Weekly Recap

A specific looped Transformer architecture and input injection strategy unlock significant gains in reasoning and context utilization, outperforming larger models with drastically fewer parameters.

Human-crafted conference posters are now directly usable as verified supervision for generating new posters, bypassing the limitations of layout-box-only datasets.

LLMs can now "think with visual latents" by generating masks directly from their internal representations, dramatically improving referring comprehension and segmentation.

Agents can now learn to design and critique 3D scenes like human designers, achieving state-of-the-art quality through complementary learning.

Packed trajectories, not sequential rollouts, unlock high-fidelity multi-turn image editing by preventing visual drift.

Human-crafted conference posters are now directly usable as verified supervision for generating new posters, bypassing the limitations of layout-box-only datasets.

LLMs can now "think with visual latents" by generating masks directly from their internal representations, dramatically improving referring comprehension and segmentation.

Agents can now learn to design and critique 3D scenes like human designers, achieving state-of-the-art quality through complementary learning.

Packed trajectories, not sequential rollouts, unlock high-fidelity multi-turn image editing by preventing visual drift.

Open-weight diffusion models can now render complex, bilingual text and control intricate layouts with near-perfect accuracy, closing a critical gap with proprietary systems.

VLM agents can achieve state-of-the-art robotic manipulation by treating failures as coding bugs, with contact precision, not reasoning, proving the primary bottleneck.

Simulating professional organizations with persistent context and delegation unlocks emergent agent capabilities, outperforming single- and multi-agent baselines on complex professional tasks.

Video world models fail at executing instructed changes, not remembering them, a bottleneck MIST and M3-Agent finally expose and overcome.

Agentic AI inference can be accelerated by 2-6x with history-free speculative decoding that eliminates KV cache overhead and boosts acceptance rates by up to 70%.

A single scene can be represented by ~100K plane primitives (5MB), enabling video world models to maintain geometric consistency across hundreds of frames and large viewpoint changes.

A specific looped Transformer architecture and input injection strategy unlock significant gains in reasoning and context utilization, outperforming larger models with drastically fewer parameters.

Explicitly grounding the goal object in view unlocks near-expert performance in object-goal navigation, revealing perception rather than policy as the key limitation.

A specialized CAD harness drastically improves LLM agent performance, cutting output tokens by up to 88% and boosting F1 scores by 13% through geometry-grounded RL.

Explicitly aligning video captions with atomic audiovisual events unlocks substantial gains in fine-grained temporal and attributional accuracy, outperforming unstructured generation by over 9% on event-centric metrics.

Remote KV caches can cripple LLM prefill latency, but sharding the KV store across parallel ranks and enabling asynchronous loading slashes p99 time from 83s to under 1s.

A novel event-centric evaluation framework, AVC-Reward, finally provides a nuanced assessment of audio-video generation beyond holistic metrics, revealing fine-grained cross-modal failures.

Targeted data augmentation and joint RL/OPD training unlock significantly stronger long-horizon multimodal search capabilities for agents.

Continuous, decomposed visual verification unlocks robust rule-governed generative world modeling, outperforming binary critics and enabling structured reasoning agents.

Synthesizing data environments from explicit world specifications, not just files, unlocks scalable training for data agents and reveals a fundamental computation-insight gap in frontier models.

A strong LLM can review grounded drafts from a weak model, retaining most task performance while halving latency and memory traffic.

Distilling explicit, reward-guided skills into VLMs without the student model ever seeing the skills themselves unlocks superior multimodal performance.

Generalization after fine-tuning is dictated by the linear structure of knowledge in pretrained models, not just factual memorization.

Generative time series models can now produce multiple plausible future trajectories as fast as quantile models query a single point, unlocking accurate dependency modeling for critical applications.

Frontier generalist computer-use agents, previously closed-source, are now open for systematic study and community development, with the first release achieving significant performance gains.

Spatial memory via frustum slab routing slashes video world model inference time by over 4x without sacrificing generation quality.

Current VLA models fail at zero-shot robotic personalization, highlighting a critical gap in embodied AI's ability to adapt to user preferences.

Explicitly supervising physical state changes, not just language or latent representations, unlocks more robust VLA model generalization and real-world robotic control.

Restricting LLM fine-tuning to the bottom singular subspace nearly eliminates catastrophic forgetting, even when protecting 99% of weights, while maintaining strong target-task performance.

Gating policy distillation with directional agreement between early and late RL shifts unlocks more effective knowledge transfer, boosting LLM performance on reasoning tasks.

Safety guard evaluations can be misleadingly inflated by superficial cues, with simple edits to tool names or command structure significantly altering measured performance and agent behavior.

