AI

  • arXiv — Study of coding-agent harnesses finds context management matters more as compute budgets tighten; planning and action-space choices should be tuned per model, not fixed globally.
  • arXiv — SoL-Pi applies recursive self-improvement at the harness layer, cutting coding-agent token traffic 44.7-49.0% and API cost roughly a third with flat performance.
  • arXiv — Prewritten task-specific plans lift stateful LLM-agent success ~7.17pp (mostly on complex tasks); a cheap verification-only terminal already rejects 61% of invalid episodes.
  • arXiv — Chronicle records LLM-agent nondeterministic boundaries as immutable envelopes for bit-stable “cut-point” regression replay, needing zero model calls and catching all mutants in guarded tool code.
  • arXiv — DeepSeek-V4.1-Flash combines cross-layer KV-cache reuse with FP4 quantization to shrink KV cache to ~1/4 of its predecessor while improving agentic-workload performance.
  • arXiv — Across 450k gender-directed GPT completions, safety training shifted rather than removed discrimination: sexual-violence content fell but women-directed completions lost 36% topic diversity vs men, showing toxicity scores are an unreliable harm-reduction proxy. Note: analysis covers generations up to GPT-4, not current models.
  • arXiv — LLM agent groups replaying human deliberation overstate consensus by 34-44pp versus real human groups, and simulated consensus doesn’t track collective accuracy.
  • arXiv — dQwen3.5 adapts a hybrid attention+RNN model into a diffusion LM, reaching a given training loss in about half the tokens of a full-attention baseline at comparable quality.
  • arXiv — “Score centering” corrects training-inference mismatch in off-policy RL, matching or beating importance sampling across 0.6B-30B models as mismatch grows.
  • arXiv — Embedding models only weakly encode physical quantities (mass, distance, time, volume); representations track surface string patterns, not measurement semantics.
  • arXiv — JEPA-Anything, a factorized predictive world-model framework, beats matched baselines across vision, biology, and weather, yielding experimentally validated biological insights.
  • arXiv — RetireOPD’s on-policy distillation retires teacher supervision once gains plateau, with the student ending up surpassing the teacher on multiple agentic benchmarks.