Can LLMs Discover Scientific Laws in Real and Parallel Worlds?
SCILAWS-BENCH evaluates whether LLMs can discover scientific laws via hypothesis generation and testing, using real and parallel worlds to avoid memorization.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
SCILAWS-BENCH evaluates whether LLMs can discover scientific laws via hypothesis generation and testing, using real and parallel worlds to avoid memorization.
SDARE-Bench benchmarks LLMs on stigma detection and response generation in conversational dyadic and group dialogue with 2,526 scenarios.
Empirical study on using LLMs to design video coding tools via generation-evaluation loops, tested on Planar mode intra prediction.
TempCloze: 1,521-video benchmark for temporal reasoning in Video-LLMs, minimizing language shortcuts via cloze-completion on egocentric clips.
Fisher geometry analysis reveals safety alignment fragility in LLMs under benign fine-tuning via selective output-side MLP resharpening.
HarnessDev benchmark evaluates LLM agents on infrastructure generation, shifting evaluation from task outputs to runnable code.
Masked-citation study shows LLMs cite less critically and with weaker rhetorical intent than human-generated scientific citations.
Self-Routing post-training framework routes LLM samples to GRPO, distillation, or regularization based on rollout correctness without external teachers.
DroneCATS benchmark evaluates MLLMs as generalist vision-language-action agents for drone control with full action-space prompting.
Latent Recurrent Thoughts framework enables chain-of-thought reasoning in continuous representation space with frozen LLM and auxiliary network refinement.
SinkPruner removes high-norm outlier visual tokens from MLLMs to reduce inference cost while maintaining multimodal understanding quality.
Analysis reveals LLMs misalign on Islamic finance questions when cultural cues override correct framework selection, exposing normative pluralism failures.
CoBRA learns tool-use boundaries in LLMs via counterfactual margins, optimizing when to invoke tools by estimating instance-level marginal benefit.
BLOOM-WILT uses logit tilting to efficiently elicit rare behaviors in deployed LLMs without model training.
S³Gym benchmark evaluates LLM agent self-improvement through coupled self-testing, self-judging, and learning from experience.
Stiefel-constrained Riemannian optimization enables parameter-efficient rotational activation steering to control LLM refusal behavior at inference.
New benchmark and synthetic training pipeline improve LLMs' self-modeling—ability to predict own behavior on counterfactual prompt edits.
MIOH: fine-grained multi-image object hallucination benchmark for MLLMs, diagnosing how visual complexity triggers factual errors across multi-image reasoning.
Retain-free identity unlearning in MLLMs via localized early-decoder MLP intervention, enabling post-deployment privacy removal without retain sets.
BiG-SURE: black-box uncertainty estimator for LLMs and VLMs using cross-temperature semantic agreement and bipartite anchor-probe graphs.
Synthetic data security risk: targeted bias injection into aligned LLMs via benign text exploiting subliminal learning.
Continued pre-training adapts LLMs to Swedish journalism domain with novel benchmark across six editorial tasks.
Q-Strata hierarchically allocates mixed-precision bitwidths across MoE LLM experts to optimize model-level quantization objectives.
Progress-augmented curriculum learning method for multi-task RL post-training of LLMs balancing task update magnitude with reward gains.
ScienceArena: olympiad-style benchmark from IPhO, IChO, IBO, USAPhO, USNCO 2025–2026 with process-credit rubrics to address contamination and saturation in frontier LLM evaluation.
Survey on tensor decomposition methods for LLMs covering tokenization, embeddings, pre-training, adaptation, inference, compression, and interpretability.
Multilingual study of solvability detection in LLMs extends ReliableMath to French and Greek, analyzing whether failures stem from belief or language-specific expression.
Study shows high-resource languages more reliably elicit mathematical reasoning in LLMs; proposes feature transfer from HRLs to activate latent LRL computations.
CE-MoE reduces training communication overhead in mixture-of-experts LLMs via heterogeneous layer patterns that concentrate expert capacity while maintaining depth.
ElephantBench probes whether LLMs retain divergent long-tail facts via 1,094 multi-account QA questions from low-exposure web corpus.