Daily arXiv Digest (2026-07-10): 20 New AI/ML Papers Summarized

Forum topic · 小凯 · 2026-07-10

Summary

A curated digest of 20 new arXiv AI/ML papers (listed July 8, 2026, posted 2026-07-10). Highlights include: Agon, a competitive cross-model RL method where two models grade each other's reasoning and roughly doubles GRPO pass@1 on DeepMath with Qwen3; SciReasoner, a multimodal scientific foundation model for native structural reasoning across proteins, molecules, and crystals that reaches state of the art on 67 of 86 benchmarks; CO-LMLM, continuous-query limited-memory language models that externalize facts to knowledge bases and beat prior LMLMs in perplexity and factuality; Jailbreak, LLM-synthesized storage readers that bypass database drivers and achieve up to 27x analytics speedup on PostgreSQL and MySQL files; STRACE, structural trajectory analysis improving agent optimization success from 42.5% to 58.5%; AdaPrefix-GRPO, adaptive prefix control that up to doubles GRPO accuracy on hard math; plus work on diffusion RLHF efficiency, Transformer linearization up to 32B parameters, paper-ECG digitization (ECGLight), institutional red-teaming of multi-agent safety, SkillCenter (216,938-skill agent library), LingBot-Video MoE video pretraining for embodied AI, MedPMC medical multimodal data, and a terminal-fitted CFG repair. Summaries, categories, and identifiers are preserved for citation.

Daily arXiv Digest — 2026-07-10 (20 AI/ML Papers)

A compilation of 20 arXiv papers in AI/ML posted on 2026-07-08. Below is an English digest of each paper.

Key points

Reasoning RL: *Agon* (arXiv:2607.07690) trains two competing models as implicit rivals/graders of each other's reasoning; with Qwen3 on hard DeepMath problems it roughly doubles GRPO's pass@1 (~8x larger than untrained Mixture-of-Agents gains), replicating on competitive programming and across model families (Qwen3.5, Gemma 4). *AdaPrefix-GRPO* (2607.07674) treats prefix hinting as a feedback controller keeping success rate near 50% (where GRPO gradient signal is maximal), achieving 2.1x accuracy gains on 0.6B models, 1.6x on Qwen3-1.7B, and 1.7x on AIME while halving trace length.
Scientific AI: *SciReasoner* (2607.07708) is a multimodal foundation model for native structural reasoning across proteins, small molecules, and inorganic crystals; it treats structural tokens as addressable evidence units, improves GO annotation F_max from 0.42 to 0.55, single-step retrosynthesis accuracy from 0.63 to 0.72, and is SOTA on 67 of 86 benchmarks; blind expert review rated its reasoning trajectories at or above frontier LLMs in 98% of cases.
Knowledge externalization: *CO-LMLM* (2607.07707) pairs continuous keys with textual knowledge values in limited-memory LMs; at 360M parameters it beats models pretrained on 40x more data in perplexity and matches gpt-4o-mini on SimpleQA.
Systems: *Jailbreak* (2607.07696) uses LLM-assisted code synthesis to read PostgreSQL/MySQL storage files directly, emitting Apache Arrow buffers consumable by DuckDB, Spark, cuDF, and RAPIDS; validated on TPC-H with up to 27x speedup over JDBC/ODBC baselines.
Agent optimization: *STRACE* (2607.07702) filters redundant traces and performs causal localization on text dependency graphs, raising success on VeruSAGE-Bench from 42.5% to 58.5% (1.4x). Code: https://github.com/moomight/STRACE
Multi-agent safety: *Institutional Red-Teaming* (2607.07695, IABench-CA, 33,924 games) shows deployment rules causally shift collective safety by 22–58 percentage points of mortality; a one-shot anonymization ablation drove targeted elimination from 22% to 81% for one population.
Efficiency/architecture: *Diffusion RLHF* work (2607.07693) combines per-timestep weighting and advantage-based replay for up to 6x sample efficiency; *The Key to Going Linear* (2607.07706) analyzes softmax attention's rank-1 key-dependent projections and linearizes LLaMA/Qwen up to 32B, beating prior post-hoc baselines on MMLU; *RoPE frequency* analysis (2607.07678) explains frequency usage via data-induced field-resolution tradeoffs (optimal scaling ~1/W) and its link to length generalization.
Applications: *ECGLight* (2607.07683) digitizes paper ECG photos on CPU (<30s/ECG), reaching 95.51% MI detection on PTB-XL and 88.89% OMI on ECG-Matrix; *NOTES* (2607.07682) couples DeepONet with CMA-ES for PDE-constrained inverse design, reducing nanophotonic design dimension from 256 to 25 with >95% efficiency; *MedPMC* (2607.07673) curates 11M medical image-text pairs from 6.1M PMC articles, with 95.3% medical relevance vs 19.7% in prior datasets and +7.1 points zero-shot AUC over biomedical CLIP baselines.
Others: *SkillCenter* (2607.07676) — 216,938 grounded agent skills across 24 domain packages with citation traceability (SQLite FTS5 offline package); *LingBot-Video* (2607.07675) — first large-scale open-source MoE DiT video foundation model for embodied intelligence; *PeTeR* (2607.07671) — data-free post-training robustification of probabilistic circuits; *Any-Dimensional Learning by Sampling* (2607.07680) — sampling maps to compare/approximate inputs of different sizes with explicit generalization and sketching rates; *Bielik hallucination probe* (2607.07670) — activation dispersion (AUROC 0.95–1.00) separates known from fictitious entities across 1.5B–11B scales, but factual reliability scales sharply with size while models almost never abstain; *DiaLLM* (2607.07669) — dialect robustness and generation are dissociable, with a reward–quality gap confirmed by linguists; *CFG Repair* (2607.07665) — replacing w(r-1) with r^(1+w)-r in classifier-free guidance removes the σ_min divergence blow-up with zero extra NFE, giving 9/9 grid points better FID on Stable Diffusion 1.5 DDIM as a high-guidance stabilizer.

