Issue 26 · Jun 22–28, 2026

Week 2026-W26

3,210 papers scanned 150 shortlisted 10 picked $10.40 spent

This week's standout is a genuine landmark in computational imaging: the first complete virtual unrolling and scholarly reading of a sealed Herculaneum scroll. Robotics is strong on two fronts—sim-to-real methods that improve real policies without real-world data, and hardware/physics work that put a robot on the table against professional table-tennis players. Neuroscience delivers several assumption-challenging causal results (proprioception vs. reaching, spike synchrony vs. rate codes), and there's a cluster of omni-modal/audio-visual architectures plus theory explaining why RLVR beats SFT.

Tech ✓ read

Complete virtual unwrapping and reading of a rolled Herculaneum papyrus

Giorgio Angelotti, Stephen Parsons, Federica Nicolardi, Youssef Nader et al.

A complete, papyrologically-reviewed virtual reading of an intact sealed scroll—moving from isolated patches to full unrolling, plus directly visible ink in another scroll and a title/author attribution. This is a demonstrated, potentially scalable capability that could unlock an entire ancient library.

Look for Check the coverage/review criteria and how much of the workflow is automated versus bespoke per-scroll, which determines whether it truly scales.

9 min read · arXiv ↗ ·PDF
Robotics ✓ read

Support-Constrained RL Enables Real-World Policy Improvement without Real-World Experience

Raymond Yu, William Huey, Mustafa Mukadam, Anusha Nagabandi et al.

SCORE constrains simulated RL to the support of a real-data generative policy via flow steering, lifting eight real dexterous tasks from 37.8% to 89.9% with no additional real-world experience or distillation. It targets the core sim-to-real exploitation failure mode with a large, multi-task hardware result.

Look for Whether the support constraint limits improvement on tasks the base policy performs poorly, and how sensitive results are to the quality of the real-data prior.

10 min read · arXiv ↗ ·PDF
Robotics ✓ read

Physics Models for Sim-to-Real Transfer in Professional-Level Robot Table Tennis

Christian Conti, Bilan Yang, Alexander Sigrist, Lorenzo Miele et al.

Broad, empirically calibrated flight and contact physics (drag/Magnus, table buckling, learned racket residuals) cut median landing error 59% and enable RL policies that compete with professional players. A rare case of a robot reaching a genuinely hard real-world skill via faithful modeling rather than benchmark gains.

Look for Details on opponent skill, match conditions, and win rates—the abstract asserts professional-level play without quantifying it.

5 min read · arXiv ↗ ·PDF
Neuroscience ✓ read

Cervical spinal cord stimulation disrupts proprioception yet improves voluntary arm reaching

Carranza, E., de Freitas, R., Verma, N., Borda, L. et al.

A clean causal test: cervical spinal cord stimulation disrupts proprioception and postural stabilization yet improves rapid reaching smoothness and accuracy, helping resolve a decades-old debate about whether proprioception is required for goal-directed movement. Directly relevant to motor control and BCI/stimulation systems.

Look for Sample size and effect sizes are absent from the abstract; scrutinize how many participants and how robust the dissociation is.

7 min read · bioRxiv ↗ ·PDF
Neuroscience ✓ read

The Importance of Synchrony in the Neural Control of Movement

Hasegawa, M., Gruszka, B., Finch, M. S., Athreya, V. J. et al.

Millisecond-precise holographic optogenetics shows that motor-cortex output depends strongly on inter-neuron synchrony even when firing rates and cell identities are held fixed—causal evidence favoring a timing code over a rate code. This could change how we think about population coding of movement.

Look for Limited behavioral scope and quantitative detail; watch whether the synchrony dependence generalizes beyond the stimulation paradigm.

6 min read · bioRxiv ↗ ·PDF
AI / ML ✓ read

MJEPA: A Simple and Scalable Joint-Embedding Predictive Architecture for Audio-Visual Learning

Revant Teotia, Adrien Bardes, Michael Rabbat, Sumit Chopra et al.

A single unified encoder trained with one JEPA objective across audio and video, where cross-modal prediction is shown to be necessary for the shared representation to beat unimodal baselines. Strong frozen-evaluation gains with 10x less data make it a compelling alternative to modality-specific contrastive/reconstruction pipelines.

Look for Breadth of downstream evaluation—gains are concentrated on audio benchmarks, so check whether video representations hold up equally.

8 min read · arXiv ↗ ·PDF
AI / ML ▲ 122 ✓ read

Wan-Streamer v0.1: End-to-end Real-time Interactive Foundation Models

Lianghua Huang, Zhi-Fan Wu, Wei Wang, Yupeng Shi et al.

A single Transformer natively streaming interleaved language, audio, and video for full-duplex interaction, replacing cascaded ASR/LLM/TTS/avatar pipelines, with ~200 ms model latency. A genuinely ambitious any-to-any real-time architecture in a direction the reader tracks closely.

Look for Latency is well specified but capability/quality evidence is thin; be skeptical about response quality and how it compares to cascaded systems.

