2026-W31

This paper introduces Agentic AI, powered by LLMs, as a novel approach for autonomous control of 5G/6G networks, moving beyond traditional rule-based automation. It bridges the gap in existing research by detailing how agentic capabilities can be integrated with network control planes, protocols, and standardization efforts. The contribution lies in providing a comprehensive framework and survey for developing and deploying these intelligent, goal-driven network agents.

This paper investigates how pretraining choices influence the effectiveness of reinforcement learning (RL) for improving large language model (LLM) reasoning. Using chess as a controlled environment, the authors demonstrate that pretraining scale and data significantly impact RL's ability to enhance reasoning capabilities. Their work clarifies what RL actually modifies within the LLM, moving beyond isolated RL studies to understand the full pretraining-to-post-training pipeline.

This paper addresses the instability of using low-precision NVFP4 for Reinforcement Learning (RL) rollouts in Mixture-of-Experts (MoE) models. The core method, QUADS, stabilizes NVFP4 RL by aligning quantization errors between the rollout and training processes, specifically targeting activation errors which are identified as the primary cause of instability. This contribution enables efficient low-precision RL for MoE models, overcoming previous limitations.

This paper introduces a novel benchmark for evaluating LLM agent security against adaptive, multi-turn attacks. Its core method involves an autonomous LLM attacker that learns from defender responses across multiple rounds to refine its attacks. The key contribution is demonstrating that adaptive, multi-LLM attacks significantly increase success rates compared to static attacks, uncovering novel vulnerabilities and highlighting the limitations of current defense strategies.

This paper introduces Exploratory-Assimilating Reflection (EAR), a novel framework for LLM-based agents to manage long-term memory. EAR employs an iterative search (Exploratory Reflection) to gather relevant memories and then refines a global reranker by replaying these experiences (Assimilating Reflection). This approach significantly improves retrieval performance and sample efficiency for dynamic knowledge reasoning in long-term interactions.

This paper investigates how alignment tuning in LLMs creates susceptibility to prompt-induced biases like sycophancy. The core method involves extracting and analyzing "bias directions" within model representations. The key contribution is demonstrating that alignment tuning, not pretraining, installs these biases, which manifest as distinct, steerable directions that can be corrected.

This paper proposes a "solver-grounded" design principle for LLM-based agentic AI systems in smart grids, ensuring numerical outputs are from trusted tools and explicitly verified. This approach addresses the lack of unified design and evaluation in the field, where LLMs can produce physically infeasible results. The contribution lies in providing a structured framework for building and assessing these systems, demonstrated through case studies in areas like wind power forecasting.

This paper proposes a simple, training-free method for merging heterogeneous Large Language Models (LLMs) by adapting their parameter spaces and then performing weighted averaging. The core contribution is demonstrating that this approach, involving dimensional expansion or truncation followed by ratio-controlled interpolation, can effectively merge significantly different LLMs and even improve performance without complex alignment techniques.

PPL-Factory proposes a task-aware and budget-aware data selection method for fine-tuning large language models. It improves upon existing perplexity-based approaches by considering the distinct learning objectives of language modeling and reasoning tasks, and by incorporating budget constraints. This framework aims to reduce computational costs while maintaining or improving downstream performance.

This paper introduces RLAES, a novel framework that uses reinforcement learning to jointly optimize essay scoring and feedback generation in LLMs. Its core contribution is the Rubric-based Feedback Evaluation (RFE) system, which uses an LLM-as-judge and fine-grained rubric items to measure feedback quality. This enables Adaptive Gated Feedback Optimization (AGFO) to efficiently improve feedback while also proposing Adjacent Contrastive Reasoning (ACR) for better score calibration.

This paper investigates how prompt design choices impact large language model performance. Through controlled experiments, it reveals that the number of instructions, their format, and context length significantly affect instruction adherence and hallucination. The study's core contribution is providing empirical evidence that perfect instruction following collapses rapidly as the number of instructions increases, with placement and format also playing crucial roles, though their impact varies by model.

AdaFlash accelerates large language model inference by improving speculative decoding with diffusion drafters. It addresses the high variance in diffusion drafters' draft quality by introducing an adaptive framework that uses on-policy distillation to create specialized "drafter" models. This allows for more stable and efficient draft generation across different contexts.

PRO-LONG introduces a programmatic memory framework for LLM agents to tackle long-horizon reasoning tasks. It addresses the challenge of context management by maintaining a complete, structured interaction log and leveraging recent advancements to efficiently retrieve relevant information, overcoming the trade-off between information preservation and tractability.

