The Archive

SFTProbe accelerates supervised fine-tuning data generation for reasoning LLMs from weeks to hours via principled sampling.

LLMs generate coherent comedy improvisations but fail spontaneous dialogue features; analysis via linguistic measures and annotation.

DecisionRank distills ranking knowledge from generative LLMs to decision models via on-policy distillation, reducing decoding latency.

ACCT enables KV cache reuse across different-scale LLMs in same family, reducing context processing overhead in heterogeneous deployments.

Zyphra introduces tree-structured hierarchical indexer achieving asymptotically subquadratic sparse attention for long-context LLMs.

Audit reveals persona-dependent variation in smoking-cessation health advice across six LLMs, raising fairness concerns.

SecPO applies adversarial training to defend LLMs against prompt injection by approximating injection-invariant reference models.

Framework aligning MLLMs with human cognition latent structure using 66-dimensional behavior-derived semantic embeddings from 4.7M Odd-One-Out judgments.

Seven frontier LLMs evaluated on molecular property prediction; most memorized ESOL but demonstrate in-context learning on clean benchmarks with self-reported uncertainty.

Repetition penalty parameter enables precision-recall trade-off control in LLM-based grammatical error correction via inference-time decoding.

Study examines whether reasoning in LLMs increases agreement or diversity, addressing monoculture vs. multiplicity trade-offs.

Analysis of discourse-level side effects from LLM activation steering, showing unintended changes beyond perplexity metrics.

Framework for targeted LLM unlearning of sensitive facts via reinforcement learning while preserving general capabilities.

Benchmark for temporal reasoning in LLMs: tests whether models can ground causal claims with statistical rigor over clinical records.

Framework measuring knowledge leakage in persona-conditioned LLM responses during user simulation tasks.

H²R: hierarchical hint-based reward framework using reinforcement learning for multi-step mathematical reasoning in LLMs.

Evidence-gated continual learning prevents LLMs from absorbing falsehoods at higher rates than random selection via surprise gating.

Fine-tuning open-weight models with five self-concept variants reveals how LLMs interpret and express identity.

LLMs interpret domain semantics to guide simulation-based engineering redesign, reducing computational cost of optimization loops.

Sleep framework enables LLMs to retain knowledge via episodic memory replay during continual learning.

Automated workflow converts natural language to power electronics schematics using LLMs with constraint verification.

TGTV framework integrates spatiotemporal graph learning with LLMs for interpretable UAV fault detection and diagnosis.

Critical evaluation reveals CLadder benchmark's procedural-generation artifacts may artificially inflate LLM causal-reasoning scores.

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 ...

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

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

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

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