AI papers — 2026-07-15

Problem: Designing a reward model for image-generation RL that is both reliable and free of expensive human-preference annotation or reward-model fine-tuning remains an open problem.

Method: The authors propose SpectraReward, a training-free reward that feeds the generated image as visual condition into a frozen pretrained MLLM and runs a single teacher-forced forward pass over the original prompt, using the mean image-conditioned prompt token log-likelihood as the scalar reward (no scalar judging or VQA-style question decomposition). Self-SpectraReward is a special case for unified multimodal models (UMMs) where the model's own understanding branch scores its own generation branch, forming a closed-loop self-improving setup with no external reward model. It is validated with the AWM RL algorithm on BAGEL and SD3.5-M across nine reward-MLLM backbones (Gemma3, InternVL3.5, Qwen3-VL, 4B–235B params).

Key results:

• SpectraReward improves BAGEL on TIIF-Bench overall short/long prompts by +10.0/+6.2 at 512 resolution, and beats AlphaGRPO by +6.3/+5.3 on TIIF-Bench short/long and +3.3 on GenEval.
• Self-SpectraReward reaches 89.8 on GenEval and 34.3 on GenEval2 at 1024-resolution inference, and 0.76 on the WISE (CoT) benchmark, outperforming both the BAGEL baseline and AlphaGRPO.
• Self-SpectraReward (using BAGEL's own 7B understanding branch) matches or beats the 30B and 235B external Qwen3-VL reward models (e.g., +1.2 GenEval, +2.1 GenEval2, +2 WISE vs. best external backbone), showing scaling the reward MLLM is non-monotonic (235B underperforms 30B).
• Ablations: scalar 1–5 scoring drops GenEval to 77.7 (vs. 89.5 for prompt-likelihood reward); sequence-level advantage outperforms token-level advantage.

Why it matters / caveats: It shows reward-policy distributional alignment (same tokenizer/vision encoder/pretraining distribution) can rival or beat sheer reward-model scale, suggesting cheaper self-rewarding UMMs can replace large external judge models for image-generation RL.

Problem: Existing benchmarks fail to measure "blind spots" — tasks that are trivial for humans (e.g., counting objects, drawing a dog with five legs) but that frontier AI models still get wrong.

Method: The authors built blind-spots-bench, 235 manually curated and annotated questions sourced from ~287 raw problems submitted by graduate AI-course students who identified questions frontier chatbots failed (October 2025), each tagged with a structured reference solution and a taxonomy of three categories (object-centric, abstract reasoning, language-and-knowledge) across three question formats (text-only, multi-to-text, image-generation). An automated Inspect-AI grading pipeline uses gemini-3-flash (with code execution) as grader, validated against human judgments, to evaluate 32 LLMs/VLMs and 6 image-generation models.

Key results:

• Grader-human agreement: 96.6% on textual-output tasks, 90.9% on image-generation tasks; no substantial pro-Google grading bias detected.
• Gemini-3.1-Pro is the best overall model: ≈83.3% accuracy on text-only, ≈66.9% on multimodal; GPT-5.5 scores ≈84% on text but only ≈58.7% on visual questions.
• Best open-weight model (GLM-5.2) reaches ≈74% text-only accuracy, roughly a 10-percentage-point gap below the top closed-source models, even at comparable scores on other benchmarks.
• Best image-generation model: Gemini-3-Pro-Image at 54.8% accuracy (vs. GPT-Image-2 at 51.2% but ~4× cheaper); fine-grained perceptual counting stays below 60% accuracy for all models.
• Enabling code-execution tools has mixed effects (e.g., Gemini-3.1-flash-lite +9.03 points, GPT-5.4 −5.32 points).

Why it matters / caveats: Blind spots persist independent of general benchmark performance (e.g., Artificial Analysis Intelligence Index), and closed-source models systematically outperform open-weight ones on this stress test even when they're comparable elsewhere — suggesting broad public benchmarks may overestimate robustness on underrepresented but "easy for humans" task types.

Problem: LLM coding agents make factual errors on software-issue resolution because they lack sufficient repository-internal understanding (cross-module dependencies, API contracts, data flow) before attempting a fix.

Method: ACQUIRE (Agent Collaboration for Question-Answer-driven Issue REsolution) splits repair into two stages: a Questioner decomposes the issue into targeted questions from a 4-category taxonomy (Mechanism & Behavior, Design & Usage, Locating & Structure, Ecosystem & Standards, derived from analyzing 116 oracle-QA cases), N Answerer agents independently explore the repository in parallel (read-only, same shell environment as Mini-SWE-Agent) to produce evidence-grounded QA pairs, and a Resolver receives the issue plus the pre-injected static QA knowledge set before running its own iterative repair loop. Evaluated on SWE-bench Verified (500 issues) with DeepSeek-V3.2 and GPT-5-mini backbones, N=2 QA pairs by default.

