Robotics
Test-Time Training pushes robot-policy context to 8,000 timesteps
NVIDIA researchers introduce RoboTTT, a robot foundation-model recipe that folds Test-Time Training into vision-language-action (VLA) policies, compressing long histories into fast weights updated by gradient descent during both training and inference. It scales visuomotor context to 8,000 timesteps — three orders of magnitude beyond current policies — with no growth in inference latency. On real-robot manipulation it reports an 87% overall improvement over a single-step-context baseline, a 62% gain over the same model pretrained with 1K-timestep context, and full completion of a five-minute, ten-stage assembly task that no baseline ever finishes. The authors also observe steady closed-loop gains as pretraining context length scales, proposing context length as a new scaling axis for robot foundation models.
arXiv:2607.15275 (cs.AI) — RoboTTT: Context Scaling for Robot Policies
Quantum Computing
Circuit compression keeps quantum-chemistry accuracy at half the gate cost
The paper exploits a robustness property of Subspace Quantum Diagonalization (SQD) — which needs only bitstrings with sufficient ground-state overlap rather than an accurate variational energy — to strip non-Clifford and variational expressivity from the sampling circuit. Two compression techniques, gradient-based operator pruning and Clifford rounding, are applied to a VQE ansatz on qubit-reduced Hamiltonians. A systematic ablation over 21 molecules finds median SQD error stays within chemical accuracy even at 50% compression on both axes, with simulation speedups up to 33x. Hardware validation on 6 molecules on IBM quantum processors confirms up to 2.8x transpiled-depth reduction with no loss in SQD accuracy.
Materials tuning boosts phonon coherence for quantum acoustics
Published in PRX Quantum, this experimental study examines loss mechanisms in high-coherence multimode mechanical (bulk acoustic) resonators coupled to superconducting circuits. By optimizing the properties of the piezoelectric transducer, the authors improve phonon coherence time and cooperativity and lay out a materials blueprint for quantum acoustics — a route toward using long-lived mechanical modes as quantum memory alongside superconducting processors.
New bounds pin the CZ-cost of preparing graph states
Published in Quantum, this work studies the complexity of preparing graph states using computational-basis measurements and single- and two-qubit Clifford operations, defining a graph state's CZ-complexity as the minimum number of two-qubit Clifford (equivalently CZ) gates needed. It gives a combinatorial characterization: a graph state can be generated from another within CZ-complexity t iff its graph follows from the other by vertex deletions, local complementations, and at most t elementary edge-complementations. Relating this to rank-width, it proves an O(rn) upper bound and an at-least n+r-2 lower bound for connected graphs, shows both are near-optimal, and gives O(n) and O(n log n) preparation algorithms for interval and circle graphs respectively.
Quantum 10, 2165 (2026) — Complexity of graph-state preparation by Clifford circuits
Aqarios becomes Germany's first listed quantum pure-play via SPAC
Quantum-optimization software provider Aqarios GmbH, based in Munich, completed a public listing on Boerse Duesseldorf via a reverse merger with the SPAC Fonterelli SPAC 4 AG. The combined entity has been renamed Aqarios Quantum Technologies AG, which the trade report frames as Germany's first listed quantum pure-play company.
Quantum Computing Report — Aqarios Enters Public Markets via SPAC
Exact quantum many-body scars constructed in a 2D gauge model
Exact quantum many-body scars (QMBS), signatures of ergodicity breaking, are rare above one dimension. Published in Quantum, this paper studies a two-dimensional Z2 gauge model dual to a two-dimensional spin-1/2 XY model on bipartite graphs, identifying exact tower-structured eigenstates of the XY model as exact QMBS and, via the duality transformation, showing the transformed exact scars of the XY (and XXZ) model are exact QMBS of the dual gauge model — a versatile construction with potential to find new scars in other higher-dimensional models.
