Nothing here yet
Long-term dependencies pose a fundamental challenge for sequential decision making because Transformers suffer quadratic attention scaling and RNNs lack expressive hidden states. This paper proposes Decision Titan (DTi), which grafts Titans-style Test-Time Training layers onto Decision Transformer to store episodic memories in fast weights updated by gradient descent at both train and test time. The authors introduce X-Maze, a controlled extension of T-Maze, and show that DTi can learn dependencies 20× longer than its context window and extrapolate to unseen episode lengths, while visualizing gate dynamics to explain how momentum propagates long-range information. Although the experiments are limited to synthetic memory tasks, the mechanistic ablations and honest characterization of catastrophic forgetting provide a valuable foundation for future research.
Does RL post-training create new agentic capabilities or just redistribute probability mass toward high-reward behaviors that a base model already possesses? This paper shows that, across 14 base/post-trained checkpoint pairs and three process-centric agentic benchmarks, pre-trained base models equipped with a lightweight harness often surpass their post-trained counterparts in solution coverage (pass@$K$) given enough test-time compute. To diagnose this coverage-accuracy trade-off, the authors propose Sharpening Tax, a single-number metric for the loss in test-time scalability after post-training, and present Posterior-Tempered Group Sampling (PTGS), a Bayesian sampler that mitigates the tax during RL training.
Emergency Department triage assigns acuity scores that can be skewed by non-clinical patient descriptors. This paper asks whether open-source LLMs change their pediatric Emergency Severity Index predictions when only demographic, socioeconomic, or contextual variables are injected into otherwise identical clinical vignettes. The authors audit ten models and find that a small fine-tuned Qwen2.5-7B is the most counterfactually stable, challenging the assumption that larger or medically pretrained models are necessarily fairer. The study matters because locally deployable open-source LLMs are increasingly considered for privacy-preserving clinical decision support, yet their fairness profiles across model families were poorly understood.
Molecular tumor boards must distinguish truly unsupported therapies from evidence-supported recommendations that still require oncologist review because of incomplete clinical information or poor performance status. This paper introduces OpenMTB-Audit, a 500-case synthetic NSCLC benchmark with four safety labels, and exposes pervasive over-refusal: every tested LLM configuration collapses the Partially Supported label, misclassifying valid-but-caveated therapies as either Unsupported or Supported. To address this, the authors propose MTB-AuditAgent, a deterministic seven-module pipeline that reduces over-refusal to 6.7% and achieves 91.2% accuracy. The work matters because label collapse in high-stakes oncology auditing can conceal unsafe endorsements or inappropriately block appropriate treatments.
iADD tackles the alignment-diversity trade-off in reinforcement-learning finetuning of diffusion models such as Stable Diffusion. Its core idea is to update early denoising timesteps rather than late ones, combined with a sparse incremental curriculum and Feynman–Kac particle branching during training. Better alignment without mode collapse matters because existing reward-optimization methods frequently degrade visual quality and diversity.
Quadrupedal locomotion requires balancing command tracking, stability, and energy efficiency, yet conventional RL bakes these trade-offs into a static scalar reward that cannot change after training. PROMO proposes a preference-conditioned multi-objective framework that takes the relative priority vector $w_t \in \Delta^{K-1}$ with $K=3$ as a runtime input to a single policy, while keeping embodiment-specific locomotion priors fixed. On a Unitree Go2, the same policy transferred zero-shot modulates behavior through preference changes alone, cutting specific energy by up to $30.4\%$, position error by $38.7\%$, and peak body-attitude deviation by $59.0\%$.
Catastrophic forgetting remains a critical barrier in multi-phase training of large language models. This paper recasts forgetting as a geometric phenomenon at the level of each weight matrix: standard gradient updates act on all inputs, not just those from the new task, corrupting prior mappings. The authors propose Local Support Learning (LSL), which pairs task-specific adapters with per-matrix Gaussian Mixture Model gates that restrict updates to activations near the current training distribution. Empirically, LSL retains pretrained math, coding, and instruction-following capabilities in models up to 7B parameters while learning new domains efficiently.
Policy-gradient methods suffer from high variance, while pathwise actor updates require accurate action derivatives that standard value-function training does not guarantee. FERPO addresses this by deriving an entropy- and KL-regularized target action distribution from critic values alone, then fitting the policy with forward KL via self-normalized importance sampling---never differentiating the critic with respect to actions. The method targets massively parallel on-policy robot learning, where reliable policy improvement is critical for both sample and wall-clock efficiency.
Where-OPD proposes a new on-policy self-distillation paradigm for multimodal large language models. Instead of giving the teacher zoomed-in image crops, it provides textual spatial hints that identify question-relevant visual elements and their coordinates, using procedurally generated synthetic counting scenes where object locations are known by design. The student learns to match the teacher’s behavior from the raw image and question alone, and the authors claim that training on just 3,000 such synthetic scenes transfers to broad real-world perception gains across 15 benchmarks.
Hierarchical Continuous Diffusion Language Models tackle a core bottleneck in non-autoregressive generation: discrete diffusion decodes parallel tokens as independent marginals, while continuous diffusion lacks valid token anchors until the final step. The authors propose HC-DLM, a hierarchy where a continuous latent trajectory is the sole persistent state and tokens are read out and fed back at every step to scaffold the next latent update. A single variational bound yields the training objective, and experiments span structured reasoning, mathematical planning, and language modeling.
