CILA: Confusion-Based Logit Adjustment for Long-Tailed Recognition
A new arXiv preprint introduces CILA, a per-instance logit adjustment method that aims to counter majority-class bias in long-tailed recognition.
Research Digest
Preprints and papers, rewritten in plain English.
New Benchmark Tests LLMs on Condensed Matter Spectral Data
CME-SpectrumBench aims to fill the gap in evaluating LLMs on experimental data analysis in condensed matter physics.
Concordance check for intervention-conditioned digital twins
A proposed plausibility check targets the counterfactual credibility of treatment predictions from human digital twins.
Can Compact LLMs Recover Missing Mobility Events from Trajectories?
A controlled study tests whether small, locally run language models can fill gaps in incomplete travel trajectories to identify air trips.
ConCI benchmark tests LLMs' contextual integrity with counterfactuals
A new benchmark uses controlled counterfactuals to assess whether language models judge information sharing by context, not just content.
When Joint, Decoupled, or Agent-Based Multimodal Models Win
A new arXiv paper compares three paradigms for combining modalities and asks when each approach is best.
Persistent Memory and Iterative Improvement in LLM Drawing Agents
A new arXiv study examines whether an LLM agent's accumulated memory from past attempts helps it solve new Chinese-character drawing tasks, and whether memory design matters.
Cache2LoRA: Testing How Well Context-to-LoRA Preserves Facts
A new preprint examines whether encoding documents as low-rank updates lets language models answer without keeping source text in context.
Self-Improving Video-to-Action Models for Embodied Agents
New research proposes video generative models as scalable visual planners that can improve their own action execution.
Training-Free Prompt Selection Cuts RLVR Costs
A new arXiv paper proposes a training-free prompt selection method using partial rollouts to reduce the computational cost of reinforcement learning with verifiable rewards.
Text-to-SQL Systems Struggle With Ambiguous Queries, New Paper Argues
A new arXiv paper highlights that natural-language queries often have multiple valid SQL interpretations, challenging the field's usual one-to-one assumption.
Tool-Using Agents Need a Stopping Rule for Uncertainty
A new arXiv paper examines how LLM agents should decide when to act, gather more information, or ask a human.
AM-Bench Targets Agentic Misalignment in LLM Agents
A new arXiv preprint proposes a unified taxonomy and evaluation suite for a failure mode where agents choose harmful actions over failure.
SteerSearch lets users redirect AI search agents mid-response
A new interactive method allows users to change a search agent's goal mid-generation, challenging the fixed-goal assumption in open training.
Cortex: A Supervisory Layer for Governed Coding Agents
A new technical report proposes Cortex, a supervisory layer that adds pre-execution governance, requirement tracking, and iterative repair to LLM-based coding agents.
LLM-as-a-Judge Can Debias Itself, New Preprint Finds
A recent arXiv paper suggests that LLM-based evaluators can reduce their own systematic biases without requiring additional external resources.
Screening LLM Judges Without Ground Truth for Dental Coaching
A new arXiv preprint proposes a staged pipeline to vet LLM judges and student models when no ground-truth labels exist, using an AI dental hygiene coach as the test case.
Fast Kernels Aren't Enough: Pipeline Latency Bottlenecks Physical AI
A new arXiv paper argues that optimizing individual kernels fails to address the real bottleneck in real-time physical AI inference.