AI Preprint Pulse

Agent Evaluation Benchmarks

• 🔥 ClawBench: Can AI Agents Complete Everyday Online Tasks?: New paper on ClawBench evaluating AI agents on everyday...

Hot trend in autoregressive LLM inference: aggressive techniques for parallel decoding and multi-token generation.

• DMax introduces aggressive...

• RLVR stalls on hard problems: No success means zero learning signal, leaving issues unsolved. Cog-DRIFT breaks this exploration barrier.
• Agentic...

• KnowU-Bench evaluates interactive, proactive, personalized mobile agents.
• ClawBench probes if agents handle everyday online tasks.
• Trend signal: Specialized evals rising for proactive, real-world agent scenarios.

New preprint rethinks poor generalization in reasoning SFT via conditional analysis of optimization, data, and model capability roles. Essential read for understanding SFT bottlenecks.

Even in agentic search, core tech like BM25 (1994) + monoT5 (2020) lets a 20B agent rival GPT-5 (2025). SIGIR2026 paper "Revisiting Text Ranking in Deep Research" proves traditional IR remains highly competitive.

AI Safety Benchmarks

• DeonticBench: A Benchmark for Reasoning over Rules: New benchmark introduced for reasoning over rules.
• The Depth Ceiling:...

New preprint spotlight:

• Real-time 4D world simulator
• Via spatiotemporal autoregressive modeling
• Join the discussion

• Massive breadth: 10,000 training/eval environments from 200 unique software packages, spanning tasks from every major sector of the economy
• Fully...

• RLVR stalls on hard problems: every rollout fails, yielding no meaningful training signal
• Cause: Zero rewards in rollouts → no advantages → no gradient updates
• Cog-DRIFT overcomes this barrier for progress on tough tasks

New preprint SEVerA introduces verified synthesis methods for self-evolving agents, targeting reliable self-improvement vital for AI safety. Join the discussion.

New preprint 'The Depth Ceiling' spotlights the limits of large language models in discovering latent planning structures, urging deeper scrutiny of LLM planning depths.

DeonticBench launches as a new benchmark for reasoning over rules, probing LLMs on deontic obligations and permissions—vital for AI alignment.

RLVR Innovations

• 🔥 Cog-DRIFT: @EliasEskin shares Cog-DRIFT, enabling RLVR to learn from zero-reward examples on hard problems by reformulating...

MegaTrain introduces full precision training for 100B+ parameter large language models on a single GPU – slashing barriers to massive model development.

Free-Range Gaussians introduce a core idea: instead of predicting Gaussians on pixel- or voxel-aligned grids, they live freely in 3D space for...

Fresh arXiv preprint Learning to Retrieve from Agent Trajectories spotlights trajectory-based retrieval for agent enhancement. Join the discussion.

• Rapid excitement builds around Cog-DRIFT, tackling RLVR's failure on tough problems where rollouts yield zero rewards (pass@64=0).
• Standard...

New preprint introduces Action Images, an approach for end-to-end policy learning using multiview video generation. Join the discussion sparking interest in this innovative method.

ThinkTwice introduces a unified approach to jointly optimizing large language models for reasoning and self-refinement. Join the discussion on this new preprint.