Company
OpenAI
Research and product company building general-purpose AI systems.
Recent coverage
Related events
Related papers
arXiv cs.AI
Local-Splitter: A Measurement Study of Seven Tactics for Reducing Cloud LLM Token Usage on Coding-Agent Workloads
We present a systematic measurement study of seven tactics for reducing cloud LLM token usage when a small local model can act as a triage layer in front of a frontier cloud model. The tactics are: (1) local routing, (2) prompt compression, (3) semantic caching, (4) local drafting with cloud review, (5) minimal-diff edits, (6) structured intent extraction, and (7) batching with vendor prompt caching. We implement all seven in an open-source shim that speaks both MCP and the OpenAI-compatible HTTP surface, supporting any local model via Ollama and any cloud model via an OpenAI-compatible endpoint. We evaluate each tactic individually, in pairs, and in a greedy-additive subset across four coding-agent workload classes (edit-heavy, explanation-heavy, general chat, RAG-heavy). We measure tokens saved, dollar cost, latency, and routing accuracy. Our headline finding is that T1 (local routing) combined with T2 (prompt compression) achieves 45-79% cloud token savings on edit-heavy and explanation-heavy workloads, while on RAG-heavy workloads the full tactic set including T4 (draft-review) achieves 51% savings. We observe that the optimal tactic subset is workload-dependent, which we believe is the most actionable finding for practitioners deploying coding agents today.
arXiv cs.AI
Designing Reliable LLM-Assisted Rubric Scoring for Constructed Responses: Evidence from Physics Exams
Student responses in STEM assessments are often handwritten and combine symbolic expressions, calculations, and diagrams, creating substantial variation in format and interpretation. Despite their importance for evaluating students' reasoning, such responses are time-consuming to score and prone to rater inconsistency, particularly when partial credit is required. Recent advances in large language models (LLMs) have increased attention to AI-assisted scoring, yet evidence remains limited regarding how rubric design and LLM configurations influence reliability across performance levels. This study examined the reliability of AI-assisted scoring of undergraduate physics constructed responses using GPT-4o. Twenty authentic handwritten exam responses were scored across two rounds by four instructors and by the AI model using skill-based rubrics with differing levels of analytic granularity. Prompting format and temperature settings were systematically varied. Overall, human-AI agreement on total scores was comparable to human inter-rater reliability and was highest for high- and low-performing responses, but declined for mid-level responses involving partial or ambiguous reasoning. Criterion-level analyses showed stronger alignment for clearly defined conceptual skills than for extended procedural judgments. A more fine-grained, checklist-based rubric improved consistency relative to holistic scoring. These findings indicate that reliable AI-assisted scoring depends primarily on clear, well-structured rubrics, while prompting format plays a secondary role and temperature has relatively limited impact. More broadly, the study provides transferable design recommendations for implementing reliable LLM-assisted scoring in STEM contexts through skill-based rubrics and controlled LLM settings.
arXiv cs.AI
Is Vibe Coding the Future? An Empirical Assessment of LLM Generated Codes for Construction Safety
The emergence of vibe coding, a paradigm where non-technical users instruct Large Language Models (LLMs) to generate executable codes via natural language, presents both significant opportunities and severe risks for the construction industry. While empowering construction personnel such as the safety managers, foremen, and workers to develop tools and software, the probabilistic nature of LLMs introduces the threat of silent failures, wherein generated code compiles perfectly but executes flawed mathematical safety logic. This study empirically evaluates the reliability, software architecture, and domain-specific safety fidelity of 450 vibe-coded Python scripts generated by three frontier models, Claude 3.5 Haiku, GPT-4o-Mini, and Gemini 2.5 Flash. Utilizing a persona-driven prompt dataset (n=150) and a bifurcated evaluation pipeline comprising isolated dynamic sandboxing and an LLM-as-a-Judge, the research quantifies the severe limits of zero-shot vibe codes for construction safety. The findings reveal a highly significant relationship between user persona and data hallucination, demonstrating that less formal prompts drastically increase the AI's propensity to invent missing safety variables. Furthermore, while the models demonstrated high foundational execution viability (~85%), this syntactic reliability actively masked logic deficits and a severe lack of defensive programming. Among successfully executed scripts, the study identified an alarming ~45% overall Silent Failure Rate, with GPT-4o-Mini generating mathematically inaccurate outputs in ~56% of its functional code. The results demonstrate that current LLMs lack the deterministic rigor required for standalone safety engineering, necessitating the adoption of deterministic AI wrappers and strict governance for cyber-physical deployments.
