Ablation studies in language models are confounded by self-repair, where downstream components appear to compensate for a removed part. This paper argues that such compensation is not dynamically recruited but is instead a standing, pre-existing opposition released by the intervention. The authors introduce a continuous dose axis for counterfactual interventions and show that downstream contributions follow an affine law predicted in part by static weights. If the finding generalizes, it would reshape how interpretability researchers infer component importance from ablations.

Intrinsic rewards in reinforcement learning aim to drive exploration by valuing experience, yet maximizing them need not yield the most informative data. This paper proposes comparing policies by counterfactual information: one policy dominates another if its interaction histories can be decoded to reproduce the other's history distributions uniformly across possible worlds. Using a simple alarm environment, the authors prove analytically that count-based, prediction-error, empowerment, and several state-based objectives can have strictly dominated optima, while also characterizing when posterior-based rewards succeed and constructing a native score $J_w$ that preserves every strict exploration improvement.

The paper tackles sampling from Gibbs distributions on matrix spaces with potentials that are neither convex nor globally gradient-Lipschitz. Its core idea is to replace the quadratic kinetic energy of underdamped Langevin dynamics with a non-quadratic, relativistic-like energy $K_\delta(p)=\delta^{-1}\operatorname{Tr}[(I_n+\delta pp^\top)^{1/2}-I_n]$, whose gradient tames momentum singular values smoothly. By carefully coupling a momentum-dependent diffusion matrix $M_\delta(p)$ and its divergence correction $\Gamma_\delta(p)$, the authors preserve the exact target Gibbs measure, recover regularised Muon in the zero-temperature limit, and prove exponential convergence of the diffusion plus uniform-in-time moment bounds for an explicit Euler-Maruyama scheme.

Most LLM agents treat repeated use of the same external source as a sequence of independent retrieval events. This paper reframes the deeper problem as source learning: progressively building a persistent, reusable source model that captures how a source’s knowledge is structured, interpreted, and applied. The proposed SourceLearn system combines self-directed study with task-guided refinement, and the authors report large gains across document, code, API, and interactive benchmarks.

This paper tackles competitive Super Smash Bros. Melee with a single Transformer policy that controls all 26 fighters from one checkpoint. The authors pretrain 10M- and 75M-parameter models on roughly 840k human replays, then apply rank-based supervised curricula, 75M-to-10M distillation, and Fox-only reinforcement learning. The resulting 10M policy wins 240 of 244 benchmark games against prior releases and runs within the 16.67 ms frame budget, while the authors openly acknowledge key confounds in evaluation.

The paper tackles how to teach LLMs new skills after pretraining without erasing old ones. While RL resists catastrophic forgetting by staying on-policy, it struggles when the model cannot generate successful trajectories on its own. The authors flip the problem: instead of changing the training loss, they use an MCMC projection sampler to reshape off-policy expert traces so they look more like samples from the base model while preserving correctness. This lets ordinary supervised finetuning match or beat RL and self-distillation on science, math, and medical tasks.

Robust Markov decision processes with polyhedral uncertainty lack exact convex representations because the natural robust Bellman inequalities become bilinear or union-of-polyhedra constraints. This paper shows how to encode a finite sequence of robust policy-iteration steps into a single exact LP whose optimal face recovers the robust optimal value and all optimal stationary randomized policies. At fixed discount, the LP has polynomial dimension and encoding length, and the authors use this representation together with a unified complexity analysis to prove strongly polynomial bounds for several standard uncertainty sets—including $\ell_1$, weighted $\ell_1$, interval, and Wasserstein—that previously lacked such guarantees.

This paper adapts Non-Linear Multi-Dimensional Concept Discovery (NLMCD) from vision to LLMs to model concepts as low-dimensional manifolds rather than linear directions. It introduces a fuzzy generalized Rand index called Concept-Based Alignment (CBA) that compares independently discovered concept spaces across layers, languages, and training stages without requiring shared tokenizers or architectures. The authors claim CBA is more sensitive than linear baselines (CKA, PCA), uses it to reveal cross-layer block structures, and applies it to study multilingual sharing, cross-model alignment, and post-training drift in Tulu-3.

