Can MechInterp explain LLM agent failures? A probe on prompt tokens predicts web action hallucinations before generation (0.80 AUC), with steering correcting 85% of errors. But formatting sensitivity reveals key limits in agent circuit discovery. https://t.co/QByikqgp34
Rethinking agentic RL scaffolds! 🤖 ModHarness-RL turns fixed agent harnesses into learnable, evolving programs that adapt as the model trains—optimizing context compression & steering guidance to boost performance across benchmarks. https://t.co/gEWDmMo5tu
IBMARL brings Imitation Bootstrapped RL to multi-agent teams! By combining round-robin human demos with Q-filtered imitation penalties, agents learn cooperative tasks up to 4.2x faster and surpass human performance. https://t.co/3lYoHNNyXT
Simulating quantum dynamics at scale just got a major boost! By partitioning Pauli strings with F2-linear hash functions, a new approach eliminates global sync bottlenecks, scaling simulations past 10^11 Pauli strings. https://t.co/EU7SuydZOU
Researchers deployed a 4-city entanglement-based QKD network in Slovakia! Features trusted-node-free passive wavelength routing and real-time key delivery benchmarked against finite-key security bounds over 100+ km of fiber. https://t.co/OARX3FCwhb
Quantum memory can hide in plain sight! Non-Markovian dynamics can mimic classical memory in 2-time tests, but reveal genuine quantum memory when evaluating 3+ times jointly. Now proven on superconducting qubits: https://t.co/FPRhDpSgDI
Word vector arithmetic (king - man + woman = queen), but for neural network weights! 🤯
New research shows meta-learning structures parameter space so symbolic addition of weight vectors creates working functions without training:
https://t.co/xMUs5ltSrH
CoReMem introduces cost-aware memory for coding agents! By valuing prospective memory usefulness against read/write costs, it boosts a 27B model's resolve rate on ChainSWE from 21% to 25.7% with 17% fewer steps: https://t.co/eXtif6U2KY
Diffusion LLMs often repeat errors when remasking text. Enter DiVR: a verifier that predicts if a partial trajectory will succeed, deciding when to regenerate blocks. It boosts LLaDA-8B accuracy from 35.3% to 43.4% at ~half the cost of search base https://t.co/FDmF8NhZa5
Can we achieve optimal quantum metrology without scaling measurements with dimension d? A new study proves a constant number of fixed bases (e.g., 5 Haar bases) suffices for near-optimal multiparameter pure-state sensitivity! https://t.co/fWPhmHdxC9
Can a fixed measurement estimate any unknown pure quantum state optimally? New research establishes sharp Fisher information bounds using projective designs, enabling state estimation with optimal O(d/N) mean infidelity! https://t.co/xmcYMhQ1Vr
Can superconducting circuits simulate early-universe particle creation? A new study shows dc-SQUID transmission lines mirror scalar fields in expanding cosmologies, using anomalous moments to witness vacuum squeezing despite dissipation. https://t.co/VLbR8oMnE2
Can AI agents engineer ML pipelines with limited labeled data? AgentSSL frames semi-supervised learning as program search, outperforming top human baselines by up to +21.8%. Program selection remains a key bottleneck. https://t.co/37zfwHEubi
Evaluating LLM architecture ideas without burning compute is hard. ArchSpace provides an open platform leveraging fitted scaling laws to predict large-scale model performance from small experiments, complete with Qwen3.5 and OLMo3 baselines. https://t.co/Vb2jHjv6aT
Do MLLMs need a standalone vision encoder? A controlled study unbundles visual encoding: pretraining vision encoders separately is unnecessary, but dedicated vision capacity is key. In-fusion vision modules in early backbone layers lower TTFT! https://t.co/O1nhxnzQAI
How do wavefunctions behave in many-body localized phases? A new microscopic theory reveals why typical Fock-space correlations decay exponentially while average ones show subexponential decay, explaining quantum multifractality! https://t.co/65ibWKGB11
How do measurements shape quantum dynamics? New work shows trajectory and projected state ensembles undergo ensemble phase transitions at the purification critical point, revealing how ergodic steering shifts in hybrid circuits! https://t.co/sltzsOTxRY
Can anti-unitary driving stabilize Discrete Time Crystals? New research shows anti-unitary Floquet driving locks subharmonic response via dynamical freezing, creating robust time-crystalline order without decay! https://t.co/BOKzPVgBC7
Predictive coding can now match Backprop on ImageNet! 🧠⚡ Natural PC uses local lateral circuits for natural gradient learning without matrix inversions—matching AdamW/BP performance at 10x lower compute cost and 100x fewer samples.
https://t.co/buMAQToXDN
Mobile GUI agents are vulnerable to prompt injections hidden in ordinary app UI. MobileProbe introduces a trajectory-aware red-teaming framework that adaptively discovers and localizes security vulnerabilities across long-horizon mobile tasks. https://t.co/D3eXeurKWx