Inference & Efficiency A

UNIEGO is a unified egocentric video encoder trained via a hierarchical multi-teacher distillation framework. Representation-specific proxy models translate knowledge from teachers spanning multiple viewpoints, modalities, and foundation models into a single egocentric space, while remaining deployable from egocentric video alone.

The paper studies multi-task Bayesian in-context learning, using in-context learning to perform Bayesian predictive inference across tasks. It targets the intractability of exact inference and the cost or restrictiveness of scalable approximations, aiming for uncertainty quantification and data efficiency.

The paper introduces Execution-State Capsules, a graph-bound mechanism to checkpoint and restore execution state for low-latency, small-batch, on-device physical-AI serving. It targets scenarios beyond the high-throughput, high-concurrency regime that paged or radix KV caches mainly serve.

Policy-adherent tool-calling agents in customer-service domains must track task state across turns while following rules. LedgerAgent introduces structured state to help such agents stay consistent and policy-compliant.

Neurosymbolic systems such as DeepProbLog combine neural perception with probabilistic logic, but standard inference has limits. DeepSWIP introduces quotient-WMC counterfactuals to enable counterfactual reasoning in neural probabilistic logic programs.

Flow-matching text-to-speech achieves strong zero-shot quality but stays static after deployment. FlowEdit uses associative memory to enable lifelong pronunciation adaptation without full retraining.

Securing AI agents that operate in complex digital environments has become critical, motivating runtime verification. This paper presents an efficient and sound probabilistic verification approach for AI agents.

The paper proposes marginal advantage accumulation, a cross-batch, operation-level mechanism for memory-driven agent self-evolution. It aims to distinguish stably effective memory operations from accidental hits, addressing contradictory feedback that the same operation can receive across different batches in trace distillation.

Context-heavy agents put unusual pressure on the key-value cache as long prefixes are reused across calls. UltraQuant applies 4-bit quantization to compress the KV cache while preserving quality.

The paper presents HEPTv2, an end-to-end efficient point transformer for charged-particle reconstruction. It targets tracking—reconstructing trajectories from sparse detector measurements under extreme combinatorial ambiguity—aiming to stay accurate and efficient at the High-Luminosity LHC.

The paper develops neural network surrogates with uncertainty quantification for inverse problems in partial differential equations. It targets the inference of unknown model parameters from noisy or incomplete observations, where traditional numerical methods are costly, particularly in Bayesian settings.

FP4 training promises large memory and compute savings for LLM pretraining but suffers from shrinkage bias. This paper analyzes its geometric origin and systemic impact and proposes a UFP4 recipe to address it.

Large language models show promise for code compilation tasks but struggle with runtime performance tuning. AutoPass uses evidence-guided LLM agents to perform compiler performance tuning.

This paper presents a robust Q-learning algorithm for discrete-time mean-field control problems with common noise, accounting for Wasserstein uncertainty in the model.

Agent skills are commonly deployed as natural-language Markdown files that encode answer policies. SoftSkill compresses such behaviors to enable more efficient contextual adaptation.

The paper proposes a token-operations-oriented framework for large-model inference optimization, presenting a four-layer technical architecture aimed at scalable, low-cost, and stable large-model services. The layers include components such as multi-model fusion.

Large language models perform strongly across language tasks but can hold conflicting parametric and contextual knowledge. This work proposes explicit knowledge conflict resolution to navigate unreliable knowledge during inference.

Using vision foundation models for camera-to-LiDAR knowledge distillation is promising. HilDA advances self-supervised LiDAR pre-training through hierarchical distillation with diffusion.

The paper characterizes when streaming tool use helps in streaming retrieval-augmented generation, which issues tool queries in parallel with ongoing user input to cut perceived latency. It argues the benefit is query-intrinsic and studies how tool intent stabilizes before an utterance is complete.

The paper proposes HydraHead, a hybrid attention design that exploits head-level functional heterogeneity to combine linear and full attention. It moves beyond the common layer-wise hybridization strategy, addressing the difficulty of integrating linear attention with full attention for efficient long-context processing.

The paper introduces GEMS, which uses geometric constraints to enable superposing multiple semantic directions in LLM activation steering. It addresses the collapse that occurs when existing single-direction steering methods inject several semantic directions at once without constraints.

The paper proposes a lightweight framework for automated pronunciation assessment based on discrete speech token surprisal, trained only on native speech resources. It operates unsupervised or with light calibration from a small set of scored utterances, avoiding costly labeled learner-error or non-native corpora.

The paper addresses the calibration gap in semantic caching, which cuts LLM inference costs by serving cached responses to semantically similar queries. It shows that evaluating with PR-AUC—which only measures ranking, not usability at a fixed threshold—leads to systematically poor deployment choices.

Post-training of reasoning models often combines supervised distillation with reinforcement learning from verifiable rewards, but distillation relies on costly chain-of-thought annotations. This work proposes rubric-conditioned self-distillation to rethink reward supervision while cutting annotation cost.

Enhancing formal math reasoning in LLMs has become a key focus, but most work relies on autoregressive generation. Diffusion-Proof offers a recipe for formal theorem proving using diffusion models, exploring proof search through a non-autoregressive framework and training approach.

Knowledge encoded in LLMs can serve as a substrate for structured reasoning over variables describing a complex world, but accessing it probabilistically is hard. This work performs structured inference over LLMs using Gibbs sampling, enabling probabilistic reasoning across interrelated variables.

Block diffusion language models speed decoding via parallel block-wise denoising, but reliably scaling them for long chain-of-thought reasoning is unresolved. The authors develop DreamReasoner-8B, using block-size curriculum learning to strengthen long-CoT reasoning in diffusion reasoning models.

Autonomous UAV operations on ships need reliable vision-based relative pose estimation, yet at-sea validation is costly, weather-dependent, and risky. This paper presents hardware- and vision-in-the-loop validation of deep monocular pose estimation for autonomous maritime UAV flight.

Model merging integrates the capabilities of several models fine-tuned from the same pretrained checkpoint into one, enabling multi-task learning. This work proposes Essential Subspace Merging, which extracts and merges each task's essential subspace to reduce interference and preserve multi-task performance.

Score- and flow-matching models often rely on preference-based RL both to align with subjective preferences and, surprisingly, to recover certain properties. This work argues the reward was in the data all along, correcting flow matching with discriminator-guided reinforcement learning.