Audio embeddings capture physical damping properties far better than single visual frames, and this gap is partially bridgeable through audio-to-vision distillation.

Assigning reward credit to specific tokens, rather than uniformly, dramatically improves multi-reward RL for LLMs.

Diffusion models can learn new objectives by stochastically routing attention within their existing architecture, achieving 4x faster RL training without sacrificing generation quality.

Robust RL policies for adaptive cancer therapy achieve superior performance and generalization by treating patients as hidden contexts, outperforming existing methods even with noisy biomarkers and shifting disease dynamics.

Uncovering crucial supporting lemmas for theorem proving by traversing multi-hop dependencies in a formal library's declaration graph.

Dual-view time-series representations unlock significant gains in temporal reasoning by preserving both absolute measurements and invariant temporal structure, outperforming dedicated TSLMs.

Regularized importance sampling in LLM RL breaks reward shift invariance, but a covariance-based estimand recovers this property, with distinct centering strategies impacting bias and variance.

Agent guards designed for English requests exhibit significant safety failures when faced with multilingual inputs, with smaller models and specialized guards showing the most pronounced language-dependent biases.

Simultaneously evaluate $K$ policies in linear contextual bandits with near-optimal sample complexity by sharing information across evaluations via a unified ridge regression estimator.

A foundation model now surpasses the Cox model on survival analysis, even for small clinical cohorts, by predicting survival curves in-context.

Downstream QA metrics hide critical memory failures: agents correctly classify memory operations less than 20% of the time, a flaw invisible to current benchmarks.

Fine-tuning wind-field super-resolution models with persistent homology and merge tree losses dramatically improves topological structure reconstruction, outperforming standard methods and even the pretrained model.

Hyperscalers can achieve 1.5x training throughput gains on AMD GPUs by moving beyond direct CUDA porting to hardware-model co-design and agentic optimization workflows.

A single model query can now estimate the full acceptance probability for parallel masked diffusion, slashing rejection sampling costs.

Lossless compression of sparse ReLU activations slashes LLM activation VRAM by up to 31% with near-zero overhead.

Exponential Hopfield networks learn exponentially more archetypes than pairwise models by leveraging softmax attention, but risk memorizing noisy examples over true archetypes.

A 2.4B VLA model achieves state-of-the-art performance by integrating a novel register-based memory into its action expert, drastically reducing oscillations and improving real-world robotic task success.

Wireless signals, unhindered by darkness or walls, can reconstruct building layouts by leveraging a diffusion model over physics-grounded representations of channel measurements.

AI weather models now forecast tropical cyclone tracks with superior probabilistic accuracy to leading physics-based systems, even while struggling with intensity.

A simple byte reordering strategy slashes multilingual boundary prediction errors by over 70% in scripts like Korean and Chinese, fundamentally altering how byte-level models perceive codepoint transitions.

Transformers trained with ABNAT learn to mimic clean attention maps, achieving up to 85.77% less attention deviation and 11.28% higher accuracy on noisy analog hardware.

Experimental data can actively harm decision-making when external, unchangeable agents act on it, necessitating a new design paradigm beyond standard Bayesian methods.

Standard MIL suppresses secondary findings in coexisting lesions; this novel framework recovers them by modeling the dominant-label annotation process.

Frozen LLMs reject sender KV caches even when geometrically similar, failing due to competitive attention, not fidelity.

Deep learning compilers are silently introducing bugs through state retention, with 35 stateful issues found in just four frameworks.

Offline-trained neural networks are mathematically proven to diverge in closed-loop control, but bounded online adaptation offers a practical solution.

A simple, fixed-budget best-arm identification algorithm can be made nearly as adaptive as state-of-the-art methods by optimizing its phase schedule, achieving $11/12$ of the optimal normalized rate for Gaussian arms.

LLMs' financial forecasts are more swayed by commentary tone and momentum than by the actual earnings surprise itself.

A standard CNN technosignature detector misses 91% of genuine signals when they deviate slightly from the training morphology, and is fooled by terrestrial interference as often as it finds real signals.

A handful of chemically diverse unlabeled molecules can boost QSAR model performance more than massive, randomly sampled datasets.

Neural networks preferentially learn low-frequency sinusoidal patterns due to the inherent graded structure of their gradient descent dynamics, explained by an orthogonal polynomial framework.

Fine-grained data balancing in heterogeneous pipelines unlocks 2x training speedups for large models by overcoming memory bottlenecks and coarse workload assignments.

Real-world robot demonstrations can now be collected at 66 demos/hour by optimizing TAMP solution quality for behavior cloning.

A novel temporal graph framework achieves over 250x fewer FLOPs and 20x fewer parameters than Transformers/LSTMs for early gastrointestinal cancer risk stratification from sparse EHRs.

