The Archive

Lawsuit alleges GPT-4o advised minor to stop seizure medication, resulting in death; raises LLM safety and liability questions.

Anthropic leaders detail Claude Opus 5.5 launch, API scaling, and managed agent architecture separating model brain from tool execution.

Large Language Models (LLMs) are increasingly used in multi-turn agentic workloads that require sustained interaction with complex environments. Existing reflective prompt optimizers such as GEPA learn from failed rollouts, but agentic failures can also come from prompt weaknesses, harness limitations, stochastic execution, or noisy judges. We argue that such failures may not provide useful learning signal, and may even be harmful; and thus it is essential to curate which tasks are provided to the prompt optimizer for better performance. We introduce GEPA-Curator, a framework that augments th...

Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can large language models (LLMs) notice unexpected behavior, such as heavier objects falling more slowly, and discover the law through experiments without being told what to investigate? To evaluate this ability, we introduce LongChang, an automated framework that creates simulated worlds with altered physics, where one or more laws differ from ordinary physics. The evaluated LLMs act as explorers, manipulating objects and building tools to investigate ...

We explore how to build performant large language model (LLM) tool-use agents that complete complex, multi-step tasks. Specifically, we study how data composition in synthetically generated tool-use environments affects the training of such agents. To do so, we propose the SynthAPI pipeline, which explicitly controls the diversity of generated environments along the axes of tools, tasks, and the databases on which the tools operate. By systematically varying SynthAPI's generation controls in 458,752 tasks across 849 trained models of four different sizes, we derive a scaling law that predicts...

MuonGS analyzes weight-norm control in optimizer Muon during LLM pretraining using accumulated effective learning rate.

Multi-State Linear Attention with generalized second-order routing improves efficiency of long-sequence LLM modeling over quadratic self-attention.

Point-in-time models and temporal variability score track how general vs. domain-specific knowledge is acquired across LLM pretraining.

CrossProfile benchmark reveals how LLM group memory heuristics (rare vs. relevant claims) fail on hidden-profile collaborative decision tasks.

VideoLVR enables latent visual grounding during chain-of-thought reasoning in video MLLMs to mitigate visual forgetting.

LCM-Promo applies LLM-based uplift modeling to promotion allocation; production A/B test shows 1% conversion gain, 6% cost reduction on churn reactivation.

Parametrized activation steering rescales intervention strength for consistent, interpretable control of LLM behavior at inference.

R package for semantic factor analysis using LLM embeddings to analyze psychometric scales; niche tooling for psychology research.

Continued-pretraining corpus embedding ecological and biosphere reasoning into LLMs via QLoRA; addresses gap in environmental decision-making.

SimpleLoRIF: gradient-based method to detect fine-tuning data that degrades LLM safety; scalable to million-example datasets.

PerceptRL: reinforcement learning approach to improve visual perception in MLLMs during post-training via multi-aspect reward design.

Audit of 41 LLM releases shows pretraining transparency declining across versions; Foundation Model Transparency Index tracks disclosure gaps.

Concordat library enforces finite-state protocol constraints on LLM agents at runtime, improving reliability of consequential actions.

Study of inter-agent communication regimes in trading swarms across 200 LLM-based agents; empirical findings on emergent coordination.

ETAMem decouples heterogeneous conversational memory for long-term LLM interactions; multi-view framework addressing entanglement and mismatch.

ChatGPT hallucinates false advice directing users to 404 Media for tech support, exposing gaps in LLM reliability and platform accountability.

NPC-RubricBench evaluates LLMs across six longitudinal cancer care phases using 941 patient-phase units from 405 nasopharyngeal carcinoma cases.

Connector-based Speech LLMs expand to new languages via targeted fine-tuning and TIES merging while preserving source-language ASR performance.

EBench evaluates GPT-6-Astra as a frontier MLLM agent controlling robots via direct control vs. policy harness, benchmarking capabilities against post-trained VLA and world action models.

Reva accelerates real-time vulnerability detection in LLM-generated code via contextual probing, reducing multi-minute reasoning loops to practical screening latency.

Study compares selection bias across Jev-style System One models, Jev reproductions, and generative LLMs using multi-option shuffling tests to evaluate whether decision structure reduces bias.

dSpark applies schedule distillation to teach diffusion LLMs when to commit tokens during parallel generation, prioritizing high-context predictions over high-probability ones for faster decoding.

Study quantifies claim smoothing in LLM research synthesis: models produce confident narratives that underreport source conflicts, discarding evidential complexity.

DECE benchmark: unified arena evaluating LLM agents across 10 Eurogames with shared strategic competition and resource constraints.