determinism

Taco 提出一种跨平台 LLM 压缩方法，通过纯整数推理、精确的 float32 到整数转换与一致性测试向量，确保编码端与解码端在不同 GPU/CPU 硬件上产生完全相同的输出，并在 7 个后端上验证了哈希一致。

该论文研究了在医疗设备合规文档评估中使用大语言模型时的可复现性问题，发现14个商业LLM部署在确定性采样设置下仍产生不一致的评估结果，并构建了一个从OCR到推理全链路确定性的流水线，实现跨机器、跨天的逐字节一致输出。

The article expands Gary Bernhardt's 'Functional Core, Imperative Shell' architecture to 'Deterministic Core, Non-Deterministic Shell,' highlighting determinism over pure functionalism for better testability and broader applicability in software systems.

The author discusses challenges in testing LLM-based agent pipelines where outputs vary with similar prompts, advocating for property-based checks over exact matches to handle non-deterministic outputs.

A blog post discusses the need to complement brilliant but clumsy LLMs with deterministic tools and formal workflows, using the author's experience developing the Beagle SCM with Fable as an example.

Using 2.3 million pieces of Claude inference trajectory data to distill the Qwen3-4B model, resulting in a small model with 100% output consistency and extremely low hallucination. Moreover, the student model was not limited by the teacher model, and it also converged to a universal truth.

The article discusses various techniques to make LLM-powered agents more deterministic, such as golden sets, guardrails, consensus mechanisms, regression tests, coded logic, and hyperparameter tuning, and asks for additional successful methods.

This paper investigates the assumption that setting LLM judge temperature to 0 ensures deterministic safety evaluations. It finds that in practice, many harnesses do not set temperature or seed, leading to high variance, and even with temperature=0, non-determinism persists due to provider-level randomness and API changes.

This speculative paper argues that the Turing Test is flawed and that both human and AI intelligence are deterministic processes, with true consciousness requiring persistent memory loops and internal dialogue, framing AI as an evolutionary phase shift from carbon to silicon.

This survey examines computational nondeterminism in financial AI systems, covering tabular models, graph networks, and LLM-based workflows, and proposes a layered evaluation framework for auditability.

The article discusses using Gemini Omni's edit mode as a deterministic transformation step in agent video pipelines, enabling reproducible state changes without full video regeneration, improving cost and speed.

Explains the Calvin protocol, which uses deterministic locking to achieve distributed ACID transactions without requiring 2PC, improving scalability and reducing contention compared to traditional approaches.

The article analyzes Factorio's deterministic lockstep architecture, explaining how it synchronizes millions of entities by transmitting inputs rather than state snapshots, utilizing custom math libraries and PRNG normalization to ensure bitwise consistency across platforms.