Research Feed Page 11

Researchers developed a method to extract near-verbatim chain-of-thought reasoning traces from black-box reasoning models using adversarial API injection techniques.

Researchers evaluated how large language models weigh conflicting evidence, finding they often over-rely on numerical data and external tool forecasts.

The paper introduces TMI, a method that translates noisy, interleaved computer-use logs into structured, hierarchical task models.

The authors introduce ConceptGuard, a benchmark that evaluates how effectively LLMs can remove harmful knowledge while retaining benign functionality for dual-use concepts.

Thinkingbox introduces a sandbox and benchmark designed to evaluate how reliably LLM agents perform multi-turn business workflows by verifying persistent backend state changes.

The paper introduces a unified framework called WithEveryone that generates group images of five to ten specific people by integrating structured planning with identity-aware constraints.

The researchers developed a framework that improves how multimodal large language models follow natural-language rules when performing spatial planning tasks.

EXIMO leverages a vision-language model to decompose complex robotic tasks into smaller steps, improving the efficiency of training vision-language-action policies.

The paper introduces Daedalus-150M, a hybrid model combining depthwise convolutions and attention mechanisms to optimize CPU inference efficiency by reducing reliance on key-value cache growth.

The researchers found that storing skills at the subtask level rather than the task level improves the performance of LLM agents, especially when using natural language notes instead of code.

4DAnyone reconstructs high fidelity 4D human models from single casual videos by using geometric guidance and optimized multiview diffusion techniques to prevent structural drift.

The paper introduces a new benchmark and evaluation protocol to accurately measure whether video object removal tools correctly eliminate both the object and its associated physical side effects like shadows and reflections.

Task-CoEvolve optimizes LLM agent harnesses by dynamically selecting a subset of validation tasks to estimate full-set performance, significantly reducing search costs.

FlashPrefill V2 improves large language model serving performance by introducing a block-sparse attention mechanism that integrates natively with paged KV cache and continuous batching systems.

ContractScrub is a new benchmark designed to evaluate how accurately large language models detect errors in legal documents.

RoMAN-Flow introduces post-training optimization and distillation techniques to eliminate the sequential sampling latency inherent in autoregressive normalizing flows for robotic control.

FlowEvo is a framework that improves AI agent efficiency by automatically compiling successful task workflows into reusable skills to avoid redundant computation.

The paper introduces MemTrapBench to quantify how retrieved memory degrades LLM reasoning and proposes AdaptiveMem as a solution to prevent memory-induced cognitive traps.

The paper introduces a method to extrapolate optimal learning rates for Mixture of Experts models using small-scale proxy runs to avoid expensive full-scale sweeps.

Repo0 treats code repository generation as a continuous structural evolution problem rather than a one-shot planning task to improve software modularity and functionality.