Building Auditable AI Text Detectors with Sparse Autoencoders
Researchers propose a pipeline using sparse autoencoders to build more transparent and auditable AI-generated text detectors.
arXiv cs.CL — Computation and Language·22 Oct 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Researchers propose a pipeline using sparse autoencoders to build more transparent and auditable AI-generated text detectors.
A paper shows document ingestion tools strip visual cues from Word files and PDFs, causing enterprise AI agents to make confident errors.
Researchers introduced P2NPCBench to test whether LLMs distinguish polynomial-time from NP-complete problems under adversarial rewriting.
Researchers introduce Quant-Arena, a benchmark and leaderboard evaluating LLM agents on end-to-end quantitative trading research loops.
Researchers propose Environment Duels, an agentic benchmark evaluating frontier LLMs on creating and solving interactive environments.
Researchers propose Discord, using reward disagreement to allocate rollouts for tool-using language models in MCP environments.
Researchers introduced CALLIOPE, a top-down framework for discovering topic hierarchies in LLM conversations with differential privacy.
Researchers propose Banzhaf Resource Attribution to measure marginal contributions of inference resources for debugging agents.
Researchers introduced VeriEvolve, a framework generating executable evaluators and challenge states to train self-improving agents.
Researchers introduce a trace-level divergence framework to evaluate reasoning paths in multilingual large language models.
Researchers propose GRAM, a group-wise rank-aware model merging technique using subspace alignment for multi-task deployment under data privacy.
Researchers propose VeeReview, applying the V-Model to measure verification evidence in agent-authored pull requests.
An arXiv paper introduces an operational taxonomy and benchmark for reasoning-search-augmented LLMs to reduce static RAG hallucinations.
Researchers introduced a task model defining boundaries for agentic auto-purchasing based on price, reversibility, and product info.
Researchers propose BIRD-RL, scaling agentic reinforcement learning over stateful data environments to address database isolation overheads.
Research paper arXiv:2610.29520 investigates whether LLMs add value to retrieval-based tabular classification compared to distance-KNN.
Perplexity released pplx-embed-v2-late, a pair of 0.6B and 9B multilingual, multimodal late-interaction retrieval models.
Researchers propose Weighted Refusal Direction Optimization to improve robustness against activation-space interventions in aligned LLMs.
Researchers propose SPARE, a training-free method for allocating sparsity across layers in Mixture-of-Experts models using routing entropy.
Researchers introduced Typed Adversarial Contracts, a framework enforcing data quality rules on LLM-generated production records.
Researchers evaluate agentic literature-searching harnesses for biomedical claim verification, highlighting the mechanics of tool-use error.
Researchers introduced a measure-based framework for AI benchmark evaluation to identify difficulty distribution mismatches and coverage gaps.
ToM2Vec introduces continuous theory-of-mind tokens to improve multi-agent collaboration and reasoning under uncertainty in latent space.
Researchers test System One decision models using LLMs to return calibrated probabilities for verifying legal authority instead of text.
Researchers propose Agent Behavior Maps to evaluate LLM agent actions using log-likelihoods without environment rollouts.
Researchers introduce AnnoCaseLaw, a benchmark of 471 U.S. civil-negligence opinions to evaluate language model legal reasoning.
A position paper argues LLM capabilities represent context-directed extrapolation from priors, offering an evaluation checklist.
Researchers use chess to study whether LLM capabilities reflect genuine reasoning or sophisticated memorization of common states.
Researchers examine Generative Engine Optimization, showing how retrieval orders in LLM-based search heavily influence product recommendations.
Researchers introduced TabAgent, a framework to replace expensive LLM routing and gating calls with distilled tabular classifiers.