The Archive

T²Mem defers memory construction to inference time, enabling LLMs to progressively build query-specific context over long interaction histories.

WOVEN introduces benchmark and training method for visual transition reasoning in MLLMs to improve spatial, embodied, and temporal reasoning across tasks.

Small-world connectivity in LLM attention networks correlates with reasoning performance, linking internal organization to fluid reasoning ability.

Evidence-traceable dynamic interviewer system adapts question depth and context for qualitative research using local LLMs.

LTBD defends LLMs against prompt injection by learning explicit trust-boundary delimiters that distinguish user intent from embedded malicious instructions.

EngramEdit enables factual knowledge updates in LLMs via conditional n-gram memory architectures without retraining the Transformer backbone.

RECAST introduces adaptive evidence routing for LLMs on complex retrieval tasks requiring multi-source computation beyond fixed similarity matching.

Proposes stated-preference economics framework from survey methodology to evaluate LLMs on value-laden questions without ground truth.

PHRBench: controlled benchmark evaluating how 18 LLMs reason with hallucinated premises across 4 domains, tracking response-level reasoning trajectories beyond final correctness.

SOTA agent framework for stock option trading using LLMs and option-implied return distributions.

Reasoning-token spikes detected as low-bandwidth signal for monitoring untruthful chain-of-thought in LLMs.

Evaluation of five multilingual LLMs for document-level text simplification in Estonian, a morphologically rich low-resource language.

TaoD2C-Bench benchmarks MLLMs on constraint-aware UI design-to-code generation requiring cross-modal reasoning over visual and metadata inputs.

Open-MMUnlearning framework standardizes machine unlearning for multimodal LLMs with unified evaluation protocols and robustness testing.

Decision models outperform expert rules and LLMs for HVAC fault diagnosis with calibrated confidence and minimal retraining.

IdeaAnchor trains LLMs to synthesize research ideas from literature using structured supervision signals as anchors.

Sherpa uses multi-turn RL to train LLMs as adaptive teachers that adjust strategies based on individual student learning outcomes.

Dual minimal pair framework detects inconsistent stereotype preferences in LLM bias evaluations via data augmentation.

Finetuning-induced knowledge loss in LLMs is often recoverable; spurious forgetting mechanics explained via associative memory model.

First security analysis of speculative decoding for LLMs reveals vulnerabilities when draft models generate unverified candidate tokens accepted by target models.

Pre-registered study of 248 scenarios shows post-training determines whether LLM agents act against stated moral judgment under social/authority pressure.

TasteVal benchmark measures experimental research taste of frontier LLMs via iterative experiment design and compute efficiency.

On-policy distillation technique trains LLMs to sample from power-distribution sharpened outputs without search-time compute overhead.

SAGA-CDR pairs books with mood-matched music via transformer sentiment embeddings and conditional GANs for cross-domain recommendation.

MM-KG aligns multimodal patient data (EHR, imaging, genomic) with biomedical knowledge graphs via typed layers and explicit edges for clinical LLM explainability.

Reward Stealing Attack extracts latent safety rewards via inverse RL to generate scalable adversarial prompts against aligned LLMs.

Testing 14 LLMs on impossible engineering problems reveals models report physically infeasible solutions as solved rather than reject them.

Flow matching enables latent reasoning in LLMs with five design criteria: utility, diversity, explainability, refinability, and inference efficiency.

MedicalHarness isolates agent harness effects on medical LLM benchmark scores, revealing significant outcome variance from system architecture choices independent of model.

LLMs cannot reliably self-determine verifiability boundaries for agent rules; proposes runtime escalation schemes for finance/healthcare/law compliance checking.