10 min read · arXiv ↗ ·PDF
AI / ML ✓ read

Provable Benefits of RLVR over SFT for Reasoning Models: Learning to Backtrack Efficiently

Stanley Wei, Juno Kim

A theoretical account of why RLVR outperforms SFT: modeling CoT as graph pathfinding, SFT on shortest paths cannot learn to backtrack from dead ends, while outcome-reward RL can, yielding an exponential inference-compute separation. A crisp explanation for a live empirical debate.

Look for How restrictive the graph/pathfinding assumptions are and whether the exponential separation reflects realistic reasoning tasks.

8 min read · arXiv ↗ ·PDF
Neuroscience ✓ read

A number simplex in the human medial temporal lobe

Zhu, H., Chericoni, A., Ismail, T., Mickiewicz, E. et al.

Human medial-temporal-lobe recordings show number representations form high-dimensional simplex manifolds rather than a 1D mental number line, with linearly transferable codes across formats and a direct parallel to LLM representations plus attention-like arithmetic. A rare brain–AI computational connection.

Look for Robustness across subjects and tasks, and whether the LLM analogy is more than superficial geometric resemblance.

9 min read · bioRxiv ↗ ·PDF
AI / ML ✓ read

Nemotron-Labs-TwoTower: Diffusion Language Modeling with Pretrained Autoregressive Context

Fitsum Reda, John Kamalu, Roger Waleffe, Mostofa Patwary et al.

TwoTower decouples context modeling (frozen AR tower) from diffusion denoising, retaining 98.7% of AR quality at 2.42x higher throughput on an open 30B hybrid MoE, with weights released. A clean architectural idea plus an open large-scale artifact make it more than a routine diffusion-LM tweak.

Look for Missing benchmark and hardware detail—verify the throughput claim reflects realistic serving rather than a favorable setting.

9 min read · arXiv ↗ ·PDF

Also notable

Hardware Design for Table Tennis Robot Capable of Beating Professional Players Robotics
Companion hardware paper to C3: an 8-DoF robot ('Ace') with 22 m/s racket velocity that reportedly beat multiple professional table-tennis players.
TurboMPC: Fast, Scalable, and Differentiable Model Predictive Control on the GPU Robotics
TurboMPC, a GPU-native differentiable MPC solver with up to 58x speedups and 8000+ knot-point horizons, open-sourced and validated on a full-scale racing car—likely to be widely adopted.
Bridging Performance and Generalization in Reinforcement Learning for Agile Flight Robotics
Drone-racing RL policy achieving 7.4x better zero-shot generalization to unseen tracks including vision-only end-to-end control, where prior fast policies crash.
Reported Confidence in LLMs Tracks Commitment More Than Correctness AI / ML
Neuroscience-inspired analysis showing LLM verbal confidence tracks commitment/abstention while token log-probs track correctness—a useful dissociation for uncertainty estimation.
Evidence for feature-specific error correction in LLMs AI / ML
Direct empirical evidence for feature-specific error correction in LLM activations, testing a concrete prediction of computation-in-superposition theory across six models.
Environmental color statistics shape the anisotropic geometry of human color discrimination Neuroscience
Evidence that human color-discrimination anisotropy reflects environmental color statistics, supported by psychophysics, MEG, and networks that reverse under hue-inverted training.
Distinct neural geometries for target position and velocity in the primate superior colliculus Neuroscience
Superior colliculus population geometry encodes target velocity as a V-shaped manifold largely dissociable from position, challenging the SC-as-location view.
Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement AI / ML
Flow Reasoning Models use denoising fixed-point stability as a self-verifier, reaching ~99% Sudoku with large search savings—interesting though confined to synthetic puzzles.
Simplified Sparse Attention via Gist Tokens AI / ML
Simplified Sparse Attention via gist tokens needs no architectural changes and reportedly beats full attention on RAG after continued pretraining, with up to 32x compression.
Superhuman AI for Generals.io Using Self-Play Reinforcement Learning AI / ML
Superhuman self-play agent for Generals.io enabled mainly by a JAX simulator ~10,000x faster than the prior one—a striking data-throughput story in imperfect-information RTS.
RAEM: random-access electron microscopy for revisitable 3D imaging Neuroscience
RAEM turns serial-section EM into query-driven multi-resolution imaging, enabling targeted revisiting in petabyte-scale connectomics volumes.
Hallucination in World Models is Predictable and Preventable AI / ML
Argues world-model hallucination is a data-coverage phenomenon that is both predictable and preventable, with adaptation from as few as 50 real trajectories.

Projects

World models dominated this week: hallucination detection, physics-reinforced simulators, and language-based environment models all shipped with real weights and data. There's also a genuinely new Transformer information-flow architecture and a clever voice-agent design for hiding cloud reasoning latency.

GitHub Robotics ★ 94 ✓ read

nicklashansen/mmbench2

Official code repository for the paper "Hallucination in World Models is Predictable and Preventable".

Hallucination is the central practical failure mode of learned world models, and this work claims it is both predictable and preventable, backed by an unusually complete release: 350M-param checkpoints across 210 tasks, 427 hours of video-action data, predictors, mitigation methods, and an interactive demo. This directly bears on whether world-model-based robot planning can be trusted, a live debate.