Quantum 10, 2164 (2026) — Exact Quantum Many-Body Scars in 2D Quantum Gauge Models
AI & ML
A 4B neuro-symbolic pipeline tops the EXACT 2026 reasoning tasks
At the EXACT 2026 competition (open-weight models up to 8B, self-hosted, with a natural-language explanation per answer), team CoTu's system has a 4B backbone write programs rather than answers: Z3 encodings whose entailment grounds deductions for university-regulation queries, and numerical Python for physics, wrapped in a shared self-correction loop with explained-JSON output. It achieved a perfect score on the physics task in both automated selection rounds and the highest final-round technical score of any team (13.44/15), placing 3rd overall once presentation was weighted in. The authors argue grounding answers in a symbolic solver yields verifiable deductions at 4B scale, with residual difficulty lying in premise selection rather than deduction.
Open-source multi-agent system automates brain-science research with audit trails
BrainPilot is an open-source multi-agent system for accelerating brain-science research. A principal-investigator agent coordinates specialist agents grounded in a curated knowledge base of 7,233 indexed items and a skill library of 72 reusable methodology units across seven domains. Every major step is recorded in a 'Graph of Trace' linking subgoals, tool use, evidence, and claims, and an Auditor agent integrates fabrication checking. Evaluated on three tasks from Agents' Last Exam plus the authors' own BrainPilotBench-v0 and end-to-end case studies, BrainPilot with an open-source backbone reports performance comparable to a state-of-the-art agent framework at lower cost.
arXiv:2607.15079 (cs.AI) — BrainPilot: Automating Brain Discovery with Agentic Research
Harness engineering lets a small model beat a bigger one on reliability
The paper argues 'harness engineering' — deliberate deterministic scaffolding around an LLM core — can matter more than model size where reliability and traceability count. It compares an unscaffolded GPT-5 chatbot (ASA) against a multi-module system (ASuS) wrapping the smaller GPT-4o-mini in a LangGraph harness with symbolic-semantic retrieval, schema-validated outputs, LLM-as-judge with bounded retry, human-in-the-loop gates, deterministic weighted risk scoring with LLM narration, and a per-node SQLite audit trail. In a blind ten-rater evaluation, ASuS outscored ASA on every one of six harness-mechanism dimensions (pooled mean 4.08 vs 1.23; 8 of 10 raters reject the null at alpha=0.05 on a paired Wilcoxon test), and an ablation indicates the harness contributions are largely model-invariant.
arXiv:2607.14707 (cs.AI) — Harnessing LLMs for Reliable Academic Supervision: A Comparative Study
Robotics
Streaming inference gives flow-matching robot policies real-time control
Reflex addresses the incompatibility between flow-matching VLA models' iterative denoising and real-time robotics. Exploiting a Timestep-Invariance Property — perception encoders are independent of the denoising loop — it partitions the attention context into static, sliding, and dynamic regions for O(1) incremental cache updates while preserving full-batch-equivalent attention outputs. An adaptive normalization layer (AdaRMSNorm) prevents BFloat16 numerical collapse and an async pipeline decouples visual encoding from action generation. On the LIBERO and Kinetix benchmarks, Reflex reports a 2.58x inference speedup, stable 50Hz streaming, and up to 54% lower reaction latency without performance degradation.
arXiv:2607.14695 (cs.RO) — Reflex: Real-Time VLA Control through Streaming Inference
RL policy harvests strawberries at 82% success on real hardware
This work tackles occlusion and deformable-plant contact dynamics in robotic strawberry harvesting with a shared interaction-aware RL policy generating Cartesian motions across obstacle separation, fruit detachment, and placement phases, coordinated by lightweight heuristics and paired with low-level Cartesian impedance control. Using feasibility-first observation alignment and domain randomization for zero-shot sim-to-real transfer, it reports 89.7% success in simulation and 82.0% in real-world experiments, with average execution time rising from 12.99s to 21.73s as occlusion increases from level 1 to 5.
arXiv:2607.14708 (cs.RO) — Reinforcement Learning for the Full Strawberry Harvesting Process