Non-Determinism of "Deterministic" LLM Settings
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:STT55N72record.jsonopen to challenge →
read the original abstract
LLM (large language model) practitioners commonly notice that outputs can vary for the same inputs under settings expected to be deterministic. Yet the questions of how pervasive this is, and with what impact on results, have not to our knowledge been systematically investigated. We investigate non-determinism in five LLMs configured to be deterministic when applied to eight common tasks in across 10 runs, in both zero-shot and few-shot settings. We see accuracy variations up to 15% across naturally occurring runs with a gap of best possible performance to worst possible performance up to 70%. In fact, none of the LLMs consistently delivers repeatable accuracy across all tasks, much less identical output strings. Sharing preliminary results with insiders has revealed that non-determinism perhaps essential to the efficient use of compute resources via co-mingled data in input buffers so this issue is not going away anytime soon. To better quantify our observations, we introduce metrics focused on quantifying determinism, TARr@N for the total agreement rate at N runs over raw output, and TARa@N for total agreement rate of parsed-out answers. Our code and data are publicly available at https://github.com/breckbaldwin/llm-stability.
This paper has not been read by Pith yet.
Forward citations
Cited by 21 Pith papers
-
FLIPS: Instance-Fingerprinting for LLMs via Pseudo-random Sequences
FLIPS identifies LLM instances with 96% closed-set and 90% open-set accuracy by exploiting biases in generated binary random sequences across 237 instances.
-
Foundation Models as Oracles for Refactoring Correctness Detection
Foundation models serve as effective oracles for detecting refactoring correctness issues in Java programs, achieving up to 93.8% accuracy in zero-shot evaluations on 226 real bugs.
-
Spider 2.0-AIFunc: Extending Real-World Text-to-SQL to AI-Native SQL Workflows
Spider 2.0-AIFunc is a 465-instance benchmark for evaluating text-to-SQL systems on queries that incorporate Snowflake Cortex AI functions, with evaluations of ten models showing proprietary models reach 67-70% accuracy.
-
Disentangling Visual and Factual Correctness in LVLMs' Visualization Literacy
Introduces CVLAT and VFRI to disentangle visual vs factual correctness in 15 LVLMs, classifies models by reliance sign, compares to human baseline, and tests prompt interventions.
-
A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving
The paper introduces a paired testing protocol for batch-conditioned refusal robustness in LLM serving and reports low rates of genuine safety-label flips after adjudication, with a batch-invariant kernel ablation eli...
-
Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation: Reproducibility Below the Rerun-Stability Baseline
Paraphrase Jaccard similarity of 0.135-0.288 falls below the 0.50-0.61 same-prompt rerun baseline on OpenAI and Anthropic models, showing prompt wording dominates buyer intent in commercial recommendations.
-
Introducing Background Temperature to Characterise Hidden Randomness in Large Language Models
LLMs show implementation-induced randomness even at T=0 that can be characterized as an effective background temperature T_bg estimated via an ideal reference system.
-
The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning
Entropy minimization on self-generated outputs elicits strong reasoning in pretrained LLMs, matching or exceeding supervised RL methods on benchmarks.
-
Multi-Agent LLM-based Metamorphic Testing for REST APIs
ARMeta uses multi-agent LLMs to identify and execute metamorphic relations for REST API testing, showing complementary coverage to scenario-based baselines on two public applications.
-
Multi-Agent Specification-based Metamorphic Testing of FMU-Based Simulations
A multi-agent LLM workflow extracts Given-When-Then metamorphic relations from specifications to generate and run tests on FMU simulations, demonstrated on a lube oil cooling system FMU.
-
Beyond Accuracy: LLM Variability in Evidence Screening for Software Engineering SLRs
LLMs exhibit substantial heterogeneity and non-determinism in SLR evidence screening, abstracts are decisive for performance, and they show no reliable superiority over classical classifiers on two real SLRs.
-
FixV2W: Correcting Invalid CVE-CWE Mappings with Knowledge Graph Embeddings
FixV2W uses knowledge graph embeddings plus longitudinal patterns to fix invalid CVE-CWE mappings, correctly predicting the right CWE for 69% of exploited cases in top-10 rankings and raising ML model MRR from 0.174 to 0.608.
-
Cross-Model Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Across Three Large Language Models
Three LLMs exhibit distinct consistency profiles in repeated exercise prescription generation, with GPT-4.1 producing unique but semantically stable outputs while Gemini 2.5 Flash achieves high similarity through text...
-
An Empirical Study of Multi-Generation Sampling for Jailbreak Detection in Large Language Models
Multi-generation sampling from LLMs uncovers more jailbreak behaviors than single generations, with the largest gains from one to moderate sample counts and diminishing returns thereafter.
-
Like a Hammer, It Can Build, It Can Break: Large Language Model Uses, Perceptions, and Adoption in Cybersecurity Operations on Reddit
Security practitioners use LLMs independently for low-risk productivity tasks while showing interest in enterprise platforms, but reliability, verification needs, and security risks limit broader autonomy.
-
Beyond Relevance: On the Relationship Between Retrieval and RAG Information Coverage
Coverage-focused retrieval metrics correlate strongly with nugget coverage in RAG responses across text and multimodal benchmarks, supporting their use as performance proxies when retrieval and generation goals align.
-
Cross-Lingual Sentiment Misalignment: Auditing Multilingual Language Models for Inversion Risk, Dialectal Representation, and Affective Stability
Multilingual models invert sentiment polarity 28.7% of the time on Bengali text and show asymmetric affective weighting plus a 57% rise in error on formal dialect compared with colloquial Bengali.
-
RAG-DIVE: A Dynamic Approach for Multi-Turn Dialogue Evaluation in Retrieval-Augmented Generation
RAG-DIVE uses an LLM to dynamically generate, validate, and evaluate multi-turn dialogues for assessing RAG system performance in interactive settings.
-
Can LLMs Make (Personalized) Access Control Decisions?
LLMs reflect users' privacy preferences in access control decisions with up to 86% agreement and can promote safer behavior, but personalization trades off higher individual match for potentially less secure results w...
-
Foundation Models as Oracles for Refactoring Correctness Detection
Foundation models achieve up to 93.8% accuracy detecting refactoring bugs across 47 types in Java IDEs via zero-shot prompting on 226 real cases.
-
Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation
Compiled AI generates deterministic code artifacts from LLMs in a one-time compilation step, enabling reliable workflow execution with zero runtime tokens after break-even.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.