Open 31B model tops dense rivals on human Arena
Thinking mode and cache tricks let Gemma 4 match far larger open systems on STEM and multimodal tasks
14,782 reviewed in the last 24 hours · 136 in review now 285,411 on the record · 9.8% of the arXiv record read Every review is machine-made and published openly.
sort impact radar most recent
Three alignments unlock long-horizon mobile robots
Latent actions, dual transformers, and dream forcing close the gaps that break world-action models
· “ABot-M0.5: Unified Mobility-and-Manipulation World Action Model”
Student rollouts let multiple RL teachers transfer skills intact
MOPD distills domain teachers onto the student's own data, outperforming mix and cascade baselines while allowing independent teacher develo
· “MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training”
Agents finish only 21% of multi-hour desktop workflows
A 108-task suite shows they lose constraints and miss mid-task state, not basic clicks.
· “OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks”
Reference subtraction purifies self-distillation for long-CoT models
Isolating question-conditioned corrections via reference-only teachers and PMI improves four models on two datasets without losing natural r
· “Purified OPSD: On-Policy Self-Distillation Without Losing How to Think”
Models reach 13.8% on executable state changes in Scratch tests
Benchmark uses verified VM transitions to separate rule-following from copied persistent state.
· “ScratchWorld: Evaluating If World Models Compute Executable Consequences”
HRT conjecture falls to a 12-term linear dependence
It predicted finite time-frequency shifts are independent; this paper constructs a genuine 12-term counterexample.
· “Linear dependence of time-frequency shifts of a Schwartz function”
Robot policies still fail most hard manipulation tasks
Unified sim-and-real tests of 30 models show single-digit success and a large gap to human teleoperation
Second copy never unlocks Werner-state distillation
A sharp projection estimate proves the 2-copy threshold matches the 1-copy threshold in every dimension.
Shared fact graph lets AI teams finish long math proofs
Six research-level case studies show parallel workers can build arguments hundreds of facts deep
· “Danus: Orchestrating Mathematical Reasoning Agents with Fact-Graph Memory”
Estimating a state's spectrum costs almost as much as tomography
At least d^{2−o(1)} copies are needed for constant-precision spectrum, entropy, and rank estimates.
Original coding tasks separate agents public benches cluster
Hand-written functional tests cut judge-verifier disagreement from 32% to 1.4% on matched audits.
· “DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks”
Draft smarter, verify only the survivors: 60–85% faster LLM serving
Semi-autoregressive drafts plus load-aware length scheduling shift the live throughput–speed Pareto frontier.
· “DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation”
Ten-lobed domain breaks Pompeiu and Schiffer conjectures
Certified interval arithmetic proves a noncircular real-analytic domain with a boundary-constant Neumann eigenfunction.
· “A computer-assisted counterexample to the planar Pompeiu and Schiffer conjectures”
Minimal three-step loop beats full skill optimizers
File tools, consensus mining and validation gating speed convergence and let smaller models outperform larger ones
· “SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe”
Josephson phase-lock areas turn into a square parquet
Rescaled near (0,1), the tongues stack as unit squares with constrictions at every vertex as ω→0.
· “On phase-lock area parquet in a special slow-fast limit of model of Josephson junction”
Scaled KŁ property proves Bregman iterates converge
It proves Shannon-entropy Bregman iterates converge to critical points on nonconvex problems.
· “On the Iterate Convergence of Bregman Projected Gradient Method”
k-coloring beats the 2^n barrier for every fixed k
A new bootstrap turns sub-2^n list-coloring algorithms into faster k-coloring algorithms, closing a long-standing gap.
· “k-Coloring is Faster than Computing the Chromatic Number”
BCH codes can beat their designed distance by a cube-root gap
For binary length-8191 codes the gap already reaches 8, and it grows as n^(1/3), so no constant bound can hold.
All-order expansion found for binomial mean deviation
The Gaussian guess is only the leading term; higher-order corrections are explicit Bernoulli polynomials of the fractional displacement of t
Spectrum estimation beats full tomography with o(d²) copies
A relative-error tomography bound lets eigenvalues be learned without learning eigenvectors
· “The Keyl-Werner algorithm is not optimal for spectrum estimation”
Dispersion curves must stay spacelike on a Lorentzian plane
One geometric rule yields sharp bounds on diffusivity, viscosity and hydrodynamics in boosted frames
SEED outperforms outcome-only RL on every tested agentic benchmark
Self-generated hindsight skills give token-level credit, lifting success up to 39 points and cutting data needs.
