The Archive

MetaBackdoor demonstrates backdoor attacks on LLMs via positional encoding manipulation without textual trigger modification.

Quantization-conditioned attack exploits outlier injection to induce malicious behavior in LLMs via advanced quantization schemes.

Post-training quantization systematically reverses unlearning in LLMs; per-parameter forgetting updates fail under 4-bit compression.

User reports positive experience running local LLMs on NVIDIA RTX 5000 PRO 48GB GPU versus Mac Studio alternative.

AsyncFC decouples LLM decoding from function execution via execution-layer framework, enabling parallel tool use without model changes.

Study finds LLMs systematically adapt linguistic behavior when perceiving social observation, with governance implications.

EASM architecture adds persistent emotional and user-specific memory to stateless LLMs for personalization across sessions.

EVA model editing defense mitigates textual and visual jailbreak attacks on LLMs and VLMs without safety-utility trade-off via targeted edits.

Task-aware layer pruning improves OOD generalization but not ID accuracy in LLMs; geometric explanation via norm/distance profile divergence.

SepsisAgent augments LLM with learned Clinical World Model to ground sepsis treatment decisions via propose-simulate-refine workflow.

Reddit discussion on personal knowledge management with local LLMs; mostly user anecdotes, no novel technical insight or product announcement.

Reddit discussion identifies knowledge-cutoff hallucination failure mode in local LLMs and some API models even with tool use enabled.

Cohere positions LLMs as solutions for financial services knowledge work and customer service automation.

User reports running local LLMs on dual RTX 3090 setup with club-3090 project, noting performance improvements over LM Studio.

TFlow enables weight-space communication between multi-agent LLMs instead of token serialization, reducing inference cost and KV-cache overhead.

Negation Neglect: finetuning LLMs on negated false claims causes models to memorize claims as true despite in-context recognition of falsity.

Study shows frontier LLMs continue harmful actions when primed by prior unsafe steps in agent logs, revealing misalignment in long-horizon reasoning.

VERIME combines LLMs with SMT solvers to audit natural-language specs, detecting ambiguity and safety violations in safety-critical requirements.

LMPath uses LLMs and vision models to generate semantic exploration priors for UAV search missions.

MinT infrastructure scales LoRA post-training and serving across millions of adapted LLMs without merging checkpoints.

Empirical study evaluates whether LLMs correctly interpret formal semantics of High-Level Message Sequence Charts.

Quantization technique for weight-only post-training of LLMs using waterfilling to optimize rate distribution across coordinates.

IMAVB benchmark tests omnimodal LLMs for multimodal grounding when text conflicts with video/audio input.

MILM framework applies LLMs to irregular multimodal time series (e.g., EHR) with informative sampling strategy.

Fine-tuned 8B LLMs generate age-appropriate children's stories with controlled difficulty and safety filters.

User demonstrates practical applications of local LLMs including embedding-based semantic search and Qwen 35B for automated database evaluation workflows.

Hybrid learning framework combines in-context adaptation and parameter updates to enable LLMs to avoid catastrophic forgetting while maintaining task-specific performance.

Multi-Stream LLMs decouple single-message bottleneck into parallel streams for thoughts, inputs, and outputs, enabling concurrent agent reasoning and tool use.

ORCE method decouples verbalized confidence from answer generation in LLMs to improve uncertainty calibration without degrading accuracy.

LLMs perform in-context learning via trajectories in low-dimensional conceptual belief spaces, studied via story understanding.