The Archive

Contrastive Activation Addition (CAA): training-free steering vectors for SpeechLLMs enable task-specific specialization without retraining.

Nullspace projection method extracts persona subspaces in LLMs for fine-grained behavioral control beyond single-direction steering approaches.

Empirical study: LLMs distort user stance in mediated communication; measures stance preservation rate across generation-extraction pipelines.

Uses LLMs to predict consequences of curriculum learning decisions in RL by exploiting richer learning signals than local gradients.

Cross-modal alignment transfer technique enables MLLMs to improve on data-scarce modalities like audio and video via model merging.

LESSER selects post-training data for LLMs using only output-layer gradients instead of full-parameter gradients, reducing computational cost.

LLM mathematical reasoning via search exhibits expert performance but fails to generalize abstraction levels across subfields.

JOVE assigns heterogeneous LLMs to task graph subtasks with asynchronous verification to optimize latency and cost trade-offs.

Mechanistic analysis of vision-language fusion in MLLMs via attention routing, alignment decoupling, and dimensionality reduction across architectures.

Task-vector merge collapse in LLMs: variance-based prediction and repair using pre-merge interference scoring.

Trigger-tag mechanisms for misuse detection in open-weight LLMs lack systematic robustness analysis; practical security implications explored.

Marginal attention weights across 60+ LLMs exhibit emergent structure conserved as token-intrinsic properties independent of model architecture.

On an afternoon in Seoul in March 2016, I watched a program I helped build put a stone on the fifth line of a Go board in what looked like a gift to its human opponent. Move 37 in game two of the five-game match looked so absurd that some commentators thought it was a…

KaliBench evaluates LLMs on executable Kali Linux cybersecurity tool commands with verifiable rewards.

Study diagnoses structural mathematical understanding in LLMs via Mathematical Primitives benchmark, proposing methods to improve post-training on mathematical reasoning.

On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher...

LLM2Jev framework extracts calibrated categorical decisions from LLM next-token probabilities without fine-tuning, enabling direct software integration.

Causal Memory Policy restores identifiability of memory utility in LLMs by intervening on retrieval rather than storage decisions.

Span-level hallucination detection framework using hidden state probing to identify structured semantic drift in LLMs.

TRACE bounds multi-turn jailbreak risk in LLMs via trajectory analysis, showing single-turn safety training alone cannot prevent multi-turn exploitation.

Federated LLM fine-tuning over mobile RANs requires coordinated transport for asynchronous model updates with shared deadlines.

Robustness evaluation of Japanese LLMs against IME-related and typographical errors across character transposition, replacement, and homophone perturbations.

DARS: dependency-aware reward shaping for RL-trained LLMs, assigning step credit via prerequisite relations to handle cascading failures.

Prompt optimization for MLLMs in clinical diagnosis using AUROC instead of accuracy to handle class-imbalanced data.

Weight tying in decoder-only LLMs (GPT-2, DistilGPT-2) under DP-SGD privacy training shows diminished benefits vs. non-private settings.

DIY framework applies five cognitive bias-mitigation interventions from psychology to LLMs via in-context learning, instruction tuning, and self-revision.

Reproduces recent AMR-augmentation claims for LLMs; finds text-only baselines match AMR gains under controlled hyperparameter protocol.

Multi-expert MoE framework for opioid overdose risk prediction via continued pretraining on diagnostic sequences; clinical application of LLMs.

Token-level reward signals using visual dependency and entropy metrics improve multimodal reasoning in LLMs via reinforcement learning.

LexReward: taxonomy-driven reward framework for legal LLMs assessing response quality across Style, Element, and Chain dimensions.