Look for Check whether the hallucination predictors generalize beyond the 350M scale and the ten benchmark domains, and whether the mitigation holds in closed-loop control rather than just open-loop rollouts.

3 min read ·GitHub ↗ ·Python·MIT
GitHub Speech / Video ★ 19 ✓ read

vysri/conversational-infill

The ConvFill Repository provides training and inference code for the Conversational Infill task proposed and implemented in Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents.

Framing the latency-accuracy tradeoff in voice agents as a trainable 'conversational infill' task — a small local Talker speaks immediately while a cloud Reasoner streams knowledge into the ongoing response — is a genuinely new architecture for real-time voice systems. The release is complete: 290k-example dataset, seven Talker models, training code, and a live demo.

Look for Test how gracefully the Talker handles cases where the Reasoner's answer contradicts what it already said aloud, and whether infill quality holds outside the training distribution.

3 min read ·GitHub ↗ ·Python·MIT
GitHub AI / ML ★ 978 ✓ read

QwenLM/Qwen-AgentWorld

Qwen-AgentWorld: Language World Models for General Agents

Training environment simulation as a native objective from continued pretraining — so one language model can simulate MCP tools, terminals, Android, web, and OS environments — is a substantive new direction for agent training, not an agent wrapper. Released MoE weights, trajectories, and a seven-domain benchmark make it testable, and it enables controllable and OOD simulation for RL without real environments.

Look for Frontier-level agent performance claims are self-reported; verify simulation fidelity on out-of-distribution tool behaviors before using it as a training environment.

3 min read ·GitHub ↗ ·Python·Apache-2.0
GitHub AI / ML ★ 23 ✓ read

lil-lab/sps

Separating persistent state storage from next-token prediction via interleaved persistent and ephemeral streams in the KV cache is a genuinely different information-flow architecture, not an attention tweak. Reported gains hold from 53M to 1.68B params at matched inference cost, with reproducible training and ablation code from a credible lab.

Look for Check whether the scaling trend holds past 1.68B and whether the gains survive on downstream tasks rather than mainly language-modeling loss.

3 min read ·GitHub ↗ ·Python·MIT
GitHub Robotics ★ 122 ✓ read

DAGroup-PKU/PhysisForcing

PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation

A training-time plug-in that reinforces physical plausibility in robotic video world models via interaction-region trajectories and relational constraints on DiT features, with zero inference overhead and released weights for Wan and Cosmos. The jump from 16% to 24% closed-loop planner success on WorldArena is a meaningful real-task result, not just video-quality metrics.

Look for Training code is not yet released and the closed-loop numbers are self-reported; verify the improvement replicates on planners other than the one evaluated.

3 min read ·GitHub ↗ ·Python·MIT
GitHub Tooling ★ 2,114 ✓ read

0xShug0/audio.cpp

An all-in-one, pure C++ inference engine for audio models, powered by ggml. Supports TTS, STT, VAD, voice conversion, music generation, and more, with highly optimized performance. No Python dependency.

A pure C++ ggml-based inference engine unifying TTS, STT, VAD, voice conversion, and music generation with GGUF support and multiple GPU backends fills the same niche llama.cpp filled for LLMs — and its 2k stars in a week suggest it will be widely adopted. For anyone deploying local audio models, this could become default infrastructure.

Look for Check which specific models are actually supported with verified output parity versus merely listed, and how quickly new architectures get ported.

3 min read ·GitHub ↗ ·C++

Also notable

Yoonkyo/mu0 GitHub ★ 39
μ0's embodiment-agnostic 3D interaction-trace world model, pretrainable from video alone, is a substantive alternative to pixel-space world models — worth a look alongside P3 and P4.
InternScience/Agents-A1 GitHub ★ 550
A 35B-A3B MoE arguing for scaling trajectory length (~45k-token trajectories) rather than parameters for long-horizon agents; strong claims that need independent verification.
aletcher/inverting-bellman GitHub ★ 37
P-learning inverts the Bellman equation to extract a usable world model from Q-values — a conceptually elegant result, though evidence is limited to toy environments.
egosteer/egosteer GitHub ★ 92
World-model-enhanced VLA with egocentric-video pretraining and released 3B dexterous manipulation models, but quantitative evidence in the README is thin.
Open-Galapagos/evolution-fine-tuning GitHub ★ 28
Evolution Fine-Tuning trains small models on evolutionary-search trajectories across 371 tasks so the model itself becomes the mutation operator — an interesting move of discovery from scaffold into weights.
ByteDance-Seed/Modular-TTT GitHub ★ 44
ByteDance's Modular-TTT reframes test-time-training memories as composable DAGs of primitive operations, a clean abstraction for studying TTT designs even without headline results.
Dewaldnel11/LightLX GitHub ★ 19
LightLX streams weights layer-by-layer from disk on Apple Silicon, running a 753B model on a Mac — very slow but a real out-of-core inference capability worth knowing exists.
Dicklesworthstone/franken_ocr GitHub ★ 312
A pure-Rust, CPU-only, framework-free port of a 3B MoE OCR VLM with custom int8 kernels — an impressive deployment feat for offline document processing.

The shortlist: top candidates that survived triage · Archive