This post is a daily digest of 20 arXiv AI/ML papers (announced July 8, 2026), auto-collected July 10, 2026. Below is an English summary of each paper.
1. Agon: Competitive Cross-Model RL (arXiv:2607.07690)
Agon lets two competing models serve as each other's implicit judges: both attempt the same problem in alternating drafter/solver roles, each rewarded for outperforming a rival that has seen its work. Reasoning is thus implicitly graded during training without process labels or reward models. Deployed as a two-stage cascade at inference, with Qwen3 on hard DeepMath problems it roughly doubles GRPO's pass@1 (about 8x the gain from untrained Mixture-of-Agents), with results reproduced on competitive programming and across model families (Qwen3.5, Gemma 4).
2. SciReasoner: Deep Native Structural Reasoning (arXiv:2607.07708)
A multimodal scientific foundation model spanning proteins, small molecules, and inorganic crystals. It discretizes coordinates, topology, and periodicity into a unified structure-aware vocabulary and treats structure tokens as addressable evidence. It raises F_max from 0.42 to 0.55 on homology-controlled gene ontology prediction, single-step retrosynthesis accuracy from 0.63 to 0.72, and achieves state-of-the-art on 67 of 86 benchmarks; blind expert review rated its reasoning traces better than or comparable to frontier LLMs in 98% of cases.
3. Co-LMLM: Continuous-Query Limited Memory Language Models (arXiv:2607.07707)
LMLMs externalize factual knowledge to a knowledge base instead of memorizing it in weights. Co-LMLM pairs continuous keys with textual knowledge values, generating flexible vector queries while retaining attributable retrieval. At 360M scale it beats models pretrained on 40x more data in perplexity and matches gpt-4o-mini on SimpleQA (above Claude Sonnet 4.5).
4. STRACE: Structural Trajectory Analysis (arXiv:2607.07702)
A framework for building high signal-to-noise optimization contexts for agent reflection: batch-level failure-pattern mining filters redundant trajectories, while causal localization on textual dependency graphs identifies true root-cause steps. It improved success rate 1.4x (42.5% → 58.5%) on VeruSAGE-Bench, optimizing a human-expert-designed agent. Code: https://github.com/moomight/STRACE
5. Jailbreak: Database Bypass via LLM-Synthesized Readers (arXiv:2607.07696)
Jailbreak bypasses database drivers by having LLMs synthesize readers directly from storage file formats (defined by source code and docs), materializing data as Apache Arrow buffers consumable by DuckDB, Spark, cuDF, and RAPIDS. Verified correct on TPC-H against JDBC/ODBC baselines for PostgreSQL and MySQL, with up to 27x end-to-end analytics speedup.
6. Institutional Red-Teaming (arXiv:2607.07695)
A method testing multi-agent deployment rules: fix agents/goals/tasks, change one rule, attribute behavioral shifts. In IABench-CA (228 contexts, 7 model populations, 33,924 games): rules causally shift average mortality by 22–58 percentage points; no universally safe rule exists and identity-targeted eliminations never proven safest; naming loss-bearers in rule text alone drove targeted elimination from 22% to 81%.
7. Sample-Efficient Diffusion RLHF (arXiv:2607.07693)
Two complementary strategies for diffusion RLHF: per-timestep weighting of denoising steps (linked theoretically to PPO's optimal convergence) and advantage-based replay prioritizing informative trajectories. Combined, up to 6x sample-efficiency improvement over standard diffusion RLHF baselines.
8. The Key to Going Linear (arXiv:2607.07706)
An analysis-driven transformer linearization under a frozen-backbone regime. Shows softmax implies key-dependent rank-1 orthogonal projections, explaining why delta-style networks beat gated accumulation. With pooling tokens, short convolutions, and fixed-budget cache routing, the method scales to LLaMA and Qwen models up to 32B, beating prior post-hoc baselines on MMLU.
9. ECGLight (arXiv:2607.07683)
A compute-light, on-device pipeline converting smartphone photos/scans of paper ECGs into calibrated 12-lead signals and screening for myocardial infarction with SHAP explainability. Trained/evaluated on 21,799 PTB-XL ECGs, it runs under 30 seconds per ECG on CPU-only resources, achieving 95.51% MI detection accuracy on PTB-XL (F1 = 0.9519) and 88.89% OMI accuracy on ECG-Matrix (F1 = 0.8862).
