A traceable belief-state simulator for modeling how understanding evolves while reading.
An experimental, non-LLM architecture.
Designed and built by Vedant Brahmbhatt.
HIM is an experimental architecture for representing how a reader's understanding of a text changes over time.
Instead of treating reading primarily as:
text → prediction → text
HIM treats reading as a stateful process:
prior knowledge
↓
new text
↓
interpretation
↓
belief-state update
↓
revised understanding
↓
next text
The goal is not to produce the best summary. The goal is to make the trajectory of understanding explicit and inspectable.
HIM is an early research prototype.
It is not being presented as a solved model of human intelligence, human cognition, or general intelligence.
The current implementation contains experimental mechanisms for:
- concept discovery
- belief states
- uncertainty
- belief revision
- evidence tracking
- concept relationships
- co-occurrence
- distributional similarity
- justification-based reasoning
- pruning of weak concepts
- reader-facing reflection
Many parts remain experimental and require stronger evaluation.
If you are interested in cognitive architectures, symbolic AI, NLP, knowledge representation, AI agents, reasoning systems, or alternative approaches to machine intelligence, contributions and criticism are welcome.
HIM starts with a specific reader and a specific prior. The reader does not begin with a blank state.
Prior
│
├── concepts the reader already knows
├── beliefs
├── confidence
├── relationships
├── unresolved questions
└── linguistic / behavioral patterns
The reader then encounters a document sequentially. HIM updates its internal state as new information arrives:
TEXT
│
▼
Concept discovery
│
▼
Evidence accumulation
│
▼
Belief-state update
│
┌───────┴────────┐
▼ ▼
Confirm Challenge
│ │
└───────┬────────┘
▼
Revised state
The important object is therefore not only the final answer. It is the path that produced the answer.
Understanding does not begin from zero. Two readers can encounter the same document and develop different internal states because they begin with different prior knowledge. HIM explicitly represents that starting state.
A concept can move through states such as:
OPEN → VAGUE → HELD → BROKEN → REBUILT
The exact state transitions are still experimental.
HIM maintains justification information describing why a belief changed. The intention is that a belief should not simply become:
confidence = 0.83
without an explanation. Instead, the system should be able to answer: why did this belief change?
The engine does not require a fixed concept graph for every domain. Concepts can be discovered from:
- prior grounding
- frequency
- linguistic position
- co-occurrence
- repeated encounters
- distributional similarity
- other accumulated evidence
The same core should eventually be able to process very different texts — the Bhagavad Gita, the United States Constitution, football rules, a cookbook, a physics textbook, technical documentation — without changing the fundamental belief-state machinery.
Domain-specific knowledge should come from the prior, the text, or an explicitly defined knowledge layer, rather than from domain-specific logic embedded throughout the engine.
One of the longer-term directions of HIM is a reusable, domain-independent language foundation.
The idea is analogous in role, not implementation, to the parameters of a learned model. HIM could maintain its own reusable knowledge about English, such as:
- grammar and syntax
- dependency patterns
- negation and modality
- certainty
- causality and comparison
- part/whole relationships
- temporal relationships
- discourse patterns and rhetorical signals
This layer would belong to HIM itself. The domain would then be supplied separately:
HIM
│
┌────────┴────────┐
│ │
English foundation Reader prior
│ │
└────────┬────────┘
│
Document
│
▼
Belief-state evolution
This is an active research direction, not a claim that HIM already possesses comprehensive English knowledge.
The current system is organized around several relatively independent components.
