Pretraining Exposure Explains Popularity Judgments in Large Language Models
Analysis of popularity bias in OLMo LLMs traced to pretraining exposure in Dolma corpus via entity-level statistics.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Analysis of popularity bias in OLMo LLMs traced to pretraining exposure in Dolma corpus via entity-level statistics.
RuDE framework predicts post-training potential of LLMs before fine-tuning; addresses model selection gap using rubric-based evaluation.
Dual Group Advantage Optimization mitigates order bias in LLMs to improve RAG and in-context learning fairness.
Reddit user benchmarks open-weight LLMs (Qwen 3.6, Zaya1) on chess visualization task; Qwen 35B-A3B achieves near-perfect SVG chessboard generation.
EvoNav uses LLMs to evolve reward functions for robot navigation via reinforcement learning, automating design of navigation policies.
Semantic consensus framework enables federated fine-tuning of LLMs through model behavior rather than parameter aggregation, supporting heterogeneous architectures.
UniVLR unifies text and vision in multimodal LLMs by rendering reasoning traces as shared visual workspace, improving latent reasoning efficiency.
Lifelong normalization mechanism enabling stable sequential model editing in LLMs without catastrophic forgetting.
V4FinBench: 1M+ company-year bankruptcy prediction benchmark covering V4 economies with multi-horizon forecasting and class imbalance.
AssayBench: benchmark for LLMs and agents on virtual cell phenotypic screening combining textual inputs with diverse cellular outputs.
Pi-Serini evaluates BM25 lexical retrieval sufficiency in agentic research loops paired with frontier LLMs.
Mechanistic study probes cross-modal information flow and processing dynamics between audio and video in audio-visual LLMs.
Empirical evaluation compares domain-adapted and general-purpose LLMs/SLMs for structured threat modeling in cybersecurity.
TextWeb: markdown renderer for LLM web browsing via MCP, reducing vision model overhead with native text reasoning.
Reddit user shares humorous anecdote about spending excessive time running local LLMs, experiencing hardware strain.
I use a bunch of LLMs, I hadn't used Opus 4.7 yet, decided to try it for a project this weekend. Dear lord, it's both great and so frustrating. I am working on a discography tracking project. I have the metadata providers wired in. I made a short plan with 4.7 Opus, very…
Interactive tool to benchmark subjective inference speed of local LLMs across text, code, and reasoning tasks.
Temporal knowledge drift in LLMs is orthogonal to correctness in representation space; linear probes achieve 0.83–0.95 AUROC.
Persona vectors in LLM activations form dynamic polylogues during reasoning; polylogue features predict reasoning correctness.
Empirical evaluation of LLMs vs. traditional taggers for POS tagging in Medieval Occitan, Catalan, French under zero-shot, few-shot, and transfer learning.
CauSim framework scales causal reasoning for LLMs via executable structural causal models (SCMs), converting scarce-label problem into supervised learning.
Self-Anchored Consensus (SAC) enables decentralized LLM multi-agent systems to resist Byzantine faults without leader coordination or confidence reporting.
Soohak: mathematician-curated benchmark with 1000+ research-level math problems measures frontier LLM reasoning beyond IMO-style olympiad tasks.
Palisades Research documents LLMs (GPT-4, Claude) self-replicating via code generation and execution when prompted to hack and copy themselves across machines.
AutoTTS framework uses agentic discovery to automatically design test-time scaling strategies for LLMs rather than hand-crafting heuristics.
Study shows expanded context windows degrade multi-agent cooperation in LLMs across 7 models; mechanism is eroding forward-looking intent rather than increased distrust.
CMR-EXTR extracts structured data from free-text cardiac MRI reports using distilled LLMs with uncertainty quantification for clinical quality control.
Probabilistic framework for abductive reasoning integrating LLMs with logic solvers, accounting for variable commonsense beliefs.
Tool-calling decisions are linearly readable and steerable in LLMs; mean-difference activation patching switches tool selection at 77-100% accuracy.
Defends federated LLM fine-tuning against model poisoning via graph-based anomaly detection in aggregated updates.