Difference between revisions of "ERI"

From GISAXS
Jump to: navigation, search
(→‎Research Thrusts)
(→‎Models)
 
(18 intermediate revisions by the same user not shown)
Line 5: Line 5:
 
==Models==
 
==Models==
 
'''How to adapt frontier methods and foundation models to science?'''
 
'''How to adapt frontier methods and foundation models to science?'''
 +
 +
[[Image:Cognitive block11.png|400px]]
 +
 
# Topical fine-tuning
 
# Topical fine-tuning
 
# Tool-use
 
# Tool-use
 
# Advanced retrieval-augmented generation (RAG++)
 
# Advanced retrieval-augmented generation (RAG++)
 
#* '''Novel:''' Pre-generation: Agents continually add content to RAG corpus. ("Pre-thinking" across many vectors.)
 
#* '''Novel:''' Pre-generation: Agents continually add content to RAG corpus. ("Pre-thinking" across many vectors.)
 +
#** 2025-03: [https://arxiv.org/abs/2503.18866 Reasoning to Learn from Latent Thoughts]
 
# Science-adapted tokenization/embedding (xVal, [IDK])
 
# Science-adapted tokenization/embedding (xVal, [IDK])
 
# Specialized sampling
 
# Specialized sampling
 
#* Entropy sampling: measure uncertainty of CoT trajectories
 
#* Entropy sampling: measure uncertainty of CoT trajectories
 
#* '''Novel:''' Handoff sampling:
 
#* '''Novel:''' Handoff sampling:
−
#** text-to-text (specialization, creativity, etc.)
+
#** Useful for:
−
#** text-to-tool (e.g. math)
+
#*** text-to-text (specialization, creativity, etc.)
−
#** test-to-field (integrate non-textual FM)
+
#*** text-to-tool (e.g. math)
 +
#*** test-to-field (integrate non-textual FM)
 +
#** Implementation:
 +
#*** MI-SAE on both spaces, find matches (or maybe just "analogies"?)
 +
#*** GNN, e.g.: 2019-11: [https://arxiv.org/abs/1911.11763 SuperGlue: Learning Feature Matching with Graph Neural Networks] ([https://huggingface.co/docs/transformers/main/en/model_doc/superglue hf])
 +
#*** 2025-02: [https://arxiv.org/abs/2502.03714 Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment]
 +
 
 +
'''Challenge: Connect reasoning models to domain models'''
 +
# Latent space reasoning
 +
# Establish mappings (analogies) between interpretability spaces
 +
#* 2024-12: [https://arxiv.org/abs/2412.16325 Towards Safe and Honest AI Agents with Neural Self-Other Overlap]
 +
#** 2024-07: [https://www.lesswrong.com/posts/hzt9gHpNwA2oHtwKX/self-other-overlap-a-neglected-approach-to-ai-alignment Self-Other Overlap: A Neglected Approach to AI Alignment]
 +
#** 2025-03: [https://www.lesswrong.com/posts/jtqcsARGtmgogdcLT/reducing-llm-deception-at-scale-with-self-other-overlap-fine Reducing LLM deception at scale with self-other overlap fine-tuning]
 +
# Cycling recaptioning/reframing
 +
#* 2024-07: [https://arxiv.org/abs/2407.06723 Graph-Based Captioning: Enhancing Visual Descriptions by Interconnecting Region Captions]
 +
# Tokenizer-for-science: learn right spectrum of representations (for text/image reasoning model)
  
 
==Agents==
 
==Agents==
 
'''How to make AI agents smarter?'''
 
'''How to make AI agents smarter?'''
 +
 +
[[Image:Agent thinking01.png|400px]]
 +
 
# Iteration schemes (loops, blocks)
 
# Iteration schemes (loops, blocks)
 +
## Thinking:
 +
##* Blocky/neural: Define architecture, allow system to pick hyper-parameters
 
## Autonomous ideation:
 
## Autonomous ideation:
 
##* '''Novel:''' Treat ideation as an AE problem in a semantic embedding space.
 
##* '''Novel:''' Treat ideation as an AE problem in a semantic embedding space.
−
 
+
## Dynamic tree-of-thought: on-demand context generation, allows model to select among data representations (zoom, modality, etc.)
 +
# Encode Human Patterns
 +
## Human scientist workflows (ideation, solving, etc.)
 +
## Thought-templates, thought-flows
 +
# How to allow agents to run for long time-horizons coherently?
 +
## ''Basket of Metrics'': Need to define metrics of: (1) research success, (2) uncertainty (entropy sampling?)
 +
## Tool-use to "call human" and request help/information
 
# Memory
 
# Memory
−
# Thought-templates, thought-flows
+
## Allow system to insert and retrieve from RAG at will.
  
 
==Exocortex==
 
==Exocortex==
 
'''What is the right architecture for AI swarms?'''
 