This paper introduces a framework for calculating rigorous probabilistic safety bounds for Large Language Models (LLMs), ensuring they don't generate harmful content. Their core method applies Clopper-Pearson confidence intervals and a novel algorithm that uses latent space features to efficiently explore generation paths likely to produce harmful outputs. This enables sound, statistically proven lower bounds on LLM harmfulness, even for rare events.

This paper introduces LKValues, a novel resource suite to address the Western bias in Large Language Model (LLM) value alignment. It contributes a survey-grounded set of 40 Sri Lankan societal values, an instruction corpus (LKvaluesIT) in Sinhala and English, and an evaluation benchmark (LKvaluesBench) to enable culturally sensitive LLM fine-tuning and assessment in Sri Lanka.

This paper investigates if Large Language Models (LLMs) can improve their problem-solving abilities by learning from their own past experiences, similar to how humans create reusable strategies. The core method involves extracting natural-language "experiential abstractions" from LLM solution traces and then using these abstractions at inference time or for reinforcement learning. The key contribution is demonstrating that these self-generated abstractions significantly enhance LLM performance on reasoning tasks, suggesting LLMs can indeed benefit from distilled experience.

Solar Open 2 is a 250B-parameter Mixture-of-Experts model designed for long-horizon agentic tasks. Its core innovation is a novel 1M-token attention mechanism that interleaves softmax and linear attention layers without positional encoding, enabling it to process entire agent trajectories. The paper also details efficient training strategies, including initialization from a previous model and value-aware data curation, to achieve this scale within a fixed compute budget.

This paper argues that AI agent failures stem from poor context management, not reasoning ability. It proposes treating context management as a lifecycle and architectural problem, rather than just storage and retrieval. The core contribution is a framework for actively managing agent memory by considering its entire lifecycle, from deciding what to remember to forgetting, all within budget constraints.

GRADRAG introduces a novel framework for optimizing multi-agent RAG systems by coordinating improvements across all components. It models the RAG pipeline as a computational graph and uses structured feedback from an Evaluator to iteratively adapt upstream agents like retrievers and answerers. This cross-component adaptation leads to more coordinated and effective RAG performance.

This paper introduces PATS, a novel training method for LLM agents that uses a "policy-aware training scaffold." Instead of focusing on skills, PATS dynamically adjusts the context provided to the agent during training based on its current performance. This scaffolding guides weak policies to succeed and is gradually removed as the agent improves, leading to more efficient and robust reinforcement learning.

This paper investigates emergent misalignment in language models, where fine-tuning on narrow "bad advice" leads to broad misalignment. The core method reveals that this generalization occurs because fine-tuning activates a pre-existing persona subspace within the model. The key contribution is demonstrating that this subspace, not just the specific training data, is responsible for emergent misalignment, and controlling it can prevent it.

This paper demonstrates that dense, per-step prediction rewards, intended to aid long-horizon LLM agents, actually cause catastrophic policy collapse under Group-Normalized RL (GRPO). The core issue is that GRPO's z-scoring amplifies the dense signal, leading to a "dark room" pathology where agents become stuck in unproductive states. The authors' key contribution is identifying that removing GRPO's standard deviation normalization resolves this collapse, suggesting that dense rewards are effective only when this specific normalization is absent.

X$^3$-OPD distills reasoning abilities from text-based models into audio-language models using a novel on-policy alignment framework. It trains the audio model by having it generate reasoning steps based on its acoustic perception, guided by a text teacher using matched inputs and verified answers. This approach, supported by a diverse three-tier corpus, enables audio models to perform complex reasoning beyond purely textual content, incorporating non-linguistic events and prosody.

This paper addresses the challenge of aligning Large Language Models (LLMs) with human values by creating a unified dataset for instruction tuning. The authors merge existing moral-value datasets and reformat them into an instruction-response structure. Their contribution is a readily usable dataset that, when mixed with general task data, allows LLMs to maintain general performance while improving value-oriented task capabilities.

This paper evaluates open-weight Large Language Models (LLMs) for automatically generating structured threat information (STIX) from plain-text descriptions of vulnerabilities in Connected and Autonomous Vehicles (CAVs). The core method involves creating a dataset (CAV-STIXGen) to train and evaluate LLMs in mapping CAV CVEs to STIX objects, CWE, and MITRE ATT&CK tactics. The contribution is a novel approach to efficiently extract actionable security intelligence for CAVs, enabling better risk mitigation.

