Speech-based Alzheimer's assessment increasingly relies on self-supervised learning models, but their vulnerability to nuisance acoustic factors remains unclear. This paper asks whether noise and reverberation are merely encoded in SSL representations or actively drive diagnostic predictions. Using controlled input- and representation-space interventions on the ADReSSo dataset, the authors show that acoustic manipulations systematically shift classifier outputs even when no baseline group difference exists, challenging the assumption that predictive accuracy implies robustness.

This paper asks how continuous latent reasoning in chemistry can be improved by training thoughts to anticipate future outcomes without verbalizing every intermediate step. The authors introduce Latent-JEPA, which augments a recurrent latent reasoner with joint-embedding prediction of both textual continuations and molecular structures using pretrained encoders. Evaluated on ChemCoTBench, the framework shows the strongest task gains in molecular optimization and yields latent representations that are more chemically structured.

Selecting the best large language model for a live stream of prompts is difficult because benchmark scores are only coarse proxies for real-world performance and oracle annotations are costly. This paper proposes Online LLM Picker, the first active model selection framework designed specifically for LLMs in streaming settings: it uses an Exponential Weights posterior and an adaptive query probability based on hypothetical response variance scaled by posterior entropy to decide which prompts are worth annotating. Evaluated across over 130 models and 10 datasets spanning question answering, dialogue, grammar correction, and math, the method claims to cut annotation cost by up to 71.67% and reduce regret by up to 2.51× compared to baselines.

QK-Wanda attacks a blind spot in activation-aware LLM pruning: Wanda scores query and key weights independently even though their dot products jointly determine attention. The paper derives closed-form single-deletion costs under a coupled QK reconstruction objective, letting queries and keys share a pruning budget without gradients or weight updates. Across 15 models up to 72B parameters, the method cuts local QK reconstruction error by roughly half, yet downstream gains are uneven—Llama 2 and Qwen2.5 improve while Llama 3 degrades—highlighting that local reconstruction fidelity is an imperfect proxy for end-to-end model quality.

AURAL tackles the tension between reasoning quality and low latency in speech language models by replacing explicit chain-of-thought with adaptive latent reasoning. Instead of generating slow text intermediates, the model predicts chunks of continuous latent states drawn from a Gaussian mixture, cutting time to the first token by more than 11x while largely matching explicit CoT accuracy across two backbones. The work also constructs AuralReason-683K, a bilingual corpus of 683K utterances with concise reasoning traces used to initialize latent training.

Most LLM benchmarks treat generalization as aggregate accuracy on fixed prompts, which can hide per-example fragility when the same user intent is rephrased. This paper reframes generalization as behavioral stability across semantically equivalent input variations and proposes SAGO, a per-prompt, multi-axis framework that tracks changes in generation consistency, internal activations, confidence/uncertainty, and response mirroring. By evaluating eleven open- and closed-source models across six datasets and three variation families, the authors show that instability is heterogeneous, axis-specific, and not resolved by scale, challenging the assumption that high benchmark scores imply reliable generalization.

The paper argues that attributing model behavior to synthetic training data requires first recording how each item was generated, because a waveform-label pair alone erases upstream history. The authors introduce a structured provenance substrate that binds scenario seeds, generated content, audio waveforms, fact checklists, quality signals, and review lineage into an immutable manifest. They instantiate this in a five-pass audit of 113 synthetic Japanese care-handoff speech assets totaling 1.552 hours, finding that while versioned, seed-disjoint manifests support some downstream attribution research, exact upstream tracing is blocked by floating generator aliases and missing per-item TTS stamps. The work reframes data attribution as a two-stage problem: establish a typed, versioned generation path $z \longrightarrow x \longrightarrow m \longrightarrow r \longrightarrow b$ before attempting causal or contributive claims.

Diffusion language models generate text through iterative denoising rather than left-to-right prediction, making attribution a spatiotemporal problem: support for any output token is distributed across both network depth and timesteps. This paper introduces Diffusion Layer Integrated Gradients (DLIG), which extends Integrated Gradients to score a DLM's commitment to its completion at any chosen layer and denoising step, decomposing it into per-token contributions. It matters because existing token-attribution tools ignore the denoising trajectory entirely, while interventional alternatives are computationally expensive.

Technical documentation preserves collaborative context, but sentence-level specificity predictors were trained on general-domain prose. This paper audits whether two such predictors—SpeciTeller and a pinned post-publication implementation of Ko et al.'s domain-adapted model—produce stable, useful signals when applied to technical corpora (GitHub, Ansible, Python docs) and used to rank LLM-generated revisions. Using a three-stage evidence structure, the authors separate score interpretation, local correspondence with humans and controlled edits, and bounded decision value for post-generation candidate selection.

Most interpretability studies ablate or probe a single training checkpoint, leaving it unclear whether observed depth-dependent sensitivities are stable or transient. This paper tracks whole-block identity bypass across five released pretraining trajectories and eleven model-domain combinations to ask what persists and what shifts. The core finding is that the relative depth ordering of block sensitivities remains recognizable even as the magnitudes redistribute, and that the intervention pathway itself evolves differently across models and perturbation scales. The work matters because it cautions against treating single-checkpoint ablation profiles as static descriptions of a model's internal organization.