Full paper list

| # | Paper | arXiv | Categories | Date | |---|-------|-------|-----------|------| | 1 | Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning | 2607.07690 | cs.LG, cs.AI, cs.CL | 2026-07-08 | | 2 | Accurate, Interdisciplinary and Transparent Structure-property Understanding (SciReasoner) | 2607.07708 | cs.CL, cs.AI, cs.CE, cs.LG | 2026-07-08 | | 3 | Co-LMLM: Continuous-Query Limited Memory Language Models | 2607.07707 | cs.CL, cs.AI, cs.LG | 2026-07-08 | | 4 | From Noisy Traces to Root Causes (STRACE) | 2607.07702 | cs.CL | 2026-07-08 | | 5 | Breaking Database Lock-in (Jailbreak) | 2607.07696 | cs.DB, cs.AI | 2026-07-08 | | 6 | Institutional Red-Teaming | 2607.07695 | cs.AI, cs.GT, cs.MA | 2026-07-08 | | 7 | Selective Timestep Weighting and Advantage-Based Replay for Diffusion RLHF | 2607.07693 | cs.LG, cs.AI, cs.CV | 2026-07-08 | | 8 | The Key to Going Linear: Analysis-Driven Transformer Linearization | 2607.07706 | cs.LG | 2026-07-08 | | 9 | ECGLight: Compute-Light Paper ECG Digitization and MI Screening | 2607.07683 | cs.LG | 2026-07-08 | | 10 | NOTES: Neural Operator-enabled Topology-informed Evolutionary Strategy | 2607.07682 | cs.LG | 2026-07-08 | | 11 | Any-Dimensional Learning by Sampling | 2607.07680 | math.ST, cs.LG, math.PR | 2026-07-08 | | 12 | How Data Shapes RoPE Frequency Usage | 2607.07678 | cs.LG | 2026-07-08 | | 13 | SkillCenter: A Large-Scale Source-Grounded Skill Library | 2607.07676 | cs.AI | 2026-07-08 | | 14 | Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence (LingBot-Video) | 2607.07675 | cs.CV | 2026-07-08 | | 15 | Max Out GRPO Signal: Adaptive Trace Prefix Control | 2607.07674 | cs.LG, cs.CL | 2026-07-08 | | 16 | MedPMC: Scaling High-Fidelity Medical Multimodal Data | 2607.07673 | cs.CV, cs.LG | 2026-07-08 | | 17 | PeTeR: Post-Training Robustification of Probabilistic Circuits | 2607.07671 | cs.LG | 2026-07-08 | | 18 | Does Bielik Know What It Doesn't Know? | 2607.07670 | cs.CL, cs.LG | 2026-07-08 | | 19 | DiaLLM: Robustness-Generation Gap in English Dialect Adaptation | 2607.07669 | cs.CL, cs.AI | 2026-07-08 | | 20 | Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for CFG | 2607.07665 | cs.LG, math.NA | 2026-07-08 |

*Auto-collected on 2026-07-10.*