arXiv cs.AI
Can AI Tools Transform Low-Demand Math Tasks? An Evaluation of Task Modification Capabilities
While recent research has explored AI tools' ability to classify the quality of mathematical tasks (arXiv:2603.03512), little is known about their capacity to increase the quality of existing tasks. This study investigated whether AI tools could successfully upgrade low-cognitive-demand mathematics tasks. Eleven tools were tested, including six broadly available, general-purpose AI tools (e.g., ChatGPT and Claude) and five tools specialized for mathematics teachers (e.g., Khanmigo, coteach.ai). Using the Task Analysis Guide framework (Stein & Smith, 1998), we prompted AI tools to modify two different types of low-demand mathematical tasks. The prompting strategy aimed to represent likely approaches taken by knowledgeable teachers, rather than extensive optimization to find a more effective prompt (i.e., an optimistic typical outcome). On average, AI tools were only moderately successful: tasks were accurately upgraded only 64% of the time, with different AI tool performance ranging from quite weak (33%) to broadly successful (88%). Specialized tools were only moderately more successful than general-purpose tools. Failure modes included both "undershooting" (maintaining low cognitive demand) and "overshooting" (elevating tasks to an overly ambitious target category that likely would be rejected by teachers). Interestingly, there was a small negative correlation (r = -.35) between whether a given AI tool was able to correctly classify the cognitive demand of tasks and whether the AI was able to upgrade tasks, showing that the ability to modify tasks (i.e., a generative task) represents a distinct capability from the ability to classify them (i.e., judgement using a rubric). These findings have important implications for understanding AI's potential role in curriculum adaptation and highlight the need for specialized approaches to support teachers in modifying instructional materials.
arXiv cs.AI
Operationalising the Right to be Forgotten in LLMs: A Lightweight Sequential Unlearning Framework for Privacy-Aligned Deployment in Politically Sensitive Environments
Large Language Models (LLMs) are increasingly deployed in politically sensitive environments, where memorisation of personal data or confidential content raises regulatory concerns under frameworks such as the GDPR and its Right to be Forgotten. Translating such legal principles into large-scale generative systems presents significant technical challenges. We introduce a lightweight sequential unlearning framework that explicitly separates retention and suppression objectives. The method first stabilises benign capabilities through positive fine-tuning, then applies layer-restricted negative fine-tuning to suppress designated sensitive patterns while preserving general language competence. Experiments on the SemEval-2025 LLM Unlearning benchmark demonstrate effective behavioural suppression with minimal impact on factual accuracy and fluency. GPT-2 exhibits greater robustness than DistilGPT-2, highlighting the role of model capacity in privacy-aligned adaptation. We position sequential unlearning as a practical and reproducible mechanism for operationalising data erasure requirements in politically deployed LLMs.
