arXiv Daily Digest (2026-06-15): 20 New AI/ML Papers

Forum topic · 小凯 · 2026-06-17

Summary

A daily roundup of 20 new arXiv papers in AI and machine learning, covering NLP, computer vision, robotics, security, and applied mathematics. Highlights include: language models encoding an internal 'value axis' that can be read and steered to control self-correction behavior; ContextRL, a context-aware RL objective that improves grounding across long-horizon reasoning and visual question answering benchmarks; BRDFusion combining physics-based and generative rendering for urban inverse rendering; an exact closed-form posterior score for linear Gaussian inverse problems; geometric foundation models repurposed as shared backbones for robotic manipulation; a backdoor attack achieving 90.3% success rate against differentially private federated learning; KVEraser enabling efficient local context erasure in long-context LLMs with only 24% latency overhead; Qwen-RobotWorld, a unified embodied world model built on Qwen2.5-VL; and HAMON, a passive diffractive-optics approach to long-horizon time series forecasting. The digest includes links, research domains, and concise findings for each paper.

arXiv Daily Digest — 2026-06-15 | 20 New AI/ML Papers

Key points

1. The Value Axis: Language Models Encode "Am I on the Right Track"

arXiv: 2606.17056 | Field: NLP

Language models internally encode a "value axis" representing the expected success probability of the current policy. Steering toward high value suppresses self-correction and reduces explanatory verbosity; steering toward low value induces backtracking and exploration. DPO training strengthens the internal value of reward behavior. Notably, Qwen3 assigns low value to politically sensitive queries — meaning internal model confidence can be read and manipulated.

2. ContextRL: Context-Aware RL for Agents and Multimodal LLMs

arXiv: 2606.17053 | Field: NLP/CV

Instead of directly supervising final answers, the model must choose which of two highly similar contexts supports a given query–answer pair. This indirect auxiliary objective forces fine-grained grounding: +2.2% average across 5 long-horizon reasoning benchmarks and +1.8% across 12 visual question answering benchmarks. Key insight: gains come from the context-selection objective itself, not the contrastive data.

3. BRDFusion: Physics × Generative Inverse Rendering for Urban Scenes

arXiv: 2606.17049 | Field: CV

A unified framework combining physically based rendering (controllable but artifact-prone) with generative models (realistic but inconsistent). The physics model provides controllable rendering; the generative model handles denoising and inpainting. Supports novel-view relighting, night simulation, and dynamic object insertion/editing. Code: https://shigon255.github.io/brdfusion-page/

4. Exact Posterior Score for Linear Inverse Problems

arXiv: 2606.17048 | Field: ML/CV

First derivation of a closed-form exact posterior score for linear Gaussian inverse problems under general Gaussian interpolation. Posterior sampling is equivalent to denoising at operation-dependent shifted pivots. The EPS training objective preserves standard pretraining input/output structure and supports training from scratch or fine-tuning. Outperforms across 5 inverse problems on FFHQ and ImageNet with ~10× fewer denoiser evaluations than gradient-based methods.

5. Geometric Action Model: Robot Manipulation with Geometric Foundation Models

arXiv: 2606.17046 | Field: Robotics/CV

Reuses a pretrained geometric foundation model (GFM) as a shared backbone for perception, temporal prediction, and action decoding. The network is split at intermediate layers: shallow layers encode observations, a causal future predictor forecasts future latent state tokens, and remaining GFM blocks decode actions. Preserves rich 3D geometric priors while enabling language-conditioned temporal world modeling; outperforms baselines on simulated and real robots.

6. HABC: Hierarchical Advantage-Weighted Learning from Sparse Robot Feedback

arXiv: 2606.17043 | Field: Robotics/RL

In online RL with binary success/failure feedback, prior methods compress sparse outcomes into a single scalar reward, conflating feasibility and efficiency. HABC trains two critic heads (feasibility and efficiency) merged via state-adaptive gating. On three bimanual contact tasks, SFT baselines of 36%/44%/12% success improve to 92%/88%/38%.

