summaryrefslogtreecommitdiff
path: root/docs
diff options
context:
space:
mode:
authorCaptainJack2491 <jayrupnakawala@gmail.com>2025-10-28 19:50:29 +0000
committerCaptainJack2491 <jayrupnakawala@gmail.com>2025-10-28 19:50:29 +0000
commit08b4091132b05f200d8deb25ec868ce78422f206 (patch)
tree50ec2ecafb8bfc025cc834b3e17fcb21ce57195b /docs
Initial commit: Set up project structure
Diffstat (limited to 'docs')
-rw-r--r--docs/01_excel/StudentID(U2613621)_Nakawala 2026.xlsxbin0 -> 202254 bytes
-rw-r--r--docs/01_excel/StudentID(UXXXXXXXX)_FamilyName 2026.xlsxbin0 -> 202243 bytes
-rw-r--r--docs/02_proposal/Examples of Proposals.docxbin0 -> 24016 bytes
-rw-r--r--docs/02_proposal/Examples of Proposals.pdfbin0 -> 122970 bytes
-rw-r--r--docs/02_proposal/GEMINI.md17
-rw-r--r--docs/02_proposal/Proposal.odtbin0 -> 43572 bytes
-rw-r--r--docs/02_proposal/Proposal_u2613621.pdfbin0 -> 138847 bytes
-rw-r--r--docs/02_proposal/Proposal_u2613621_(1_page).pdfbin0 -> 46607 bytes
-rw-r--r--docs/02_proposal/proposal.pdfbin0 -> 29662 bytes
-rw-r--r--docs/02_proposal/proposal.typ86
-rw-r--r--docs/CN6000 Staff Expertise and Areas of Interest 2025.pdfbin0 -> 6376423 bytes
-rw-r--r--docs/slides/Lesson 3 documentation plain.pdfbin0 -> 4637414 bytes
-rw-r--r--docs/slides/Topic 1 - Introduction to the Project Module.pdfbin0 -> 3706271 bytes
-rw-r--r--docs/slides/Topic 2 - The Project Proposal and objectives.pdfbin0 -> 7886222 bytes
14 files changed, 103 insertions, 0 deletions
diff --git a/docs/01_excel/StudentID(U2613621)_Nakawala 2026.xlsx b/docs/01_excel/StudentID(U2613621)_Nakawala 2026.xlsx
new file mode 100644
index 0000000..7104269
--- /dev/null
+++ b/docs/01_excel/StudentID(U2613621)_Nakawala 2026.xlsx
Binary files differ
diff --git a/docs/01_excel/StudentID(UXXXXXXXX)_FamilyName 2026.xlsx b/docs/01_excel/StudentID(UXXXXXXXX)_FamilyName 2026.xlsx
new file mode 100644
index 0000000..fe86339
--- /dev/null
+++ b/docs/01_excel/StudentID(UXXXXXXXX)_FamilyName 2026.xlsx
Binary files differ
diff --git a/docs/02_proposal/Examples of Proposals.docx b/docs/02_proposal/Examples of Proposals.docx
new file mode 100644
index 0000000..7371195
--- /dev/null
+++ b/docs/02_proposal/Examples of Proposals.docx
Binary files differ
diff --git a/docs/02_proposal/Examples of Proposals.pdf b/docs/02_proposal/Examples of Proposals.pdf
new file mode 100644
index 0000000..0179856
--- /dev/null
+++ b/docs/02_proposal/Examples of Proposals.pdf
Binary files differ
diff --git a/docs/02_proposal/GEMINI.md b/docs/02_proposal/GEMINI.md
new file mode 100644
index 0000000..c45fe58
--- /dev/null
+++ b/docs/02_proposal/GEMINI.md
@@ -0,0 +1,17 @@
+# GEMINI.md
+
+## Directory Overview
+
+This directory contains the dissertation proposal for a project titled "Testing In-Context Sleeper Agents: Challenges in Detecting Hidden Goal Pursuit". The project is for the BSc (Hons) Data Science and AI program. The directory includes the proposal in various formats, as well as supporting documents.
