Overview
Long-horizon LLM agents cannot keep full histories in context. Recent systems compress interaction into belief states or summaries: ABBEL maintains natural-language belief bottlenecks; CoACT optimises observation compression for next-action preservation (NAP); other work uses mutual-information rate between raw context and compression as a proxy for compressor quality (ABBEL, arXiv:2512.20111; CoACT, arXiv:2607.02911; information-theoretic agentic design, arXiv:2512.21720). Passing NAP or improving accuracy is not the same as preserving an action-sufficient representation: a summary $R = f(O)$ such that $p(y\mid o, a) = p(y\mid R, a)$ for consequences $y$ of available actions. Summaries can preserve the next click while discarding distinctions needed for later escalation — a microscopic form of agentic debt. This project builds audits that separate those failure modes.
Prerequisite
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
FAQs
- What are the prerequisites?
L172 Information, Energy and Intelligence (IEI), or equivalent preparation in information theory, maximum entropy, and information geometry.
- What will I learn in this Project?
You will learn information-bottleneck and rate–distortion ideas applied to LLM agents, how to evaluate context compression beyond task success, and how to design causal attacks that separate next-action preservation from consequence-preserving (action-sufficient) representations. You will implement diagnostics on modern agent scaffolding (belief summaries, memory compressors).
- What is the objective of the project?
(1) Reproduce one published compressor (ABBEL-style belief bottleneck or CoACT-style observation compression) on a fixed agent benchmark. (2) Implement action-sufficiency and consequence-equivalence tests, not only next-action match or end-task reward. (3) Design distribution-shift attacks: new tools, reordered observations, distractors that preserve NAP but break consequence distributions. (4) Deliver a diagnostic report showing compressions that pass NAP yet fail action-sufficiency, and argue how the lost distinctions would matter at a judgement / escalation junction. Deliverables: audit code, attack suite, and thesis chapter suitable for a short conference paper.
- How does this fit into the bigger picture?
The judgement layer is the set of junctions that must preserve the ability to contest and escalate. Agent memory compression is where that requirement meets engineering practice. This project links IEI information-bottleneck material to instrumentation of agentic systems and to the governance question: what must not be summarised away?