7. MetaSyn: Evaluating LLM Agents with Nature-Paper Meta-Analyses

arXiv: 2606.17041 | Field: NLP/IR

A benchmark of 442 Nature Portfolio meta-analysis articles, each with research questions, PI/ECO criteria, a 140K-PubMed retrieval corpus, validated positive studies, and hard negatives. Reveals a key bottleneck: retrieval recall ceiling is 90.9%, yet no system recovers more than 52.7% of ground-truth literature. LLMs fail to reliably distinguish studies meeting PI/ECO criteria from topically similar but non-qualifying ones.

8. R2RDreamer: 3D-Aware Data Augmentation for 2D Manipulation Policies

arXiv: 2606.17040 | Field: Robotics/CV

Applies lightweight 3D augmentation on incomplete point clouds (editing trajectories and point clouds), projects results into masked image space, and completes temporally consistent RGB observations with a dense-control image-to-video model. Separating 3D geometric consistency from 2D visual completion lets 2D policies benefit from 3D augmentation.

9. Phase Matters: Phase Determines Identity in Neural Representations

arXiv: 2606.17037 | Field: CV

Inspired by Oppenheim & Lim (1981) — natural images are reconstructible from Fourier phase alone — the study swaps phase and magnitude across images in PRISM2D, GFNet, and ViT-B/16. Predictions follow the phase donor; removing image-specific magnitude information barely changes accuracy. ResNet-50 appears to break the pattern, but analysis reveals strong latent sign coding before ReLU. Different architectures share phase/sign identity encoding, exposed differently by rectification and readout geometry — explaining the texture–shape gap between CNNs and attention models.

10. Your Privacy My Cloak: Backdoor Attacks on DP Federated Learning

arXiv: 2606.17035 | Field: ML/Security

Counterintuitive finding: differential privacy (DP), expected to harden federated learning against backdoors, instead masks statistical signatures of malicious updates, disabling existing defenses. The proposed RING attack exploits DP noise to hide malicious contributions and reconstruct strong backdoor signals at aggregation: 90.3% average attack success rate across 4 datasets and 6 SOTA defenses — a 26.08× improvement over baselines. The DP–security trade-off is tighter than assumed.

11. KVEraser: Learned KV-Cache Manipulation for Efficient Local Context Erasure

arXiv: 2606.17034 | Field: NLP

Long-context LLMs need post-hoc erasure of outdated facts, erroneous tool observations, and harmful injections — but local edits have global consequences. KVEraser replaces KV states of the erased span with a learned steering state, fully reusing the rest of the cache. Two-stage training (general span-neighbor pretraining + task-specific fine-tuning) achieves near-full-recompute performance on 1K–32K contexts with only 24% latency overhead (vs. 17.6× for full recompute).

12. Qwen-RobotWorld: A Unified Embodied World Model

arXiv: 2606.17030 | Field: CV

A 60-layer dual-stream MMDiT coupling frozen Qwen2.5-VL semantics with video VAE latents. Through a natural-language unified action interface, it predicts future visual trajectories for robot manipulation, autonomous driving, indoor navigation, and human-to-robot transfer. Three designs: Double-Stream MMDiT, an 8.6M video–text corpus (200M+ frames), and a generalist-to-expert progressive curriculum. Ranks 1st on EWMBench and DreamGen Bench; surpasses all open-source models on WorldModelBench and PBench.

13. DeepRubric: Evidence-Tree Rubric Supervision for Deep Research Agents

arXiv: 2606.17029 | Field: NLP

Instead of having LLMs generate rubrics from queries (often incomplete), DeepRubric first determines what an evidence-backed report should be evaluated on, then synthesizes aligned query–rubric pairs. It builds evidence trees from seed topics (recursively expanding evidence-backed sub-questions), with leaf nodes as atomic verifiable evaluation targets. With 9K supervised samples + GRPO, it matches SOTA on 3 benchmarks using 13× fewer RL GPU-hours.

