LLM unlearning is usually tested by whether a model refuses harmful queries and retains clean-task accuracy. The authors highlight a gap: models can pass those checks while failing when benign prompts sit just outside the forgotten domain, a failure they term forget-adjacent utility loss. They propose RA2, which redirects forget-token hidden states toward nearby retain-supported anchors instead of simply pushing them away. Experiments on WMDP and MUSE show this directional choice meaningfully reduces neighborhood-level collateral damage.

This paper investigates whether high-performing humanoid control networks can be synthesized directly in weight space rather than trained conventionally. The authors adopt a latent diffusion framework---pairing a cluster-wise linear autoencoder with a Diffusion Transformer---to generate weights for 3-layer MLP policies on a Unitree G1 robot, achieving zero-shot sim-to-real transfer for walking, dancing, and spin kick. Alongside generation, they analyze the learned weights and identify a kinematic prior in first-layer parameters that emerges under both reinforcement and imitation learning, suggesting embodiment-specific structure in policy weights.

This paper tackles digital-domain self-interference cancellation in in-band full-duplex radios, where conventional adaptive filters degrade when the desired signal is much stronger than the residual interference.
The authors reformulate cancellation as an operator-learning problem: a Transformer neural operator learns the temporal evolution of the self-interference channel from historical observations and reconstructs the current interference from the known transmit signal.
Trained in a desired-signal-free setting and applied during contaminated reception, the method reduces reliance on corrupted instantaneous updates and reports strong high-SIR stability.

Neural symbolic regression models are bottlenecked by training datasets riddled with structural redundancy and simplicity bias. This paper proposes ExCaUS, a deterministic grammar-based framework that exhaustively enumerates canonical univariate skeletons under explicit unary-count and depth bounds, eliminating duplicate functional forms. The authors pair this with a derivative-driven sampling strategy and a 150-problem benchmark, demonstrating that curated structural diversity exposes severe weaknesses in state-of-the-art SR methods.

The paper investigates when jointly optimizing imputation with a downstream classifier improves robustness under train–test mismatches in missingness rates or mechanisms. Instead of treating imputation as fixed preprocessing, the authors cast missing values as learnable variables regularized around a reference imputation, giving both amortized and latent instantiations. Through 98 tabular classification datasets—and targeted shift experiments on 36 datasets—they show that reconstruction quality is a poor proxy for predictive utility and that task-aware corrections matter most in hard regimes such as MNAR and severe distribution shift.

Most standard diversity measures violate at least one of the Mironov–Prokhorenkova axioms (monotonicity, uniqueness, continuity), and the only known measures that satisfy all three are NP-hard to compute. This paper introduces Geometric Kernel Diversity (GKD), which defines diversity as the volume of the union of radial-kernel under-graph solids for a point set in Euclidean space. For a canonical Student-$t_2$ kernel, the authors prove that GKD satisfies all three axioms on Euclidean-realizable inputs, give exact polynomial-time algorithms for dimensions $d \leq 3$, and provide a randomized $(1\pm\varepsilon)$-approximation for arbitrary dimensions. The contribution is a theoretically clean first positive answer in the restricted Euclidean setting, though the axiomatic question on general distance matrices and exact computation in $d \geq 4$ remain open.

Conflicting studies on whether neural networks converge to a shared representation have used different alignment metrics and reached opposite conclusions. This paper proposes the Lattice Representation Hypothesis (LRH), which formalizes alignment criteria into a lattice ordered by strength, and argues that disagreements arise because prior work evaluates criteria at different points on this lattice. It also proves in idealized models that the training procedure and objective determine which criteria are met reliably across independent runs, supported by experiments on dozens of pretrained vision and language models.

GRPO for long-horizon agents suffers from sparse terminal rewards that yield all-fail rollout groups and zero advantage signals early in training. This paper studies guide-assisted curricula that replay successful reference prefixes before the learner continues, systematically varying the initial position, handoff granularity, and movement rule. Across Qwen3 models from 1.7B to 8B on BFCL and AppWorld, the authors find that starting near task completion and using outcome-informed call-level selection improves independent execution while reducing training cost. The work usefully frames curriculum design as a practical lever for GRPO-based agent post-training.

This paper studies noiseless efficient global optimization via Gaussian Process Thompson Sampling (GP-TS) in the frequentist RKHS setting. It proves that standard GP-TS suffers an intrinsic $\sqrt{\log N}$ simple-regret penalty for Mat\'ern kernels, yielding $O(N^{-\nu/d}\sqrt{\log N})$, and proposes Capped Thompson Sampling (GP-CTS), which achieves the minimax-optimal $\Theta(N^{-\nu/d})$ rate using a known RKHS norm bound. The analysis covers both discrete and continuous domains and is notable for matching upper and lower bounds in a non-bandit setting.

RM-OPD targets a real bottleneck in generalist reward modeling: training on mixed-domain data with standard RL produces cross-domain interference and sparse outcome rewards, so capabilities like math verification, code review, and safety guardrails cannot coexist without degrading each other. The core idea is a two-stage framework—first cultivate domain-specialist reward teachers via single-domain GRPO, then distill them into one student through routed on-policy distillation, domain contribution reweighting, and criteria context distillation. This matters because it swaps coarse verdict-level supervision for dense token-level guidance, raising the multi-domain performance ceiling for generative judges.

