Affective-Sedimentation-LLM-Agent-CSAI-2026

简体中文 · Architecture · Validation

Accepted at CSAI 2026; not yet published.

All rights reserved. Public viewing only; other uses require prior written permission, subject to existing legal, platform and earlier-license rights. See RIGHTS.md.

This project separates event memory from an externally computed slow affect state, then tests whether state injection changes an LLM agent's choices. It addresses a research problem: distinguishing state-channel effects from the memory text an agent receives. Explicit controls make that interface measurable without changing model weights.

Zexian Xiong — first author. Primarily responsible for mechanism design and validation, with involvement throughout the research. Yan Li — supervisor, second author and corresponding author, providing research guidance. The implementation, data collection, analysis and paper are presented as collaborative research outputs.

Interaction events → fast/medium/slow P/A/D state → prompt metadata → A/B choice → paired analysis

P/A/D represents valence, activation and perceived control. State is computed outside the model and supplied as background metadata that names no action. The key design choice is separating event memory from state, so each channel can be examined with explicit controls:

• E — label control: affect-labeled state versus numerically identical sensor metadata, without event memory.
• N — state ablation: history-derived state versus zero state, with identical event memory.

AB/BA option counterbalancing, fixed seeds, frozen materials and predefined gates constrain the comparisons. These experiments test the prompt interface; they do not isolate a unique causal effect of the three-layer recurrence. The shared implementation combines NumPy dynamics, deterministic seeded histories, a six-worker Qwen runner, resumable logs and SHA-256 provenance. These provide auditable experiment execution. See the architecture.

756 formal observations, plus 80 development/stability observations: 836 preserved ledger records. Both primary gate suites pass on the frozen grid. CS denotes the counterbalanced choice score.

Raw ledger · Locked results

Scope: one Qwen snapshot and authored forced-choice probes. Nine of 15 E probes are saturated; none of the E axis-wise tests passes Holm correction. N's mean only narrowly exceeds the fixed 0.10 engineering threshold, and three of eight seed means fall below it. These results support a limited interface-level effect, not internal emotion, persistent personality or cross-model generality.

Engineering evidence: the recorded validation includes 264 passing tests on Windows and Ubuntu with Python 3.11.9. The credential-free demo checks the numerical controller and two mock observations; it demonstrates functionality, not a new behavioral result. Validation details.

Only for rights holders and users with prior written permission. Commands document the authors' procedure and grant no execution, reuse or modification permission.

Python 3.11.9 is verified on local Windows, Ubuntu 24.04.5 and Windows Server 2025; macOS is untested. No GPU, API key, model weights or paid requests are needed below. Internet is needed for installation and four tokenizer files (about 15 MB); allow several hundred MB for the environment. Tests take roughly 1–3 minutes locally.

git clone https://github.com/Sean-xzx/Affective-Sedimentation-LLM-Agent-CSAI-2026.git
cd Affective-Sedimentation-LLM-Agent-CSAI-2026
python -m venv .venv

Activate on Windows PowerShell:

.\.venv\Scripts\Activate.ps1

Or on Linux/macOS:

source .venv/bin/activate

If PowerShell activation is blocked, use .\.venv\Scripts\python.exe instead of python. Then run:

python -m pip install -r requirements-public.lock
python -m tools.reproduce demo
python -m tools.reproduce verify
python -m tools.prepare_resources
python -m tools.reproduce test
python -m tools.check_publication

Success means every command exits 0: Demo PASS, numerical half-times 1/8/69, two mock observations and live_api_attempts=0; Verify PASS reproduces the results above and checks 33 protected files; resource preparation verifies four tokenizer hashes; the suite reports 264 passed. Tests can rebuild identical result files and generate ignored helper SVGs, so use a disposable clone if temporary writes are unacceptable.

The frozen config fixes 200-event histories, seed 20260804, probes, budgets and the recorded model qwen3.7-flash-2026-07-15. Resource metadata pins Qwen/Qwen3-0.6B as a proxy tokenizer, not weights or a proven identical tokenizer for that hosted snapshot. Cache files stay in ignored .cache/. The publication lock supplies the complete test environment; historical dependency files remain unchanged.

sprint/ implements state, histories, rendering and collection; tools/ analyses the ledger and provides the safe entry points; tests/ protects behavior; data/ and reports/ hold recorded evidence. See Architecture for module relationships and Reproducibility for seeds, units, offline resources, and hashes.

Live hosted collection was not rerun during publication and can incur charges. .env.example describes the optional DASHSCOPE_API_KEY; the adapter reads environment variables and does not load .env on import. Keep scientific code/data frozen. Run modules from the root and prepare the tokenizer before tests.

Questions, permission requests and authorized-run reports: Issues. No unlicensed patches or functional changes are invited; include sanitized output, OS, Python version and commit SHA. No maintenance or hosted-service availability promise is made.

Non-neutral appraisal values are attributed to Gebhard and Kipp (2006), Table 2; see resource provenance. Cite the repository commit used.

No open-source license is offered for this revision. RIGHTS.md defines the permission policy and the limits arising from earlier MIT grants and GitHub platform rights. Third-party dependencies and resources retain their own terms.