arXiv cs.AI
Cooperative Memory Paging with Keyword Bookmarks for Long-Horizon LLM Conversations
When LLM conversations grow beyond the context window, old content must be evicted -- but how does the model recover it when needed? We propose cooperative paging: evicted segments are replaced with minimal keyword bookmarks ([pN:keywords], ~8-24 tokens each), and the model is given a recall() tool to retrieve full content on demand. On the LoCoMo benchmark (10 real multi-session conversations, 300+ turns), cooperative paging achieves the highest answer quality among six methods -- outperforming truncation, BM25, word-overlap retrieval, a search-tool baseline, and full context -- on four models (GPT-4o-mini, DeepSeek-v3.2, Claude Haiku, GLM-5), confirmed by four independent LLM judges ($p=0.017$, paired bootstrap). We then study the paging design space with a 5x4 ablation over boundary strategies and eviction policies (3,176 synthetic probes, 1,600 LoCoMo probes). Key findings: (1) coarse fixed-size pages (fixed_20) reach 96.7% while content-aware topic_shift collapses to 56.7%; (2) eviction policy choice is data-dependent (FIFO best on synthetic, LFU on LoCoMo); (3) two bookmark generation strategies improve over the heuristic baseline (+4.4 and +8.7 E2E points); (4) the remaining bottleneck is bookmark discrimination -- the model triggers recall() 96% of the time but selects the correct page only 57% when bookmarks are insufficiently distinctive. Keyword specificity alone accounts for a 25 percentage point accuracy difference.
arXiv cs.AI
Mining Large Language Models for Low-Resource Language Data: Comparing Elicitation Strategies for Hausa and Fongbe
Large language models (LLMs) are trained on data contributed by low-resource language communities, yet the linguistic knowledge encoded in these models remains accessible only through commercial APIs. This paper investigates whether strategic prompting can extract usable text data from LLMs for two West African languages: Hausa (Afroasiatic, approximately 80 million speakers) and Fongbe (Niger-Congo, approximately 2 million speakers). We systematically compare six elicitation task types across two commercial LLMs (GPT-4o Mini and Gemini 2.5 Flash). GPT-4o Mini extracts 6-41 times more usable target-language words per API call than Gemini. Optimal strategies differ by language: Hausa benefits from functional text and dialogue, while Fongbe requires constrained generation prompts. We release all generated corpora and code.
arXiv cs.AI
ReflectCAP: Detailed Image Captioning with Reflective Memory
Detailed image captioning demands both factual grounding and fine-grained coverage, yet existing methods have struggled to achieve them simultaneously. We address this tension with Reflective Note-Guided Captioning (ReflectCAP), where a multi-agent pipeline analyzes what the target large vision-language model (LVLM) consistently hallucinates and what it systematically overlooks, distilling these patterns into reusable guidelines called Structured Reflection Notes. At inference time, these notes steer the captioning model along both axes -- what to avoid and what to attend to -- yielding detailed captions that jointly improve factuality and coverage. Applying this method to 8 LVLMs spanning the GPT-4.1 family, Qwen series, and InternVL variants, ReflectCAP reaches the Pareto frontier of the trade-off between factuality and coverage, and delivers substantial gains on CapArena-Auto, where generated captions are judged head-to-head against strong reference models. Moreover, ReflectCAP offers a more favorable trade-off between caption quality and compute cost than model scaling or existing multi-agent pipelines, which incur 21--36\% greater overhead. This makes high-quality detailed captioning viable under real-world cost and latency constraints.
arXiv cs.AI
Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning
We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.
arXiv cs.AI
IDEA: An Interpretable and Editable Decision-Making Framework for LLMs via Verbal-to-Numeric Calibration
Large Language Models are increasingly deployed for decision-making, yet their adoption in high-stakes domains remains limited by miscalibrated probabilities, unfaithful explanations, and inability to incorporate expert knowledge precisely. We propose IDEA, a framework that extracts LLM decision knowledge into an interpretable parametric model over semantically meaningful factors. Through joint learning of verbal-to-numerical mappings and decision parameters via EM, correlated sampling that preserves factor dependencies, and direct parameter editing with mathematical guarantees, IDEA produces calibrated probabilities while enabling quantitative human-AI collaboration. Experiments across five datasets show IDEA with Qwen-3-32B (78.6%) outperforms DeepSeek R1 (68.1%) and GPT-5.2 (77.9%), achieving perfect factor exclusion and exact calibration -- precision unattainable through prompting alone. The implementation is publicly available at https://github.com/leonbig/IDEA.