The Archive

LLM-SR-UQ framework quantifies structural uncertainty in LLM-guided symbolic regression via conformal intervals.

TradeWar benchmark evaluates LLM agent negotiation via tariff policy execution with economic outcome metrics.

Logit corrections and temperature adjustments recover <10% of instruction-tuned LLM performance loss, showing limited efficacy.

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

Gradual pruning with sparse adaptation preserves LLM quality at extreme sparsity better than one-shot methods.

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

Stepwise Querying distills reasoning from black-box LLM teachers by extracting step-specific explanations without chain-of-thought access.

Cisco enables engineers to use unlimited AI tokens for legacy code modernization, demonstrating enterprise-scale LLM adoption for codebase migration.

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

Task Vectors decomposed into Task Recognition and Task Learning components via unembeddings, revealing interpretable mechanisms for steering LLM behavior.

Simple syntactic features outperform 20 of 24 open-weight LLM-based AI control monitors on backdoor detection via label leakage.

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.

SWIFT applies small-world network topology to sparse parameter tuning, improving feature interaction in LLM fine-tuning with fewer parameters.

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.

Pre-registered audit of 25 cognitive modules in LLM-agent systems shows module names do not guarantee actual computation execution.

Reliability-Weighted DPO improves LLM alignment robustness by learning weights for noisy preference comparisons.

ORION applies LLM advice to RL-based multi-domain 6G slice placement, addressing NP-hard online network scheduling.

SCRBench: 67.6k Python pairs for evaluating LLM semantic change reasoning beyond equivalence checking.

Compares LLM strategies (CoT, LoRA, SLM-enhanced) for extracting entity relations from academic paper abstracts and full text.

DAPS uses dependency graphs instead of similarity for retrieving lemmas in Lean 4 theorem proving with large proof libraries.

Neuro-symbolic framework combines LLM decoding with grammar refinement to guarantee formal specification satisfaction.

Wikimedia Foundation files amicus brief arguing LLM training on copyrighted works deserves legal protection to preserve open-web training data.

Guide on scaling AI-native engineering orgs: hiring practices, LLM-integrated tooling, culture, and development workflows for teams building with AI.

, the open-source AI penetration testing tool, can now run on any of the 400+ models on Vercel .StrixAI Gateway Strix can reach models from Anthropic, OpenAI, Kimi, GLM, Deepseek, and others through a single API with no markup, removing the need to sign up for multiple providers. Usage and cost are tracked across all models in AI Gateway. Strix selects the Gateway through , where the model is the standard AI Gateway ID behind a prefix:STRIX_LLM / vercel_ai_gateway/ See to configure it and get started today.Strix's AI Gateway setup guide Read more

CATT benchmark exposes LLM data attribution failures via fictitious knowledge with ground-truth labels.

ShallowDefend defends multi-turn LLM jailbreaks by exploiting concentrated safety tokens in early layers.

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

TextSeal2: distortion-free LLM watermarking using copula mixing survives speculative decoding without output latency.