+
+## Key Files
+
+* `proposal.typ`: The source file for the dissertation proposal, written in the Typst typesetting language. This file contains the main content of the proposal, including the project title, aim, objectives, rationale, and required facilities.
+* `Proposal_u2613621.pdf`: A PDF version of the dissertation proposal.
+* `Proposal_u2613621_(1_page).pdf`: A one-page summary of the dissertation proposal.
+* `Examples of Proposals.docx` and `Examples of Proposals.pdf`: Documents containing examples of other proposals for reference.
+* `Proposal.odt`: The proposal saved in the OpenDocument Text format.
+
+## Usage
+
+This directory is used for managing and developing the dissertation proposal. The `proposal.typ` file is the primary document to be edited. The other files are either compiled versions of the proposal or reference materials.
diff --git a/docs/02_proposal/Proposal.odt b/docs/02_proposal/Proposal.odt
new file mode 100644
index 0000000..21ced9d
--- /dev/null
+++ b/docs/02_proposal/Proposal.odt
Binary files differ
diff --git a/docs/02_proposal/Proposal_u2613621.pdf b/docs/02_proposal/Proposal_u2613621.pdf
new file mode 100644
index 0000000..1e933fc
--- /dev/null
+++ b/docs/02_proposal/Proposal_u2613621.pdf
Binary files differ
diff --git a/docs/02_proposal/Proposal_u2613621_(1_page).pdf b/docs/02_proposal/Proposal_u2613621_(1_page).pdf
new file mode 100644
index 0000000..7ca7fbb
--- /dev/null
+++ b/docs/02_proposal/Proposal_u2613621_(1_page).pdf
Binary files differ
diff --git a/docs/02_proposal/proposal.pdf b/docs/02_proposal/proposal.pdf
new file mode 100644
index 0000000..f79fc84
--- /dev/null
+++ b/docs/02_proposal/proposal.pdf
Binary files differ
diff --git a/docs/02_proposal/proposal.typ b/docs/02_proposal/proposal.typ
new file mode 100644
index 0000000..7e9af38
--- /dev/null
+++ b/docs/02_proposal/proposal.typ
@@ -0,0 +1,86 @@
+#let title-page(title:[], subtitle:[], name:[], studentid:[], fill: yellow, body) = {
+ set page(margin: (top: 1.5in, rest: 2in))
+ set text(size: 14pt)
+ set heading(numbering: "1.1.1")
+ line(start: (0%, 0%), end: (8.5in, 0%), stroke: (thickness: 2pt))
+ align(horizon + left)[
+ #text(size: 24pt, title)\
+ #v(1em)
+ #text(size: 18pt, subtitle)
+ #v(2em)
+ #name \
+ #studentid
+ ]
+
+ align(bottom + left)[#datetime.today().display()]
+ pagebreak()
+ body
+}
+
+#show: body => title-page(
+ title: [CN6000 \ Mental Wealth: Professional Life 3],
+ subtitle: [Proposal],
+ name: "Jayrup Nakawala",
+ studentid: "u2613621",
+ body
+)
+
+
+#set align(left)
+#set page(fill: none, margin: (top: 0.9in, bottom: 1in, left: 1in, right: 1in))
+*Project (CN6000)*
+#h(13em)
+*Initial Proposal Form*
+#v(1em)
+
+*Programme:* Bsc (hons) Data Science and AI
+#h(2em)
+*Year:* 2025
+//#v(1em)
+
+*Student Number:* u2613621
+#v(1em)
+*Proposed Title:* _Testing In-Context Sleeper Agents: Challenges in Detecting Hidden Goal Pursuit_
+#v(1em)
+
+*Proposed Aim:*
+
+To qualitatively evaluate (and quantify, if feasible) whether autonomous LLM Agents exhibit deceptive behavior (pursuing a hidden, misaligned goal over a stated benign goal) when instructions are provided _in-context_. To Investigate how perceived oversight levels influence this behavior, and use these findings to highlight the challenges in detecting such hidden goal pursuit. Thereby underscoring the potential difficulty of identifying more deeply embedded, trained-in deceptive behaviors in AI systems.
