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 $59.0\%$.

LESS introduces a lightweight neural architecture search pipeline that marries a brief fair supernet warm-up with candidate-conditioned hard-path updates inside a CMA-ES loop. The core claim is that just six balanced SGD steps per proposal are enough to sharply reduce selection regret, enabling NAS-Bench-201 searches in roughly seven minutes on a single RTX 4090 without retuning across datasets. If this speed–accuracy trade-off holds more broadly, it could lower the barrier to reproducible NAS experiments.

This paper studies how to make Quantum Reinforcement Learning (QRL) for combinatorial optimization more robust on near-term quantum devices. Its core idea is APGEM, a LinUCB contextual-bandit controller that dynamically selects among error-mitigation techniques—Zero-Noise Extrapolation, Probabilistic Error Cancellation, Clifford Data Regression, and Readout Error Mitigation—based on a cost-aware utility that balances fidelity gain against shot cost and estimator variance. The work matters because static error mitigation cannot adapt to changing noise regimes or training phases, and adaptive orchestration could improve the reliability of hybrid quantum-classical loops.

Unified language models must handle heterogeneous capabilities such as mathematics, coding, instruction following, and controllable reasoning, yet standard single-trajectory training is notoriously sensitive to ordering and data ratios. Naive weight-space merging of single-objective experts offers a modular alternative but often collapses distinct behaviors or degrades cross-domain performance. This paper proposes Mixture-Trained Merging (MTM), which trains each branch on an objective-biased data mixture rather than a pure objective so that the branches stay in mutually compatible regions of parameter space and can be merged at deployment time.

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.

The paper solves the generalized Schrödinger bridge on finite graphs—transporting probability mass between distributions while charging costs for visited states—by showing that state costs fold into the reference process via a Feynman–Kac tilt. This reduces the problem to solving a plain bridge against the tilted generator $A = Q - \operatorname{diag}(f)$ with sparse matrix-exponential iterations, eliminating the need for learned controls or tuned temporal-difference penalties. The method scales to millions of nodes and extends to congestion costs through a provably convergent damped best-response outer loop.

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.

This paper tackles handwritten Bangla word recognition by casting it as character-level object detection rather than whole-word classification. The authors ensemble SSD-MobileNetV2 and Faster R-CNN-InceptionResNetV2 to spot individual characters and diacritics in unsegmented word images, then rearrange the predictions into text. They introduce a modified Non-Maximum Suppression and a heuristic called Preserving Necessary Overlap (PNO) to cope with the cursive matra line and overlapping diacritics that make Bangla OCR notoriously difficult.

LLMs can hit high accuracy on hard math problems while lacking genuine structural insight. This paper proposes the Mathematical Primitive---a concise, non-procedural statement of the key idea that makes a solution possible---and introduces Prim, a benchmark that disentangles structural understanding from procedural execution across Discovery, Generation, Digestion, and Execution. The authors show that independent Discovery is the dominant bottleneck and introduce Absorb, a primitive-privileged self-distillation framework that selectively transfers structural guidance into the student model without requiring primitives at inference time.

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.

Mechanistic interpretability tries to identify sparse circuit subgraphs that explain model behavior, typically evaluating candidates via intervention-defined faithfulness. This paper demonstrates that faithfulness objectives can prefer an equally sized circuit that reproduces intact-model behavior worse than an alternative, because ablating excluded components distorts the computational context of retained ones. The authors argue this creates an objective-level recovery gap: better discovery algorithms cannot recover true mechanisms if the evaluation metric itself systematically rewards the wrong candidate.

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.

Most trajectory generators require check-ins from the deployment city, which is unrealistic when only a public POI map is available. This work studies target-trajectory-free generation: synthesizing human mobility in a new city using only source trajectories and the target city's POI coordinates and categories. Its key insight is to replace absolute destination prediction with transferable relative transitions—semantic and geographic displacements between place contexts—and ground them onto the target POI graph via a behavior graph and an exploration–return walk. This factorization yields a theoretically principled and empirically evaluated method for zero-shot mobility synthesis.

The paper asks whether stochastic rounding (SR) can improve low-precision transformer inference over round-to-nearest (RN), and shows that the benefit is site-dependent: SR outperforms RN in long MLP reductions because its $O(\sqrt{n}u)$ probabilistic error envelope beats the $O(nu)$ deterministic drift of RN, while RN is superior at the language-model head because SR's non-uniform variance is heavily penalized by the Fisher-weighted cross-entropy. The authors support this with a fast variable-precision emulation tool (PRISM), a vector Bienaymé–Chebyshev bound, a second-order loss decomposition, and experiments on DistilGPT-2 that motivate mixed-precision rounding recipes.

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.

This paper studies gradient descent with a large constant stepsize on linearly separable logistic regression, a canonical model for the Edge of Stability phenomenon. It proves that the worst-case exact transition time from oscillatory loss to monotone decrease scales as $\Theta((\log \eta)^{\min\{n-2,d-2\}})$ uniformly over dimension $d \ge 2$, disproving a conjecture that stepsize independence holds beyond $d=2$. The result translates into substantially stronger finite-horizon optimization guarantees by permitting exponentially larger constant stepsizes than prior $O(\eta)$ transient bounds.