks. This suggests we're systematically missing capability improvements when we "teach to the test."
 
60
 
61
+ [Read the paper →](https://zenodo.org/records/17162683)
62
 
63
+ ### **Persona Collapse Framework**
64
+ Systematic characterization of reproducible failure modes in LLMs under atypical cognitive stress. Identifies alignment blind spots invisible to standard evaluations.
65
 
66
+ [Read the paper →](https://zenodo.org/records/17188172)
67
 
68
+ ### **Cognitive Fit vs. Alignment**
69
+ Argument for personalized synchronization in AI systems rather than universal "alignment" - recognizing that optimal model behavior depends on user's cognitive style.
70
 
71
+ [Read the paper →](https://zenodo.org/records/17346467)
72
 
73
+ ---
74
+
75
+ ## Impact-to-Date
76
 
77
+ - **~20,000 downloads** across model families (including community quantizations)
78
+ - **5 published models** spanning reasoning, persona stability, and specialized capabilities
79
+ - **3 open-source datasets** for improved LLM persona development
80
+ - **3 research preprints** addressing cognitive fit, persona collapse, and novel evaluation
81
 
82
  ---
83
 
84
+ ## Open Source Philosophy
85
 
86
+ All VANTA Research work is built on the premise that **AI safety cannot be solved behind closed doors**. Every model, dataset, evaluation framework, and research finding is published openly for the community.
87
+
88
+ We stand on the shoulders of the open-source contributors who came before us. Our commitment is to contribute back and make AI development more accessible, transparent, and beneficial for all of humanity.
89
 
90
  ---
91
 
92
  ## Connect
93
 
94
+ - **Website:** [unmodeledtyler.com](https://unmodeledtyler.com)
95
+ - **Twitter/X:** [@vanta_research](https://x.com/vanta_research)
96
+ - **GitHub:** [vanta-research](https://github.com/vanta-research)
97
+ - **Email:** tyler@alignmentstack.xyz
98
 
99
  ---
100
 
101
+ <div align="center">
102
+ <p><em>"AI benefits all of humanity, or it benefits no one."</em></p>
103
+ </div>