Hybrid Blockchain Paper Proposes Balanced Architecture for Adoption
A revised arXiv paper argues that a balanced hybrid blockchain design could help overcome barriers to government and institutional use.
Research Digest
Preprints and papers, rewritten in plain English.
Interventional Step Credit Targets Efficient Reasoning
A new arXiv paper proposes using step-level signals to reduce wasteful chain-of-thought in large reasoning models.
Arabic NLP Methods for Linking Financial Figures to Narrative
A new arXiv paper outlines how Arabic natural language processing can be used to study the connection between numerical accounts and the explanations that accompany them.
MITRA-lexicon links Sanskrit attestations to digital texts
A new resource aims to make Sanskrit lexicography more transparent by connecting attested translations to digitized sources.
LEAPS Benchmark Targets Long-Context Pattern Learning and State Tracking
A new arXiv preprint introduces LEAPS, a benchmark for evaluating how long-context language models adapt to patterns and track state over long trajectories.
Truncated Reasoning Budgets: Cost vs. Accuracy in LLM Decisions
A new arXiv study evaluates how limiting reasoning tokens affects decision accuracy, highlighting System-1 models as a possible solution.
LLMs Accelerate Drift in Scientific Prose, arXiv Analysis Finds
A time-series study suggests that LLM-based writing tools are speeding up changes in the style of scientific texts.
Speech LLMs That Breathe: Modeling Non-Verbal Vocalizations
A new arXiv paper highlights the need for speech language models to account for non-verbal vocalizations and the norms governing human dialogue.
Study Probes Rubric Control in LLM Essay Scoring Across 25 Models
A new arXiv paper examines whether LLM-based automated essay scoring can be trusted for high-stakes testing by looking inside the models' rubric-following behavior.
DeepSEA Benchmark Targets Southeast Asian Language Gaps
A new evaluation suite aims to address the underrepresentation of Southeast Asian languages in language model testing.
Where RL Breaks: Targeted Pretraining to Repair Reasoning Failures
A new preprint argues that RLVR mostly sharpens existing skills and proposes targeted continued pretraining to fix where it fails.
Agent-Based Voice System Targets Rigid Menus
A new end-to-end voice interaction system, VISAGENT, aims to overcome the rigidity and fragmentation of current public-service voice platforms.
Balance-Aware Pivot Selection for Parallel Diffusion Decoding
A new preprint proposes balance-aware pivot selection to improve parallel token commitment in diffusion language models.
New Benchmark Tracks AI Detection of Indirect Self-Harm Language
A new dataset aims to measure how well AI systems recognize culturally situated, indirect expressions of self-harm distress.
Rethinking BPE: Choosing Tokenizer Merges as Model Selection
A new paper argues that the number of BPE merges should be chosen via model selection, not a fixed heuristic budget.
Heartbeat Thinking: Voice Agents Act While Users Are Still Speaking
A new arXiv paper proposes 'Heartbeat Thinking' to let realtime voice agents start tool use during speech, cutting post-request waiting.
Rethinking LLM Efficiency: Optimize Time-to-Answer, Not Just Decoding
A new arXiv preprint argues that the interval users actually experience—from request submission to answer—should be the primary optimization target, proposing streaming generation to improve it.
TempoMoE Targets Overthinking in Multimodal Reasoning Models
A new approach aims to cut token cost and latency by steering reasoning effort to match task difficulty.