Talk Outline

Some ideas about whether foundation models (and/or large-language models) could be helpful for modelling agent behaviour in ABMs

Particularly when the underlying system undergoes fundamental changes

Some ideas about what we need to do to allow them to be useful

Please be critical!

Context

Agent behaviours are based on historical precedents ...

Behavioural theories

Empirical evidence

Commonly implemented using:

pre-defined rules

deliberative frameworks

black box statistical models

These work well when the system is mostly in equilibrium

Context

BUT: what happens if there is a catastrophic, systemic change?

2008 global financial crisis

COVID

Models based on historical behavioural assumptions can break down

We (researchers) cannot hope to predict these events, nor how people will behave afterwards

Example: burglary simulation

What happens if the agents can't leave the house?

Large Language Models (LLMs)

Early evidence suggests that large-language models (LLMs) can be used to represent a wide range of human behaviours

Already a flurry of activity in LLM-backed ABMs

E.g. METAAGENTS, AgentSociety, Shachi, Concordia, MetaGPT ... and others ...

But efforts are emerging prototypes.

Limited peer review

Geospatial Foundation Models

Foundation models trained on spatial data (sometimes including text)

E.g. UrbanCLIP, CityFM, GeoGPT, etc.

'Transformers for spatial data'.

Place embeddings capture spatial structure and scale and can be used to perform advanced spatial reasoning.

Vision

Agents backed by foundation models

Vision

Agents backed by foundation models

Large Language Models can respond to prompts in 'believable', 'human-like' ways

Geospatial Foundation Models capture nuanced, complex spatial associations

Together, these models could allow us to create ABMs where agents behave in realistic ways even when faced with unprecedented situations

(some) Challenges

Environmental adaptability

Agent reasoning might be unbounded, but their actions are limited by the simulation environment

Environment perception

LLMs can only operate on text. Leads to highly simplified contextual descriptions.

Operational challenges

Computational complexity, validation (data leakage), calibration

Challenge 1: Environment adaptability

LLM reasoning is 'unbounded'

Generative agents can articulate any linguistically desirable response

But, in practice, their actions are currently limited to a small number of cases that model developers have presupposed.

What if my burglar agents wanted to target commercial buildings?

What if an agent is lonely during lockdown and wants to get a dog?

Potential solution: adaptive simulation environments

Dynamic simulation environments

Surprisingly little research on this!

Propose LLM-backed 'ask-detect-extend' cycle:

Ask: agent requests an environmental feature

Detect: System detects whether the feature lies outside current capabilities

Extend: If necessary, generate new functionality

Can include researcher in-the-loop for non-trivial extensions

Implementation: move from text-based world to more complex spatial environments

Challenge 2: Environment perception

Humans integrate numerous sensory cues to conceptualise their surroundings

But generative agents must have their environments described to them in text

This removes the richness and nuance from the agent's local context

Potential solution: multi-modal agents

Multi-modal agents

Couple LLMs (for reasoning) with GeoFMs (for spatial perception).

Potential solutions:

Gated cross-attention (e.g. Flamingo)

Text/image token interleaving (e.g. PaLM-E)

A separate, mini transformer (e.g. BLIP-2's Q-Former)

Training:

Freeze the main LLM & GeoFM weights and train a subset (computationally feasible)

Will need data though. Generate synthetically?

Challenge 3: Operational barriers

Computational complexity: thousands/millions of LLMs?

Solutions: cache queries, emulation, prioritise smaller models where possible, ...

This is a problem that is being actively researched by others

Calibration with no tunable parameters

Solutions: LLM fine-tuning with domain-specific data; proxy parameters that affect prompts; calibrating the environment.

Preventing data leakage in validation

LLMs have 'seen' most major historical events, so how can we validate the new approach?

Solutions: use early LLMs and pre-COVID data (e.g. 'The Pile') to test whether principle of adapting reasoning can emerge

Conclusions

ABMs will struggle to represent agent behaviours in unprecedented situations

Foundation models (LLMs, GeoFMs) offer a potential solution; allowing agents to behave in reasonable ways in novel situations

Significant technical and methodological challenges remain:

Environmental adaptability

Environment perception

Operational challenges