Hand-picked recent LLM frontier papers, with practitioner notes on what to use and skip.
arXiv 2605.06208
2026-05-10
Jin Park, Aldous Foster, Maya Sundaram, et al.
Empirical bound on multi-step agent coherence: success rate drops sharply at the 7th tool call across Claude 4.7, GPT-5, Gemini 3 Pro. Mitigations via tool-trace summarization recover most of the loss.
arXiv 2605.05117
2026-05-09
Hyeonji Lee, Tara Mukherjee, Daniel Roa-Bell
Trace analysis of 14M production LLM requests: 73% of input tokens repeat across requests within a 5-minute window. Quantifies cost savings from prompt caching across providers — Anthropic 5-min TTL captures 81% of savings.
arXiv 2605.04421
2026-05-08
Sasha Rubinstein, Mira Wong, Ali Akbarian, et al.
Empirical scan of 200,073 public MCP servers across npm/PyPI registries: 14.3% expose tool descriptions vulnerable to prompt-injection hijack; 4.8% leak credentials in unfiltered logs. First quantitative MCP threat model.
arXiv 2604.24763
2026-04-27
Zhiheng Liu, Weiming Ren, Xiaoke Huang
Native unified multimodal model that encodes images with simple patch embeddings — no VAE, no separate vision encoder. Wins at scale on fine-grained perception.
arXiv 2604.22074
2026-04-23
Qinan Yu, Alexa Tartaglini, Peter Hase
Introduces CIR (Causal Importance of Reasoning) and SR (Sufficiency of Reasoning) metrics. RLVR improves accuracy but does not reliably improve CIR or SR.
arXiv 2604.19295
2026-04-21
Qingyang Zhang, Xinke Kong, Haitao Wu
Test-time training framework that interleaves policy refinement on unlabeled questions with periodic critic recalibration. OLMO3-7B 33.0% → 51.1% on AIME 2024.
arXiv 2604.18788
2026-04-20
Afsara Benazir, Felix Xiaozhu Lin
Runtime that handles MoE on Apple NPUs via offline calibration, static capacity tiers, and load-aware graph residency. 1.32-5.55× lower latency on M-series chips.
arXiv 2604.16529
2026-04-16
Joongwon Kim, Wannan Yang, Kelvin Niu
Argues test-time scaling for long-horizon coding agents is a representation problem. Reports Claude Opus 4.5 going from 70.9% to 77.6% on SWE-Bench Verified.
arXiv 2604.13120
2026-04-13
Rajesh Kumar, Waqar Ali, Junaid Ahmed
Proposes execution-grounded verification as a first-class principle: every code change must survive Docker-sandboxed execution. Reaches 40.0% on SWE-Bench Lite.
arXiv 2604.12710
2026-04-13
Junxiao Yang, Haoran Liu, Jinzhe Tu
Identifies a "semantic bottleneck" layer where representations are language-neutral. Anchoring alignment there drops attack success on LLaMA-3.1-8B from 24.7% to 2.8%.
arXiv 2604.10261
2026-04-11
Zae Myung Kim, Dongseok Lee, Jaehyung Kim
Builds a DAG-puzzle Wikipedia benchmark. Across 1,400 instances and three frameworks, the best system hits 37.2% — navigation errors dominate, not tool-use errors.