Autoregressive transformers trained on limit order book streams are usually evaluated only by output realism, leaving their internal mechanics opaque. This paper inverts the problem by training GPT-style models on synthetic order-flow simulators with known latent parameters, enabling exact verification against Bayes-optimal predictors and oracle entropy floors. It finds that next-token pretraining implements amortized Bayesian inference and online filtering, that latent states are linearly decodable from frozen activations, and that regime diversity—not model scale or token count—drives generalization.

Multimodal large language models suffer large accuracy drops when their visual inputs are corrupted by noise, blur, or weather artifacts. This paper asks whether test-time feature alignment—matching activation statistics to clean source moments, a technique that works well for CLIP-like classifiers—can recover that lost performance without labels or weight updates. Across three modern MLLMs and 15 standard corruptions, the authors find it cannot: sample-blind, position-shared corrections recover only about one percentage point of an order-of-magnitude larger drop. The core insight is that corruption shifts in MLLM visual token sequences are nearly orthogonal across images and dominate the residual term in a per-token squared-shift decomposition, whereas pooled classification embeddings exhibit a shared, low-rank shift direction. This matters because it establishes that robustness in generative MLLMs is a per-token restoration problem, not a simple domain-statistics-matching problem.

Latent world models for control are routinely selected and checkpointed by one-step next-embedding prediction loss, on the assumption that lower loss implies better planning. This paper tests that proxy in JEPA architectures by introducing a frozen-decoder protocol that reports one-step error in interpretable physical units, then applies fine-tuning interventions and measures goal-conditioned planning with paired seeds. It finds the relationship saturates after a single epoch within a task and has no stable exchange rate across tasks, implying that latent prediction loss is an unreliable stand-in for planning competence.

This paper identifies progress stagnation—unproductive reasoning-action loops—as a critical failure mode when distilling multi-turn agent capabilities from large teachers into small language models via supervised fine-tuning. To mitigate it, the authors propose Progress-Aware Distillation (ProD), which measures the stagnation gap between student-generated and teacher trajectories and encodes that discrepancy into an adaptive margin within a SimPO-style preference objective. The contribution matters because prior agent distillation improves tool-use accuracy yet collapses on longer horizons, directly undermining practical deployment of efficient small-model agents.

Steady PDE surrogates on unstructured meshes must resolve multiscale features from a single coarse discretization. This paper proposes routing a mixture of vanilla transformer experts using a learned discrete Hodge-Laplacian controller that measures high-frequency content per vertex. Each expert is biased toward a different spectral scale through auxiliary band targets, and the combined output predicts pressure, velocity, and turbulence fields in one forward pass. On a remeshed AirFRANS benchmark, the architecture cuts mean relative $L_2$ error by roughly a third compared with a Transolver baseline.

This paper tackles unsupervised pre-training for reinforcement learning: how to learn from reward-free trajectories so that a good policy can be extracted once a task reward is revealed. The core idea is to replace costly, iterative flow-based occupancy models with one-step drifting models, which makes it possible to extract policies by differentiating directly through the generator instead of distilling a separate critic. This matters because it cuts inference cost, removes an entire distillation stage, and improves performance on standard continuous-control benchmarks.

The paper studies how to learn interpretable binary classifiers as sparse Boolean rules by optimizing the weights of a fixed-structure DNF network. It proposes mixed integer linear programming (MILP) formulations that trade off training misclassification against rule sparsity, with an advanced formulation that dramatically reduces integer variables and redundant constraints while preserving exact Boolean semantics. A combined warm-start and dropout strategy is introduced to break symmetries and push scalability to larger datasets.

CoPE attacks the sample-efficiency bottleneck in training long-horizon language agents with reinforcement learning. Standard agentic RL such as GRPO optimizes only a sparse terminal reward $Y(\tau) \in \{0,1\}$, leaving the rich temporal structure of intermediate steps unused. The paper introduces CoPE, which augments GRPO with success-conditioned multi-step action forecasting: on successful on-policy trajectories, the shared policy backbone learns to predict future action chunks $a_{t:t+K}$ from each history $h_t$, so every action is supervised from both the moment it is taken and from the $K$ preceding histories. This implicit temporal ensemble yields stronger aggregate performance and up to $2\times$ sample-efficiency gains across five benchmarks, with no added inference cost.

Data-quality gates conflate within-region example selection with across-region mass redistribution, so a gate-versus-random comparison cannot isolate what improved downstream performance. This paper introduces Counterfactual Gate–Support Decomposition (CGSD), a four-arm, budget-matched design that decomposes the gate effect into within-support selection $E$, between-support redistribution $S$, and their interaction $I$ by crossing eligibility and support-share factors. The core contribution is not the factorial algebra itself, but operationalizing it for finite pools while characterizing when the required arms are unrealizable.

INTACT attacks the inference cost of chain-of-thought reasoning by replacing early textual reasoning steps with compact continuous latent vectors while keeping the backbone LLM frozen. A small 107M-parameter encoder iteratively builds a latent prefix from the frozen model's own hidden states, first distilling the effect of written reasoning steps and then refining via group-relative policy gradients. Across Llama and Qwen backbones from 1B to 8B, the method improves accuracy and shortens generation length, and a single encoder trained on math, science, and code transfers to twelve held-out benchmarks including commonsense without adaptation.