Schrödinger bridge (SB) transport assumes a fixed initial distribution, which breaks at test time when inputs shift. This paper introduces Distributionally Robust Schrödinger Bridge (DRSB), which learns a single controller by minimizing the worst-case sum of control energy $\int_0^1 \frac{1}{2}\|u_t^\theta(X_t)\|^2 dt$ and a terminal KL penalty $\rho \mathrm{KL}(p_1^{u,\mu}\|\nu)$ over an ambiguity set $\mathfrak{M}_\epsilon(\hat{\mu})$ of initial distributions. The authors derive a variational reformulation, connect it to stochastic optimal control and distributionally robust optimization, and develop Wasserstein and Sinkhorn variants with alternating block updates. Experiments on 2D transport and FFHQ image-to-image translation show improved robustness to input perturbations, with a tradeoff in nominal performance.
CARM targets off-policy drift in LLM reinforcement learning by preventing signed token log-ratios from canceling in sequence-level response masks. The core idea is to average absolute log-ratios rather than signed ones, which stops bidirectional policy drift from being hidden by the geometric mean. It also proves a joint bound on the fraction and severity of out-of-band token ratios for any accepted response. The method is evaluated on Qwen3.5 models and shows consistent gains over geometric-mean and token-level baselines.
While 3D Gaussian avatars render photorealistic humans efficiently, their animation pipelines are often throttled by costly per-frame neural inference. This paper introduces GALA, which shows that the animation outputs of diverse pretrained avatar networks can be closely approximated by an identity-shared linear blendshape basis extracted via distillation, coupled with a shallow coefficient predictor. By constructing the basis with block-local PCA under a rendering-aware metric and a strict memory budget, the method reduces CPU animation cost by up to three orders of magnitude and enables 60 fps on mobile devices with only modest quality loss.
Training-data selection for LLMs is usually heuristic or sample-specific, which prevents generalization to unseen data. This paper tackles the harder problem of learning a shared meta-network that scores any example, and shows that plugging such a network into the standard bilevel objective (ScaleBiO) causes unstable optimization and poor transfer. The authors trace the failure to two properties of the ScaleBiO loss—competitive weight suppression and easy-feature dominance—and propose TESS, a method that replaces the bilevel objective with a simple pointwise MSE regression on pseudo-labels derived from loss gaps. TESS achieves strong transfer across datasets, corpus scales, and model sizes while cutting peak memory by over 20%.
Recent reasoning models show impressive capabilities thanks to RLVR post-training, but their chain-of-thought traces increasingly drift into unusual, non-standard language. This paper formally proves that RLVR optimization allows unbounded language drift while supervised fine-tuning imposes a hard finite bound, and shows empirically that such drift arises specifically when RLVR must discover novel behaviors on tasks beyond the base model's existing capabilities. The results suggest a fundamental tension: improving chain-of-thought monitorability by constraining drift necessarily limits the expected reward a model can achieve.
SILSA tackles the cost and topological fragility of high-resolution 3D generation by replacing dense or sparse voxel tokens with compact sliding-window slice latents along the three canonical axes. Each of the 384 fixed tokens summarizes a local depth window, preserving cross-sectional continuity while enabling a single-stage rectified-flow generator. Topology-aware slice supervision and a Volumetric Anchor Lattice then maintain geometric and structural coherence. The paper reports large efficiency gains—over 98% fewer tokens than sparse voxelizers—and solid improvements in PSNR, coverage, and Betti error.
This paper introduces Higher-order Grammar Representation (HGR), a method that lifts molecular graphs into combinatorial complexes and parses them into compact sequences of context-free production rules, making higher-order topology such as ring systems tractable for standard sequence models. To reduce benchmark bias toward simple rings, the authors also construct RingDiv, a ring-enriched benchmark of 1.18 million molecules. HGR is instantiated in generative models (HGR-VAE and HGR-LDF) and a foundation model (HGR-FM), with claims of 100% validity by construction and state-of-the-art distributional alignment and transfer performance.
This paper targets few-shot ransomware detection by pairing an Autoencoder Feature Extractor (AFE) with a Model-Agnostic Meta-Learning (MAML) classifier. The idea is to compress 72 tabular behavioral features into a 64-dimensional latent space via the autoencoder, then let MAML adapt a binary classifier from just a handful of labeled malware-versus-benign samples. The authors pitch this as a rapid-deployment defense for healthcare and critical infrastructure settings where new ransomware variants appear with little warning.
Judgment models that return option probabilities without reasoning text are attractive as fast agent policies, but their competence boundary is unclear. This paper argues that such models excel at evaluation (judging given options against described criteria) yet fail at simulation (predicting missing consequences like opponent moves or physical dynamics). By cleanly separating the two, the authors show that supplying simulated outcomes via code restores strong performance across matrix games, text adventures, and robot control. The result is a sharp diagnostic of where fast judgment models can be trusted and a principled architecture for combining them with external simulators.
Asynchronous LLM post-training mixes stale rollouts from earlier policies, making GRPO-style optimization unstable. This paper derives a convergence bound that exposes a second-moment versus bias tradeoff, then proposes Group-Mass Capping GRPO (GMC-GRPO), which rescales groups when their average importance ratio exceeds a cap. The result is tighter delay dependence than per-trajectory clipping and stronger empirical robustness under large rollout delays on Qwen3 models.