Daily Paper Digest: 20 New AI/ML Papers from arXiv (June 15, 2026)

A curated daily digest of 20 new AI/ML papers from arXiv (June 15, 2026), sourced from cs.AI, cs.LG, cs.CL, and cs.CV.

Key Papers

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

arXiv: 2606.17056 | NLP | Nick Jiang et al.

Language models contain an internal "value axis" encoding the expected success of the current policy at achieving its goal. 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-seeking behavior. Notably, Qwen3 assigns low value to politically sensitive queries—meaning the model's internal "confidence" is readable and steerable.

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

arXiv: 2606.17053 | NLP/CV | Peiyang Xu et al.

Instead of directly supervising final answers, the model chooses which of two highly similar contexts supports a 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 contrastive data.

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

arXiv: 2606.17049 | CV | Yi-Ruei Liu et al.

A unified framework combining physics-based rendering (controllable but artifact-prone) with generative models (realistic but inconsistent). The physical model provides controllable rendering; the generative model handles denoising and repair. 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 | ML/CV | Abbas Mammadov et al.

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 can be trained from scratch or fine-tuned. Outperforms prior methods on 5 inverse problems across FFHQ and ImageNet, using ~10x fewer denoiser evaluations than gradient-based approaches.

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

arXiv: 2606.17046 | Robotics/CV | Jisang Han et al.

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

6. HABC: Hierarchical Advantage Weighting from Sparse Robot Feedback

arXiv: 2606.17043 | Robotics/RL | Tongyan Fang et al.

In online RL with binary per-episode success/failure, compressing outcomes into a single scalar reward conflates feasibility and efficiency. HABC trains two critic heads (feasibility and efficiency) merged with state-adaptive gating. On three bimanual contact tasks, SFT baselines at 36%/44%/12% success rise to 92%/88%/38%.

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

arXiv: 2606.17041 | NLP/IR | Anzhe Xie et al.

442 meta-analysis articles from Nature Portfolio, each with research questions, PI/ECO criteria, a 140K-PubMed retrieval corpus, verified positive studies, and hard negatives. Key bottleneck: retrieval recall ceiling is 90.9%, yet no system recovers more than 52.7% of ground-truth literature. LLMs cannot reliably distinguish studies meeting PI/ECO criteria from topically similar but non-qualifying ones.

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

arXiv: 2606.17040 | Robotics/CV | Xiuwei Xu et al.

Performs lightweight 3D augmentation on incomplete point clouds (editing trajectories and points), projects into masked image space, then 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 Network Representations

arXiv: 2606.17037 | CV | Alper Yıldırım

Inspired by Oppenheim & Lim (1981)—natural images are reconstructable from Fourier phase alone—in experiments on PRISM2D, GFNet, and ViT-B/16, grafting image A's phase onto image B's magnitude makes predictions follow the phase donor. Removing all image-specific magnitude information barely changes accuracy. ResNet-50 seems to break the pattern, but strong latent sign encoding exists before ReLU. Conclusion: 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 Differentially Private Federated Learning

arXiv: 2606.17035 | ML/Security | Xiaolin Li et al.

Counter-intuitive finding: differential privacy (DP), expected to harden federated learning against backdoors, actually masks statistical signatures of malicious updates, disabling existing defenses. The proposed RING attack exploits DP 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 26.08x improvement over baselines. Lesson: the DP–security trade-off is tighter than assumed.

11. KVEraser: Learned Local Context Erasure in KV Caches

arXiv: 2606.17034 | NLP | Mufei Li et al.

Long-context LLMs need to erase outdated facts, wrong tool observations, or harmful injections—but local edits have global consequences since all subsequent tokens must be recomputed. KVEraser replaces erased intervals' KV states with learned guidance states while reusing the rest of the cache. Two-stage training: generic span-neighbor pretraining + task-specific fine-tuning. Matches full recomputation quality on 1K–32K contexts with only 24% added latency (vs 17.6x for full recomputation).

12. Qwen-RobotWorld: A Unified Embodied World Model from Qwen

arXiv: 2606.17030 | CV | Jie Zhang et al.

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

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

arXiv: 2606.17029 | NLP | Minghang Zhu et al.

Existing methods have LLMs generate rubrics from queries, but these are often incomplete. DeepRubric reverses this: first determine what an evidence-backed report should be evaluated on, then synthesize aligned query–rubric pairs from those targets. It builds evidence trees from seed topics (recursively expanding evidence-supported sub-questions), with leaves as atomic verifiable evaluation targets. With 9K supervised samples + GRPO training, it matches SOTA on 3 benchmarks while cutting RL GPU-hours by 13x.

14. HAMON: Long-Horizon Forecasting with Passive Optics

arXiv: 2606.17028 | ML/Optics | Alper Yıldırım

A bold attempt: if the core forecasting operator is low-complexity and near-linear, can learned digital temporal mixing be avoided? HAMON uses passive diffractive optics as the forecasting core—history encoded into an optical aperture with future positions left dark, cascaded trainable phase masks + free-space diffraction shape predictions directly in the output field. Inference is a single passive optical propagation with no trained parameters. It beats the strongest digital baselines on all ETTm2 horizons and improves MSE up to 14% on ETTh2.

15. MeshLoom: Feedforward Non-Rigid Registration of Mesh Sequences

arXiv: 2606.17027 | CV | Jianqi Chen et al.

Existing methods require costly per-instance optimization, narrow categories, and pairwise-only inputs. MeshLoom is a feedforward 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 deformations. 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 | ML | Violet Xiang et al.

Sparse-reward RL success depends heavily on the base model's coverage. 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 with the reference, assigning outcome- or process-level dense rewards. This upweights partial progress, useful intermediate reductions, and productive reasoning that sparse rewards cannot reinforce. Outperforms SFT, sparse GRPO, and self-distillation on math reasoning and provides better initialization for later sparse RL.

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

arXiv: 2606.17022 | Math/ML | Gary P.T. Choi et al.

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 tooth evolution), and computer vision. Core challenge: complex, nonlinear, often unaligned geometric variation; opportunity: ever-larger geometric datasets.

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

arXiv: 2606.17020 | CV | Jiaju Han et al.

The first large-scale RGB–infrared–text remote sensing dataset, built by converting RGB images to infrared-style images to form aligned RGB–IR pairs, each with a standard scene caption and an IR-aware caption describing infrared-specific visual attributes. Key finding: IR-aware captions are critical for strengthening infrared–language alignment; modality-specific text supervision matters for scalable RGB–infrared vision-language learning.

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

arXiv: 2606.17016 | NLP | Buqiang Xu et al.

Dual-granularity context management. Global: Ingestion-Aware Compaction stabilizes the prompt prefix and removes open-world environment noise. Local: Lifecycle-Aware Eviction monitors residual utility of context segments and offloads in batches only when task relevance expires. Existing text pruning and dynamic memory eviction change layout, causing prefix mismatches and cache invalidation. TokenPilot cuts costs 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 | ML | Yannick Limmer

A frozen state-space filter emits one-step predictive means and covariances; split-conformal calibration is applied to Mahalanobis scores. The filter selects the ellipsoid shape, conformal calibration selects the scalar radius. The core difficulty: filter scores are dependent, and learned recursive filters need not contract on raw latent states. The analysis studies contraction in the observable predictive-law quotient, identifying latent states yielding identical future Gaussian law sequences. GCN–GRU filters give sharper ellipses 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*