14. HAMON: Long-Horizon Forecasting with Passive Optics

arXiv: 2606.17028 | Field: ML/Optics

If the core forecasting operator is low-complexity and approximately linear, is learned digital temporal mixing needed? HAMON uses a passive diffractive-optical forecasting core: historical values encode into an optical aperture, future positions stay dark, and cascaded trainable phase masks + free-space diffraction shape predictions directly in the output field. A single passive optical propagation at inference, zero trained parameters. Beats the strongest digital baselines on all ETTm2 horizons and improves MSE by up to 14% on ETTh2 — defining a concrete target for optical hardware.

15. MeshLoom: Feed-Forward Non-Rigid Mesh Sequence Registration

arXiv: 2606.17027 | Field: CV

Prior methods require expensive per-instance optimization, narrow categories, and pairwise inputs. MeshLoom is a feed-forward network registering multiple meshes in seconds: topology-aware point representations encode reference mesh topology into vertex features; a multimodal encoder fuses shape latents and image features into a compact global motion embedding; a lightweight decoder queries per-vertex deformation. The global embedding–query paradigm naturally supports intermediate-time interpolation for motion interpolation and mesh deformation. Code: https://meshloom.github.io/

16. ExpRL: Exploratory RL for LLM Mid-Training

arXiv: 2606.17024 | Field: ML

Sparse-reward RL success depends heavily on base-model coverage, and mid-training requires manually specifying what to learn. ExpRL uses reference answers as reward scaffolding (hidden from the policy, used only to build question-specific rubrics); an LLM judge compares sampled reasoning traces against reference answers, assigning outcome-level or process-level dense rewards. This upweights partial progress, useful intermediate reductions, and productive reasoning that sparse rewards miss. Outperforms SFT, sparse GRPO, and self-distillation on math reasoning and provides better initialization for subsequent sparse RL.

17. Learning the Geometry of Data: A Mathematical Survey of Shape Space Analysis

arXiv: 2606.17022 | Field: Math/ML

A survey organized around a unified pipeline: shape representation and parameterization → rigorous construction of robust geodesic metrics → statistical analysis on shape spaces → geometry-aware learning. Applications span biology (subcellular morphology), anthropology (primate dental evolution), and computer vision. Core challenge: complex, nonlinear, often unaligned geometric variability; opportunity: increasingly large and diverse geometric datasets.

18. FusionRS: A Large-Scale RGB–Infrared Remote Sensing Dataset

arXiv: 2606.17020 | Field: CV

The first large-scale RGB–infrared–text remote sensing dataset. RGB images are converted to infrared-styled images to form aligned RGB–IR pairs, each with a standard scene description and an IR-aware caption explicitly describing infrared-specific visual properties. Key finding: IR-aware captions are essential for strengthening infrared–language alignment; modality-specific text supervision matters for scalable RGB–IR vision-language representation learning.

19. TokenPilot: Cache-Efficient Context Management for LLM Agents

arXiv: 2606.17016 | Field: NLP

Dual-granularity context management: globally, Ingestion-Aware Compaction stabilizes prompt prefixes and removes open-world environment noise; locally, Lifecycle-Aware Eviction monitors residual utility of context segments and batch-evicts only when task relevance expires. Key insight: existing text pruning and dynamic memory eviction change layout, causing prefix mismatches and cache invalidation. TokenPilot cuts costs by 61%–87% on PinchBench and Claw-Eval while maintaining performance. Integrated into LightMem2: https://github.com/zjunlp/LightMem2

20. Filtered Conformal Ellipsoids: Joint Prediction Sets for Graph-Native Time Series

arXiv: 2606.17014 | Field: ML

A frozen state-space filter emits one-step predictive means and covariances; split-conformal calibration is applied to Mahalanobis scores. The filter selects ellipsoid shape; conformal calibration selects the scalar radius. The core difficulty: filtered scores are dependent, and learned recursive filters are not guaranteed to contract on raw hidden states. Analysis in the observable predictive-law quotient identifies hidden states producing identical future Gaussian law sequences. A GCN-GRU filter yields sharper ellipsoids than static-covariance and unfiltered baselines on METRLA-20 and PEMSBAY-50.

---

*Collected: 2026-06-17 | Source: arXiv cs.AI, cs.LG, cs.CL, cs.CV*