Style, Not Self: Surface Cues Explain Zero-Shot Code Attribution by Large Language Models
LLMs cannot reliably recognize their own code (49-58% balanced accuracy), relying on surface style cues rather than ownership detection.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
LLMs cannot reliably recognize their own code (49-58% balanced accuracy), relying on surface style cues rather than ownership detection.
SWE-Flux benchmark evaluates LLM execution reasoning on 480 repository-level Python instances with gold labels from instrumented test execution.
Fine-tuning LLMs for translation causes catastrophic forgetting of MT-specific instructions; general mitigation methods fail to preserve formality/gender/length control on Llama 3.1/3.2.
Fixed-budget revision protocol with deterministic verifiers reveals LLM closed-loop revision success ranges 17.4%-99.8% across 19 models under exact feedback.
Evaluates five open-weight 7B-8B LLMs on Turkish domain documents in resource-constrained settings using industrial and public-sector benchmarks.
SkillGym framework internalizes human-written agent skills into LLMs via training on 2,756 executable task environments across 12 categories.
Compares nine LLMs and Jev on ContractNLI legal document inference, isolating decision consistency from answer accuracy.
FakeContextBench evaluates whether LLMs can distinguish when contextual information should override learned patterns, revealing gaps in in-context learning robustness.
NAFBench evaluates whether LLMs can adopt specified negation semantics across legal, regulatory, and medical domains.
Flash-dLLM optimizes KV caching and parallel decoding for diffusion LLMs via IO-aware techniques, addressing inference bottlenecks in non-autoregressive text generation.
LYRA identifies Proximity Trap: irrelevant nearby context overshadows distant evidence in long-context LLMs; proposes t-distributed relevance alignment.
On-policy distillation (OPD) fixes quantization exposure bias in sub-3-bit LLMs; restores math and code reasoning in low-bit models.
Theoretical physics study on learning-theoretic conditions for bosonic Gaussian states; unrelated to AI systems or LLMs.
Epistemological assessment of LLMs used in infant syntax acquisition research, examining dataset construction and model implementation in the BabyLM challenge.
Face/Off evaluation framework reveals LLMs overweight lexical cues in code comprehension tasks despite access to semantic program structure.
Survey of LLM security, privacy, and reliability risks in mobility/automotive sector, covering 1.5B vehicles and EU AI Act compliance.
Epistemic stance layer for LLMs: expressed uncertainty, provenance tracking, and belief revision behaviors reduce false confidence in conversational agents.
ClusterFewshot improves LLM few-shot prompt optimization by combining semantic task structure with utility scoring, outperforming random and metric-based selection in DSPy.
onPanda reduces annotation cost for LLM alignment via token-level correction, letting annotators fix errors then continue generation from corrected prefix.
Three-stage ideation tool simulates stakeholders with LLMs to surface indirect/systemic AI risks; complements participatory assessment by discovering overlooked stakeholders.
Instruction-tuned LLMs for argument mining jointly segment and classify argumentative components as generative task vs. pipeline/sequence-labeling baselines.
Answer-Basin Representation Hypothesis: concept linear structures in LLMs organized by continuation probability distributions, not directional probing.
Explanation-aware post-training quantization for medical LLMs preserves rationale quality alongside answer accuracy in multiple-choice QA.
Verification-driven fine-tuning approach enabling LLMs to generate syntactically correct FORM code for symbolic particle physics calculations with zero-shot failure.
ValueDiff proposes value-geometric KV cache eviction for LLMs with weak attention sinks, ranking tokens by value-vector L2 deviation from cache mean.
CROSS-MAP: privacy-preserving LLM inference via semantic decoupling while preserving structural reasoning cues.
Gemini Hacked Three Companies in First Known Breakout by Google’s AI Gemini finally caught up on Felony Bench ! Tags: security , ai , generative-ai , llms , gemini , accidental-cyberattacks
“It’s important for service members to understand the uncertainty inherent to LLMs," a GovAI research scholar warns.
Probe of Internal Recognition detects concealed knowledge in LLMs using forensic methods adapted from psychology.