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 constrained contextual bandit algorithm for personalized LLM routing that balances inference costs and user preference.
Researchers propose measuring autoregressive language models' harmful reachability rather than relying solely on refusal rates.
Researchers propose transferring scale calibration updates between model representations without recipient-side optimization.
A preprint study demonstrates that LLM-based causal discovery agents rely heavily on semantic variable names rather than true structural inference.
Researchers propose MetaCurate to filter agentic coding trajectories for supervised fine-tuning, moving beyond simple success metrics.
Research reveals RL-tuned search agents can boost accuracy while hallucinating ungrounded evidence due to flawed credit assignment.
Researchers introduced a Tool Recommendation system enabling multi-step AI agents to dynamically access over 8,200 tools via three fixed interfaces.
Researchers propose a logic-guided graph diffusion model using First Order Logic to enforce structural constraints in generated graphs.
Researchers introduced a fact-level memory gating strategy using delayed feedback to optimize personalized context in AI agents.
Academic research demonstrates that LLM pricing agents can sustain tacit collusion in repeated Bertrand competition markets.
Research demonstrates cosine-similarity attacks can recover private item identities from language embeddings in recommender systems.
Research on edge GPU query-based segmentation decoders shows memory bandwidth bottlenecks invisible to standard FLOP analysis.
Researchers propose a method to rank vision-language models by domain-specific cue reliance rather than generic leaderboards.
Research on causal foundation models shows unit-level aggregation breaks coverage, challenging calibrated confidence intervals.
Researchers introduced RECAP, a framework for long-horizon AI agents to dynamically manage context history and prune failure records.
Researchers propose a conformal policy selection framework to prevent adverse individual outcomes when deploying AI decision models.
Researchers introduce PersonaCUABench to evaluate computer-use agents handling diverse user personas, constraints, and approval requirements.
A new arXiv preprint surveys and categorises Retrieval-Augmented Generation techniques to help navigate architectural selection.
Researchers propose QD-Rec, combining open-ended search and foundation models to diversify high-stakes recommendation decisions.
Researchers examine layer-wise representations of the WavLM speech model for spoofing detection using the ASVspoof 5 database.
Researchers propose HardEval, an execution-verified benchmark for assessing AI models on OS security hardening and compliance remediation.
Researchers propose RAW-Explainer, a post-hoc explainability method for path reasoning graph neural networks in knowledge graph tasks.
Researchers propose neural operators to simultaneously solve infinite families of convex programs with linear complexity in finance.
Researchers introduce Cast-R1, exploring whether smaller models can learn agentic time-series forecasting via trajectory distillation.
Researchers propose AOI, a multi-agent framework aiming to automate Site Reliability Engineering through failure-driven training signals.
Researchers propose CastFlow, an agentic time-series forecasting framework combining tool interaction and numerical prediction.
Researchers propose STRIDE, a reasoning-aware training method addressing the modality gap between time-series forecasting and LLMs.
A study on semantic invariance in reasoning benchmarks shows model rankings shift significantly when flawed paraphrases are removed.
New research explores the theoretical trade-offs between information loss and generalization in sparse attention mechanisms for Transformers.
Researchers propose an adaptive self-speculative decoding method to improve inference throughput for vision-language models.