Key results:

• ACQUIRE reaches 70.8% Pass@1 with DeepSeek-V3.2, a +4.4 percentage-point gain over the Mini-SWE-Agent baseline (66.4%), and +3.8 points over GPT-5-mini's baseline (58.4%→~62.2%), beating LocAgent, CoSIL, LingmaAgent, and SWE-Debate on both backbones.
• An oracle-QA injection experiment recovers 26 of 116 previously-failed Mini-SWE-Agent instances using just one privileged question/answer.
• Ablations: replacing question decomposition with a single-pass proposal drops Pass@1 to 66.0% (−4.8 pp, even below the no-QA baseline); removing the category-guided template drops it to 67.0% (−3.8 pp).
• Sweeping N: N=1 already gives 69.0% Pass@1 for only +$0.005 cost; N=2 peaks at 70.8%; N=3 falls back to 69.0% (diminishing/overlapping returns).
• Cost/time overhead is modest: ~$0.030/$0.018 extra and ~115s/227s extra per instance for the two backbones respectively; the QA stage is parallelizable across Answerer instances.

Why it matters / caveats: Explicit, structured pre-repair knowledge acquisition — rather than fix-oriented localization or structural summaries — is shown to be a more effective and generalizable lever for agentic bug-fixing than prior pre-repair exploration methods, at modest added cost.

Problem: A large body of deep RL research implicitly assumes algorithm performance rankings are monotonic across sample-complexity regimes (low-data vs. asymptotic/high-data), and the paper investigates whether this assumption is valid and what conclusions it has distorted.

Method: The paper provides a theoretical sample-complexity analysis (via a proposition on distributional value-learning error and total-variation-distance bounds for C51 with fixed support size k, and unknown-support algorithms like QRDQN/IQN) showing the number of samples needed for correct action-ranking scales differently across regimes. It backs this with a large-scale empirical study on the Arcade Learning Environment (ALE), comparing DQN, Double-Q, Dueling, Prior (prioritized replay), C51, QRDQN, and IQN in both the 100K-interaction low-data regime and the 200-million-frame high-data regime, using Haiku/Optax/RLax.

Key results:

• Human-normalized median score in the ALE-100K benchmark: Dueling 0.2304±0.0061 — highest among all tested algorithms, ahead of C51 (0.0941±0.0081), Double-Q (0.0920±0.0181), QRDQN (0.0820±0.0037), Prior (0.0840±0.0018), IQN (0.0528±0.0058), and DQN (0.0481±0.0036).
• The simple 2016 Dueling baseline — historically excluded from many low-data-regime comparison benchmarks — performs 15% better than the DRQ-NeurIPS reimplementation and only 11% worse than DRQ-ICLR, contradicting an originally reported 82% gain for DRQ over the baseline.
• The theoretical proposition shows unknown-support models (QRDQN/IQN) can require asymptotically more samples (Ω(M/ϵ²)) than fixed-support C51 (k/ϵ²) to reach the same total-variation accuracy needed for correct action ranking.

Why it matters / caveats: The paper argues systematic omission of core baselines under a false monotonicity assumption has produced incorrect conclusions across several years of low-data-regime RL research, and proposes five principles (e.g., mandatory direct comparison to core algorithms) for a more reliable evaluation framework.

Problem: General-purpose autonomous ML-engineering (MLE) agents don't transfer well to medical imaging model development, which requires modality-specific method choices and strict, verifiable validation/prediction-artifact protocols.

Method: AMID is an autonomous multi-agent framework combining (1) Data-Conditioned Method Planning (DCMP), which profiles task data (modality, geometry, label structure, evaluation metric/submission format) and grounds candidate approaches in resources like nnU-Net, MONAI, and a foundation-model registry (MedSAM, MedSAM2, TotalSegmentator, VISTA3D) to build parallel, executable "method lanes"; and (2) Verification-Guided Two-Stage Optimization, moving from broad behavior-gated exploration (with an independent reviewer certifying validation protocol, metric computation, and artifact validity before any attempt counts) to selective exploitation of promoted candidates, run over a self-organizing multi-agent loop (backend-agnostic: Codex, Claude Code, OpenCode, Cursor Agent, Kiro) managed by a central lifecycle manager.

Key results:

• Across the 20-task ReX-MLE benchmark, AMID (Codex+GPT-5.5) produces valid accepted results on every task and beats the strongest baseline (AIDE, ML-Master, or R&D-Agent) on 19 of 20 tasks, tying on TopCoW-CTA-Cls.
• Segmentation gains: SEG.A Dice +0.89 (0.91 vs. best baseline 0.02), ISLES'22 Dice +0.67 (0.71 vs. 0.04); previously FAIL/0.00 PUMA tissue-segmentation runs become valid Dice scores of 0.56.
• Detection gains: DENTEX AP +0.40 (0.49 vs. 0.09, matching the 0.40 human reference score); PUMA-T1 nuclei-detection F1 +0.46 (0.54 vs. 0.08).
• Under a backend-matched comparison (all agents on GPT-5.5): AMID reaches 0.49 AP vs. best baseline 0.08 AP on DENTEX, 0.56 vs. 0.52 Dice on PUMA-T1-Seg; with Claude Code (Opus-4.8) backend AMID reaches 0.64 Dice / 0.50 accuracy on PUMA-T1-Seg / TopCoW-MRA-Cls.
• Weaker areas: graph classification, small-vessel-structure analysis, and ultrasound enhancement — AMID does not reach expert-level performance uniformly across all 20 tasks.

Why it matters / caveats: The paper positions rigorous, independently-reviewed verification (not just score optimization) as essential for turning bespoke medical-imaging engineering into a trustworthy, auditable agentic workflow; the authors note this is "an ongoing preliminary technical report."