RPTune addresses product search for small merchants whose catalogs fit inside long-context LLMs. Rather than relying on retrieval pipelines, the authors show that feeding the full catalog works only when items are learned to be pruned, ranked, and positioned with downstream LLM feedback, after which the LLM itself is post-trained on these curated contexts using purely synthetic supervision. The result is a fully automated pipeline that achieves large accuracy gains and latency reductions across 7 merchants and 9 model families.

In-context learning (ICL) lets LLMs adapt to new tasks from demonstrations, but reprocessing all examples for every query is costly and prior compression methods fail on complex, input-dependent tasks. This paper introduces Task Operator (TO), a training-free method that extracts per-head affine transformations — scaling weights and bias vectors — from the ICL forward pass and replays them during zero-shot inference by updating attention output projections. By treating the ICL effect as a dynamic, state-dependent modification rather than a fixed activation vector, TO substantially narrows the gap between zero-shot and few-shot performance while revealing sparse, task-specific circuits.

This paper identifies empty commitments: promises that LLM agents make about future actions their runtime cannot actually execute. Unlike broken promises, these are structurally impossible from the moment of speaking because the agent lacks the required tools, scheduling, memory, or delegation affordances. The authors define a formal taxonomy of failure types and propose a measurement protocol, though no empirical results are presented.

Multi-agent LLM systems are usually judged by end-task accuracy, which conflates genuine information sharing with extra reasoning or fortunate architecture. This paper introduces Independent–Communicate–Revise (ICR), a controlled audit that freezes each agent’s initial trajectory and measures answer revision conditional on both agents’ initial correctness. By separating correction from preservation and comparing against a no-message control, the work shows that aggregate accuracy can hide large, policy-dependent trade-offs between fixing errors and corrupting correct answers.

A long-term assistant must recall memories whose cues are associatively related to, yet not semantically similar to, the current user turn. Madeleine reframes this involuntary recall as a learnable relevance—the pointwise mutual information of memories under simulated life trajectories—and distills it into a query-side residual adapter, $s_\theta(q,m)=\cos(E(q),E(m))+\cos(E_\theta(q),E(m))$. The system is trained offline on synthetic (trigger, cue) pairs written by an LLM life simulator, and at runtime it plugs into existing vector memories with zero LLM calls, cutting answer context by roughly $20\times$ while setting a new state of the art on LoCoMo-Plus.

Multi-agent workflows can solve complex tasks that stump a single LLM, but current methods typically train only the workflow generator while leaving execution agents frozen despite their shared influence on outcomes. This paper proposes FloWright, a paradigm that treats the workflow itself as a shared harness to optimize any role—generator, executor, or skill—using a hierarchical, structure-aware reward derived from execution traces, without extra models or labels. It also introduces DataWright, a rule-based data hardening procedure that converts single-agent datasets into workflow-level benchmarks. On 44 evaluation arms across documents, slides, charts, code, math, and finance, the authors show that co-evolving multiple roles yields larger gains than evolving one alone.

Knowledge distillation risks propagating the reversal curse when a teacher can recall a relation forward but not backward. This paper proposes directional label distillation: frozen teachers score candidate answers in the direction they know, and the highest-scoring candidate becomes the student's training target. On parent-child facts from Wikidata, this yields 13-15 point gains in student open-ended accuracy and shows that verification can provide accurate supervision where generation fails.

The paper asks when a nominally weaker LLM can reliably audit patches produced by a stronger coding agent. Its central claim is that groundability—the property that evidence exposes an independently checkable claim about missing behavior—matters more than reviewer scale. On a benchmark of 411 real execution-labeled traces and 101 controlled cases, the authors show that weak reviewers given official execution evidence can achieve near-perfect catch and near-zero over-rejection, while a deployable cascade of static checks and generated tests recovers only part of that upper bound. The finding that parameter count is not a monotonic predictor of review quality reframes the oversight problem from building bigger judges to building better checks.

Post-training quantization (PTQ) for LLMs is typically performed sequentially, block by block, by minimizing a reconstruction loss. The authors show that the standard MSE loss creates an Optimization Imbalance: because gradient magnitude scales with the square root of loss, early stages receive much weaker updates than later ones. Replacing MSE with RMSE implicitly normalizes gradients across stages, improving quantization accuracy—especially at aggressive bit widths like W2A16 and W4A4—with no extra parameters or meaningful compute overhead.

JoinGR attacks a real bottleneck in Text-to-SQL pipelines: dense retrievers rank tables independently and miss ones only identifiable through multi-hop join relationships. The method builds a column-level join graph with typed intra-table and foreign-key edges, selects semantic anchor tables via dense similarity, then propagates query-conditioned relevance deposits along edges using a lightweight learned scorer. Final table scores aggregate these deposits, making schema structure a first-class retrieval signal rather than an afterthought, which matters most for enterprise settings with cryptic table names and implicit relational dependencies.