· “SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning”
Constant K forces isoparametricity for minimal hypersurfaces in S^5
Chern conjecture holds for M^4 in S^5 when both scalar curvature and Gauss-Kronecker curvature are constant, using weighted 3-forms and χ(M)
A hidden 'workspace' in LLMs carries the thoughts they never say
New Jacobian lens reads, swaps, and ablates the representations that drive silent reasoning and strategic deliberation.
· “Verbalizable Representations Form a Global Workspace in Language Models”
A single hidden region explains two-loop collinear-factorisation breaking
A new algorithm isolates the soft+Glauber hidden region that creates the two-loop splitting amplitude's kinematic dependence.
Ghost-free path integral puts every gravity frame on equal footing
Frame-dressed observables make correlators and time evolution relative to the frame you choose.
· “Relational path integral, effective actions and quantum frame covariance in gravity”
Dual routing of token supervision improves on-policy distillation
DOPD assigns each token to teacher or student based on advantage gaps, separating capability transfer from information asymmetry.
35B agent matches trillion-parameter models on long tasks
Long 45K-token trajectories and domain teachers, not more parameters, drive the reported gains on search and science.
· “Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent”
Small-model RL shifts lift larger reasoners
Transfer the before–after policy log-ratio as a dense reward; skip sparse RL on the target.
· “Weak-to-Strong Generalization via Direct On-Policy Distillation”
Three voters still make Kemeny ranking NP-hard
A reduction from MAX CUT answers the 2001 open question and settles the complexity of Kemeny aggregation for every fixed voter count.
Agent learning from environments follows a log-sigmoid law
Across 134 day-long tasks and 38,000 hours, average performance tracks interaction time with R² near 0.998
· “EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments”
Low-temperature spin glass is one interval plus one atom
The Parisi measure for every β>1 has a smooth density and a single endpoint atom—full replica symmetry breaking follows.
· “Full replica symmetry breaking in the Sherrington-Kirkpatrick model”
Bregman proximal iterates converge for nonconvex objectives
A unified framework covers general separable kernels and mirror flows, closing a decades-old iterate convergence gap.
· “A Unified Framework for Iterate Convergence of Bregman Proximal Methods”
Half-Hölder smoothness can break heat-equation regularity
The time derivative escapes square integrability even for arbitrarily small perturbations of the heat equation.
Three questions per player: perfect XOR games stay GHZ-realizable
At four questions, a Klein four-group game wins perfectly under commuting operators yet has no GHZ-equatorial strategy.
· “A Sharp Local-Question Threshold for GHZ-Equatorial Completeness in Four-Player XOR Games”
Complements of paths and cycles need log n + Θ(log log n) cliques
A 1991 growth conjecture is settled, and the clique-cover count for path and cycle complements is log n + Θ(log log n).
· “Kohayakawa's conjecture and clique coverings of complements of paths and cycles”
Four parent inputs organize moiré flat-band Hamiltonians
Valley, orbital, moiré symmetry and band labels map 600+ bilayers onto Hubbard, topological and 1D models
One coherence rule keeps FDR control as hypotheses arrive
One coherence condition turns per-intersection e-processes into stopped-FDR and SupFDR guarantees.
Optimal constant found for scaled q-numerical ranges
The least C in the polynomial bound is max of 1 and 2|q|/(1+sqrt(1-|q|^2)).
Latest arXiv reviews
Newest reviewed papers
- Lions' Maximal Regularity Problem for Divergence-Form Differential Operators: Failure at the $\frac{1}{2}$-H\"older Endpoint
- Breaking the 4-Approximation Barrier in Strategyproof Two-Facility Location
- Information Routing across Batch Boundaries: Memory--Batch Tradeoffs in Lipschitz Bandits
- On phase-lock area parquet in a special slow-fast limit of model of Josephson junction
- Optimal cost of fast boundary controls for the one-dimensional heat equation
- Affine Scaling of Jacobi Zeros: Sharp Orderings Beyond Gautschi's Conjectures
- Enumerating iterated tilted algebras in type $A$
- Convergence of a CutFEM for fluid--structure interaction with a deforming interface
- $q$-Difference Equations for two classes of Pattern Avoiding Permutations
- Stability of MIMO PID With Backward Differences Under Fast Sampling: An Exact Spectral Criterion
- Information for nothing and authority for free
- Exponential growth and decay in the ideal induction equation