When a machine-learning benchmark improves, multiple mechanisms could explain why. This paper studies how an autonomous agent should select diagnostic probes from a finite library to discriminate among competing explanations under a tight budget. MECHVAR scores each probe by the posterior-weighted variance of its predicted responses, showing that under a shared Gaussian predictive scale this is exactly proportional to the classical Box–Hill pairwise-KL criterion. The reformulation yields $O(KE)$ vectorized rescoring, an explicit pairwise audit trail, and orders-of-magnitude faster acquisition than quadrature-based expected information gain.

FedMIX-P tackles preconditioner coordination in federated learning by mixing a shared reference preconditioner with each client's local adaptive state at every local step. The core idea is to control the operator mismatch that arises when heterogeneous clients transform gradients differently, and the paper demonstrates empirical improvements on vision and language models under severe data heterogeneity. However, the convergence analysis is restricted to linear positive-definite preconditioners, while the headline experiments rely on nonlinear methods such as Muon, creating a notable gap between theory and practice.

Diffusion models amplify demographic biases present in training data, and current debiasing methods usually require costly attribute annotations, model retraining, or sacrifice output diversity. This paper proposes Debias Anything, a training-free guidance method that steers both unconditional and text-conditional diffusion models toward arbitrary target proportions for any sensitive attribute described in free text. By linking the frozen denoiser to a vision-language embedding space via a lightweight adapter, it jointly enforces batch-level fairness through optimal assignment and restores diversity through a semantic disagreement score, offering broad applicability without labeled data.

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.

Contextual linear optimization faces a discontinuity: small prediction errors can flip the LP solution to a different vertex, making regret non-differentiable pointwise. This paper proves that, after averaging over the data distribution, the population regret is locally quadratic with curvature concentrated on the walls of the normal fan. The authors derive the exact curvature measure $K_0$, provide a tractable single-projection approximation $\mathsf{K}_{\lambda,s}$, and apply it to scenario generation, showing a $30.8\%$ regret reduction on a battery arbitrage task.

Medical imaging models often fail when deployed across scanners or hospitals because they overfit to domain-specific textures and appearance statistics. PhaseAT addresses this by adversarially perturbing the Fourier phase of the luminance channel while keeping the amplitude spectrum fixed, generating hard training views that stress spatial organization under matched style. By encouraging the model to ignore brittle amplitude-driven correlations and focus on stable geometric structure, PhaseAT achieves large gains on both single-source and multi-source domain generalization benchmarks.

Can response safety be scored by cosine similarity to a safe-response embedding centroid? The paper shows that such a raw positive prototype identifies a location, not a separating direction, and formalizes why positive-only observations cannot determine the sign of the safe-unsafe gap. On three prompt-controlled safety corpora and four frozen encoders, the unreferenced prototype is weak or inverted while an explicit safe-minus-unsafe reference recovers useful ranking; the result is carefully scoped to the centroid rule, not all one-class methods.

CI pipelines for LLM training frameworks are drowning in GPU-heavy end-to-end tests that can take hours per run. FastCI attacks this bottleneck by replacing coarse file-level mappings with a runtime evidence graph built from lightweight stack sampling, then pruning redundant tests via caller/callee context, scheduling risky tests first, and shrinking workloads along dimensions outside each test's validation scope. Deployed at ByteDance, it claims a 77.5% latency reduction and 63.9% GPU savings, though the evaluation rests on an internal codebase with no open-source artifacts.

SkillEvoLean tackles the challenge of improving LLM agents for Lean theorem proving without updating model parameters. The core idea is to evolve natural-language procedural skills through progressive refinement from execution trajectories and, crucially, a mutation operator that perturbs skills in a semantic concept space to escape local optima when all proof attempts fail. This matters because existing skill-evolution methods stagnate on hard theorems where no sampled trajectory succeeds, leaving the solver with no signal for improvement.

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.

SIEVE tackles selective unlearning in vision-language models (VLMs), where a model must forget specific sensitive image-question-answer associations about an individual while retaining permitted facts about the same person. The core idea is to suppress attention-value representations for forget examples toward zero and anchor retain-example values to a frozen reference model, coupled with sequence-level negative cross-entropy. This matters because VLMs can recover biographical information from faces, creating privacy risks, and prior methods struggle to disentangle closely related forgotten and retained knowledge without degrading utility.