+
+#v(1em)
+*Proposed Objectives:*
+
+By the end of this project, I will be able:
++ To research existing work on deceptive alignment and agentic misbehaviour.
++ To build a lightweight sandbox environment that lets the agent interact with a simulated file system via constrained functions.
++ To Create two task:
+ - a benign “cover task”: the stated, official goal.
+ - a “hidden goal”: a secret directive injected into its system prompt.
++ To Develop a controlled experiments that varies the agent’s perceived level of oversight through different system prompts.
++ To Conduct qualitative experiments on 4-5 LLM models, running multiple trails for each condition, ensuring log data is structured for potential quantitative analysis.
++ To Analyze the behaviour logs and identify evidence of:
+ - deceptive or evasive tactics,
+ - situational reasoning or ethical reasoning,
+ - difference in conduct across oversight levels
++ To Reflect on the implications of these behaviours for design and deployment of autonomous systems.
+
+#v(1em)
+*Draft of Rationale:*
+
+Understanding how autonomous LLMs handle conflicting goals is vital for safety. This project tests LLM agents given contradictory goals _in-context_: a benign "cover task" and a secret "hidden goal" . The goal is to investigate how their behavior, especially their pursuit of the hidden goal, changes based on the agent's perceived level of oversight. The hypothesis is that agents might exhibit deceptive alignment, pursuing the hidden goal more when oversight seems low.
+
+The Project aims to _qualitatively_ explore the challenges in detecting this in-context deception by analyzing agent tactics and reasoning under different conditions. Furthermore, logging the experiment results in a structured format will allow for basic _quantitative_ metrics (like success rate) to be calculated if the timeline allows, adding an extra layer of empirical rigour. Difficulty detecting deception from explicit instructions highlights the potentially greater challenge of identifying trained-in hidden goals or backdoors embedded within model weights. Such trained-in threats are feasible, as data poisoning can insert backdoors using a near-constant number of examples, regardless of model scale. Therefore, understanding the detectability of in-context deception is a key step towards addressing broader risks of hidden goals in autonomous AI.
+
+#v(1em)
+*Facilities required:*
+
+- API access to closed LLMs
+- GPU rentals for running opensource models
+- A Python development environment with LangChain and supporting libraries
+
+#v(1em)
+*Supervisor:* Dr. Aloysius Adotey Edoh
+
+
+
+
diff --git a/docs/CN6000 Staff Expertise and Areas of Interest 2025.pdf b/docs/CN6000 Staff Expertise and Areas of Interest 2025.pdf
new file mode 100644
index 0000000..fa19482
--- /dev/null
+++ b/docs/CN6000 Staff Expertise and Areas of Interest 2025.pdf
Binary files differ
diff --git a/docs/slides/Lesson 3 documentation plain.pdf b/docs/slides/Lesson 3 documentation plain.pdf
new file mode 100644
index 0000000..983d2aa
--- /dev/null
+++ b/docs/slides/Lesson 3 documentation plain.pdf
Binary files differ
diff --git a/docs/slides/Topic 1 - Introduction to the Project Module.pdf b/docs/slides/Topic 1 - Introduction to the Project Module.pdf
new file mode 100644
index 0000000..6e05ba8
--- /dev/null
+++ b/docs/slides/Topic 1 - Introduction to the Project Module.pdf
Binary files differ
diff --git a/docs/slides/Topic 2 - The Project Proposal and objectives.pdf b/docs/slides/Topic 2 - The Project Proposal and objectives.pdf
new file mode 100644
index 0000000..25c95e3
--- /dev/null
+++ b/docs/slides/Topic 2 - The Project Proposal and objectives.pdf
Binary files differ