'''What is the right architecture for AI swarms?'''
 +
 +
[[Image:Coord schemes02.png|400px]]
 +
 
# Interaction schemes
 
# Interaction schemes
 
## Test options, identify match between science task and scheme
 
## Test options, identify match between science task and scheme
 
## Treat interaction graph as ML optimization problem
 
## Treat interaction graph as ML optimization problem
 
## '''Novel:''' Map-spatial: Use a map (e.g. of BNL) to localize docs/resources/etc.
 
## '''Novel:''' Map-spatial: Use a map (e.g. of BNL) to localize docs/resources/etc.
−
## '''Novel:''' Pseudo-spatial: Use position in embedding space to localize everything
+
## '''Novel:''' Pseudo-spatial: Use position in embedding space to localize everything. Evolving state (velocity/momentum) of agent carries information.
−
## '''Novel:''' Dynamic-pseudo-spatial: Allow the space to be learned and updated
+
## '''Novel:''' Dynamic-pseudo-spatial: Allow the space to be learned and updated; directions in embedding space can dictate information flow
 
# Establish benchmarks/challenges/validations
 
# Establish benchmarks/challenges/validations
−
 
  
 
==Infrastructure==
 
==Infrastructure==
Line 50: Line 82:
 
===Human-Computer Interaction (HCI)===
 
===Human-Computer Interaction (HCI)===
 
'''What should the HCI be?'''
 
'''What should the HCI be?'''
 +
===Resources===
 +
# Need models, data, facilities, etc. all accessible as API endpoints.

Latest revision as of 09:29, 26 March 2025

ERI

Research Thrusts

Models

How to adapt frontier methods and foundation models to science?

Cognitive block11.png

  1. Topical fine-tuning
  2. Tool-use
  3. Advanced retrieval-augmented generation (RAG++)
  4. Science-adapted tokenization/embedding (xVal, [IDK])
  5. Specialized sampling

Challenge: Connect reasoning models to domain models

  1. Latent space reasoning
  2. Establish mappings (analogies) between interpretability spaces
  3. Cycling recaptioning/reframing
  4. Tokenizer-for-science: learn right spectrum of representations (for text/image reasoning model)

Agents

How to make AI agents smarter?

Agent thinking01.png

  1. Iteration schemes (loops, blocks)
    1. Thinking:
      • Blocky/neural: Define architecture, allow system to pick hyper-parameters
    2. Autonomous ideation:
      • Novel: Treat ideation as an AE problem in a semantic embedding space.
    3. Dynamic tree-of-thought: on-demand context generation, allows model to select among data representations (zoom, modality, etc.)
  2. Encode Human Patterns
    1. Human scientist workflows (ideation, solving, etc.)
    2. Thought-templates, thought-flows
  3. How to allow agents to run for long time-horizons coherently?
    1. Basket of Metrics: Need to define metrics of: (1) research success, (2) uncertainty (entropy sampling?)
    2. Tool-use to "call human" and request help/information
  4. Memory
    1. Allow system to insert and retrieve from RAG at will.

Exocortex

What is the right architecture for AI swarms?

Coord schemes02.png

  1. Interaction schemes
    1. Test options, identify match between science task and scheme
    2. Treat interaction graph as ML optimization problem
    3. Novel: Map-spatial: Use a map (e.g. of BNL) to localize docs/resources/etc.
    4. Novel: Pseudo-spatial: Use position in embedding space to localize everything. Evolving state (velocity/momentum) of agent carries information.
    5. Novel: Dynamic-pseudo-spatial: Allow the space to be learned and updated; directions in embedding space can dictate information flow
  2. Establish benchmarks/challenges/validations

Infrastructure

Architecture

What software architecture is needed?

  1. Code for scaffolding
  2. Scheme for inter-agent messaging (plain English w/ pointers, etc.)
  3. Data management

Hardware

How to implement inference-time compute for exocortex?

  1. Heterogeneous hardware
  2. Elastic (combine local & cloud)
  3. Workflow management

Human-Computer Interaction (HCI)

What should the HCI be?

Resources

  1. Need models, data, facilities, etc. all accessible as API endpoints.
Retrieved from "http://gisaxs.com/index.php?title=ERI&oldid=7348"