Hundreds of billions of midtraining tokens on a curated software engineering corpus unlock substantial gains in agentic capabilities, outperforming standard SFT alone.

A single weight-space encoder now masters 100 transformer architectures, outperforming specialized models and handling unseen, larger networks with less memory.

Agents can now learn reusable knowledge from experience by converting it into ontological axioms, enabling LLMs to dynamically propose and verify hypotheses for complex tasks.

Continual learning, aiming to improve deepfake detector accuracy over time, systematically intensifies unfairness across demographic groups.

Causal direction can be identified by minimizing the Dirichlet energy of one variable treated as a signal on a kernel graph of the other, bypassing regression and independence tests.

Logarithmic sampling regret is achievable in online learning by leveraging squared confidence widths, a fundamental departure from standard regret minimization.

A single INR can now model complex, periodic deformations across entire image sequences, enabling self-regularization and outperforming prior methods.

Active policy selection, not just expert samples, is essential for strategically identifying rewards in multi-agent games, enabling near-optimal downstream decision-making.

Existing benchmarks fail to accurately measure edge unlearning effectiveness, with current methods often slower than retraining and standard accuracy metrics proving misleading.

A two-bit code and shared reconstruction program slash LLM conditional memory storage by 75%, enabling 4x larger tables on GPU.

A 32.5x smaller grid foundation model achieves superior reactive power prediction and significant active power improvements over its larger teacher, enabling real-time edge inference.

Adaptive planning horizons emerge naturally from policy confidence, outperforming fixed discount factors on challenging credit assignment problems.

Non-monotonic and fine-granular quantized activations are now efficiently implementable on custom DNN accelerators, unlocking broader hardware optimization.

A closed-form depth-width law for neural collapse reveals how representation compression scales with the L/N ratio in deep linear networks.

A single diffusion model can forecast sky images and irradiance simultaneously, recovering ~46 W/m² of predictive power lost by cascaded frame-to-irradiance methods.

Concept interventions can actively worsen model predictions, and existing "benign leakage" criteria fail to prevent this, necessitating new training objectives.

Single-qubit quantum circuits, despite universal function approximation, get stuck in training due to "phase-locking," requiring orders of magnitude more optimization steps than multi-qubit equivalents.

Shared experts in MoEs paradoxically limit the very exploration fine-grained segmentation needs, but MirrorMoE unlocks this potential by adaptively pruning the routing space.

A novel PI-DeepONet implementation slashes training time by up to $3.75\times$ and makes 3D physics-informed learning feasible, overcoming critical computational barriers.

Specialized experts in MoE LLMs, not dense model neurons, are the key to interpretable and controllable personalization.

Robots can now ask humans precisely *what* they don't understand, leading to 30.8% better reward recovery with fewer queries.

Autonomous ML engineering agents can achieve frontier-model performance by learning from long experimental trajectories using outcome-guided supervision and windowed reinforcement learning.

A single adversarial token can bypass most safety filters.

Byte-level LLMs can achieve a 25% jump in math accuracy by patching where text is *new*, not just surprising.

Sharing LLM internal states is efficient, but only if you augment the KV cache and replace activations, not swap them.

Deep learning architectures designed for semantic understanding can now solve complex problems in algebraic topology, identifying topologically equivalent finite group actions on Riemann surfaces.

LLM agents can be reliably steered back to task goals with a learned, dual-path intervention mechanism that generalizes across tasks and executors.

LLM agents can learn to proactively build optimized, hierarchical file system workspaces that anticipate future queries, outperforming reactive retrieval by 33 points on BrowseComp-Plus.

Asymmetric temporal dynamics and expanded receptive fields are unlocked in spatio-temporal graph forecasting by a unified spectral operator on cellular sheaves.

Decentralized agent swarms can learn to self-orchestrate their collective behavior and individual capabilities simultaneously, outperforming fixed coordination strategies.

Neural audio watermarking detectors can now reliably detect legacy content even after updates, while simultaneously rejecting unauthorized embedders, by explicitly training for angular separation in the embedding space.

Rational synthesis for multi-agent systems with unknown dynamics and $\mathsf{PLTL}_f$ objectives is achieved by integrating solution concepts like Nash equilibrium with temporal logic satisfaction.

Hundreds of adversarial honesty targets are sufficient to teach LLMs to confess their hidden behaviors and resist pressure to lie, outperforming existing methods without sacrificing utility.

A single adversarial token can bypass most safety filters.

Consolidating specialist LLMs into a single, more accurate generalist is possible with *zero* gradient updates, relying solely on forward passes and input statistics.

Prioritizing upstream causal drivers over downstream effects, Causal Integrated Gradients accurately attributes model predictions along known causal pathways at scale.