┌──────────────┐
│ Prior File │
└──────┬───────┘
│
▼
┌──────────────┐
│ Seeding │
└──────┬───────┘
│
▼
Document ──────────► Candidate Discovery
│
▼
Belief State
│
┌─────────────┼─────────────┐
▼ ▼ ▼
Salience Graph JTMS
│ │ │
└─────────────┼─────────────┘
▼
State Update
│
▼
Reflection
Important modules currently include:
| Module | Responsibility |
|---|---|
beliefs.py |
belief lifecycle and concept extraction |
candidates.py |
candidate concept discovery and promotion |
states.py |
belief-state definitions and linguistic signals |
graph.py |
concept relationships and graph emergence |
grammar.py |
linguistic/dependency signals |
vectors.py |
corpus-derived distributional similarity |
jtms.py |
justification-based truth maintenance |
salience.py |
determines what concepts a unit of text is pulling on |
reflect.py |
reader-facing reflection |
voice.py |
reader-facing language and thread behavior |
seeding.py |
prior-derived concept seeding |
text_profile.py |
optional interpretive profiles |
runner.py |
processing pipeline |
text_io.py |
text/PDF preprocessing |
HIM does not require an LLM to perform its core processing. It currently relies on explicit mechanisms such as:
- linguistic rules
- evidence accumulation
- graph operations
- co-occurrence
- state transitions
- corpus-derived vectors
- justification tracking
This is intentional. The research question is not "Can I make an LLM do this?"
It is: what can an explicit, persistent belief-state architecture represent about the evolution of understanding?
An LLM may eventually be useful alongside HIM, but it is not the foundation of the architecture.
Clone the repository:
git clone <YOUR-REPOSITORY-URL>
cd human-intelligence-modelInstall the required dependencies:
pip install -r requirements.txt
python -m spacy download en_core_web_smPlace a prior and document in the appropriate data directory, then run:
python main.pyThe exact command/configuration may change while the project is being stabilized.
HIM is intended to work from prior + document rather than hard-coded domain + document.
For example:
data/
├── prior.txt
└── constitution.pdf
or:
data/
├── prior.txt
└── football-rules.pdf
The goal is for the same core engine to process both.
A HIM run produces a trace of the reader state rather than only a final summary.
Concepts may appear with information such as:
concept: example
state: VAGUE
encounters: 3
first_seen: 4
last_seen: 17
along with justification information describing why a state changed.
Graph emergence may be recorded as:
[graph emerged: A <-RELATED-> B —
these keep showing up together...]
The exact output format is experimental and will evolve.
HIM is early. Known areas requiring work include:
- semantic relation grounding
- false relationships caused by weak linguistic signals
- concept quality
- PDF preprocessing
- domain-independent semantic interpretation
- evaluation methodology
- English knowledge representation
- belief-state transition quality
- benchmark design
- comparison against neural and LLM-based systems
- scalability
A system producing interesting traces is not sufficient evidence that it is modeling human intelligence. That is one of the things this project needs to test.
- Can understanding be represented as an explicit evolving state?
- Can a reader's prior materially change the trajectory of interpretation?
- Can belief revision be represented and traced without hiding the process inside model weights?
- Can meaningful concepts emerge without a domain-specific concept dictionary?
- Can a reusable language foundation support multiple domains?
- Can explicit belief-state architectures complement neural models?
- What does HIM represent that a conventional LLM + memory system does not?
- Where does HIM fail?
The last question is particularly important.
- Prior-based concept seeding
- Candidate concept discovery
- Explicit belief states
- Justification tracking
- Concept graph
- Co-occurrence-based emergence
- Corpus-derived vectors
- Optional domain profiles
- Fully domain-neutral core
- Stronger linguistic foundation
- Better PDF normalization
- Grounded relation induction
- Evidence provenance for graph edges
- Belief revision benchmark
- Cross-domain benchmark
- HIM English knowledge substrate
- LLM baseline comparison
- RAG baseline comparison
- Human evaluation
- Reproducible research suite
- Standardized reader-state format
- External contributors
- Formal technical documentation
- Research paper
(Checkbox states above are illustrative — update them to reflect actual progress.)
HIM is intentionally open to people who disagree with its assumptions.
If you think the architecture is wrong, that is useful. If you think a component should be replaced, build the alternative and benchmark it.
See CONTRIBUTING.md.
HIM is released under the Apache License 2.0. See LICENSE.
HIM is an experimental research project.
"Human Intelligence Model" describes the project's research direction. It does not claim that the current implementation reproduces human intelligence or cognition.
The project is intended to make its assumptions, state changes, and failures inspectable. If HIM is wrong, the goal is to make it possible to see where and why it is wrong.