10. NOTES: Neural Operator Evolutionary Strategy (arXiv:2607.07682)
Couples a DeepONet neural operator with CMA-ES in a compact topology-aware latent space for PDE-constrained inverse design. On Maxwell-constrained nanophotonic beam deflectors it reduces design dimensionality from 256 to 25 with >95% efficiency, beating CMA-ES and topology optimization baselines.
11. Any-Dimensional Learning by Sampling (arXiv:2607.07680)
A unified framework using stochastic sampling maps (generalizing sampling with replacement, random binning, and species sampling) to compare inputs of different sizes, yielding explicit generalization and sketching rates for function classes on sequences, graphs, and tensors, including moment polynomials on measures, homomorphism densities, and permutation-invariant transformers.
12. How Data Shapes RoPE Frequency Usage (arXiv:2607.07678)
Proposes a data-centric account of RoPE frequency usage: frequencies are selected to match the relative-distance structure of training data, with an optimal scaling of 1/W for data-induced dependency contour width W. Explains the observed mid/low frequency band in LLMs and clarifies when positional-interpolation-style frequency scaling helps or fails length generalization.
13. SkillCenter (arXiv:2607.07676)
A large open agent skill library: 216,938 structured skills in 24 domain packages, with 114,565 source-grounded skills (traceable via exact citations to peer-reviewed papers, arXiv, and 24,000+ technical sources) plus 102,373 community skills, curated through the SkillGate LLM quality gate. Distributed as an offline-searchable SQLite FTS5 package.
14. LingBot-Video (arXiv:2607.07675)
A DiT-based video pretraining paradigm for embodied intelligence using Mixture-of-Experts for a better capability/efficiency trade-off, a data-curation engine adding robot-oriented footage (manipulation, navigation, egocentric), and a multi-dimensional reward system enforcing physical plausibility. Released as the community's first large-scale open-source MoE video foundation model.
15. AdaPrefix-GRPO (arXiv:2607.07674)
Turns correct-prefix hinting into a feedback controller that keeps each problem's success rate near 50% (where GRPO's gradient signal is maximal), then withdraws assistance. At matched training FLOPs it improves held-out accuracy 2.1x on 0.6B models, 1.6x on Qwen3-1.7B, 1.7x on AIME, while roughly halving trace lengths.
16. MedPMC (arXiv:2607.07673)
An automated, continuously updatable framework converting licensed PMC literature into high-fidelity medical multimodal data: 11M image-text pairs from 6.1M articles, with 95.3% of images rated medically relevant (vs. 19.7% in prior PMC-derived datasets). CLIP-style models trained on it beat biomedical CLIP baselines by 7.1 points average zero-shot AUC across 26 benchmarks, using less than half the pairs. Framework, corpus, benchmarks, and pretrained models are released.
17. PeTeR (arXiv:2607.07671)
A data-free, post-training framework that robustifies pretrained probabilistic circuits against distribution shift via distributionally robust optimization, matching or beating data-dependent robust training baselines across density-estimation benchmarks without retraining from scratch.
18. Bielik and Hallucination Detection (arXiv:2607.07670)
On four Polish Bielik models (1.5B–11B), unsupervised activation-dispersion measures separate known from fictitious entities with AUROC 0.95–1.00 before the answer token. Notably, entity familiarity saturates at 1.5B while factual reliability scales sharply (0, 2, 10, 19 of 42 known athletes answered correctly at 1.5B/4.5B/7B/11B), and models almost never abstain (2 refusals in 2,520 answers).
19. DiaLLM (arXiv:2607.07669)
A controlled study of English dialect adaptation (Australian, Indian, Northern UK): continued pretraining shapes benchmark robustness while alignment visibly reshapes generation. Explicit variety-targeted adaptation produces reliably recognized dialect output, but the most aggressively dialect-reward-optimized methods are not preferred by human evaluators, revealing a reward-quality gap.
20. Terminal-Fitted Repair for CFG (arXiv:2607.07665)
Analyzes classifier-free guidance numerically: guidance rigidifies the discriminative subspace with exponent 1+w, causing DDIM's guided residual to diverge as sigma_min → 0. A single-coefficient, zero-extra-NFE repair replaces CFG's w(r−1) with r^(1+w)−r, removing the divergence and giving 9/9 better FID points on tested grids (including Stable Diffusion 1.5 DDIM), though it is a high-guidance stabilizer rather than a universal quality knob.
*Auto-collected July 10, 2026.*