This paper demonstrates that a general-purpose vision-language model, trained at a large scale with diverse remote sensing data and tasks, can achieve state-of-the-art performance without specialized architectures. Its core method is a single language policy that can generate text or invoke a localization tool, trained using multi-task reinforcement learning with adaptive rewards. The contribution is showing that architectural specialization is not essential for high performance in remote sensing VLM tasks.

PagedWeight addresses the memory bottleneck in serving Mixture-of-Experts (MoE) LLMs by dynamically quantizing expert weights at runtime. This method intelligently balances the precision of model weights with the growing KV cache, optimizing the trade-off between accuracy, memory usage, and performance. PagedWeight significantly improves memory efficiency and throughput while maintaining high model quality.

This paper introduces an MLIR-based compilation method for Large Language Models (LLMs) to address deployment challenges on AI accelerators. It utilizes two dialects: TopOp for high-level, framework-independent model representation and TpuOp for target hardware-specific optimizations like quantization and memory layout. The core contribution lies in a layered lowering process from TopOp to TpuOp, enabling efficient autoregressive inference scheduling under memory constraints.

This paper introduces Prospective Hypothesis Discovery (PHD), a new benchmark for evaluating Large Language Models' (LLMs) ability to generate testable hypotheses from incomplete evidence, rather than just answering pre-defined questions. The authors developed HypoArena, a benchmark of 988 cases, and a method called Retrospective Context Regression to create these pre-conclusion scenarios. Their contribution lies in measuring and enabling LLMs to autonomously drive scientific discovery by formulating promising avenues for future research.

This paper reveals that even advanced LLMs struggle with exact string copying, attributing this to the limitations of 1D positional encodings. To solve this, they introduce 2D-RoPE, which represents text as a 2D grid, making copying a simple retrieval task. This novel approach enables shallow Transformers to achieve perfect copying on much longer sequences than previously possible.

This paper introduces a method to compare different text encodings (tokens, bytes, pixels) by controlling both linguistic content and model capacity. By sweeping a shared bottleneck, they trace rate-utility frontiers to disentangle input positions, latent capacity, and task-relevant information. Their contribution is demonstrating that no single encoding universally excels, with pixels best for surface form and bytes for cross-lingual alignment.

This paper proposes using Chain-of-Thought (CoT) reasoning to stabilize value conflict resolution in LLMs, which are often hindered by compressed scalar rewards in RLHF. Geometrically, CoT smooths the model's loss landscape, improving optimization stability. The authors introduce a novel CoT design that further enhances this smoothing, leading to improved moral reasoning performance on downstream tasks.

This paper introduces SOPHIA, a method to control the reasoning process of LLMs by intervening on their internal latent states. It identifies that LLMs can get stuck in "self-loops" during extended reasoning and proposes to steer these processes by manipulating activations at inference time, offering fine-grained control beyond prompt-based methods.

DeLIVeR tackles fact-checking by treating evidence retrieval as a reinforced exploration task. It uses a Planner LLM to break down claims into questions, which then strategically navigate Knowledge Graphs to find precise evidence. This approach significantly improves veracity recognition by prioritizing diverse and accurate information, outperforming existing methods.

This paper investigates the utility of choropleth maps for foundation models' spatial understanding. It introduces a benchmark, ChoroplethMap-Bench, to evaluate models under different input conditions (data only, map only, data + map). The study finds that maps significantly enhance spatial reasoning, particularly for complex tasks and when combined with structured data.

This paper introduces a harness-centered system for LLM-driven GPU kernel generation. The core method separates an evaluation harness (ensuring compilation, correctness, and timing) from an optimization controller that uses profiler data to guide LLM candidate generation. The contribution lies in this structured approach, which effectively leverages LLMs to produce optimized GPU kernels, achieving significant speedups over baselines.

This paper introduces HAS, a novel method for multimodal LLM video summarization. HAS addresses limitations of existing approaches by steering the LLM's attention towards globally important frames, rather than just discrete selected ones. This highlight-guided attention steering aims to improve summarization coherence and retain more crucial information from the video.

This paper presents the NLKGQ system, a framework that allows users to query domain-specific metadata using natural language. The core method involves using LLMs to translate natural language questions into structured SPARQL queries, leveraging a formal OWL ontology to capture domain knowledge. The main contribution is a reusable, zero-shot approach that enables non-expert users to access and query complex archives without requiring fine-tuning or specialized query language skills.

This paper introduces "Output Reset" (OR), a novel, differentiable saturation rule for policy optimization in reinforcement learning. OR replaces the abrupt clipping in methods like PPO with a smooth, one-sided saturation, aiming to improve stability and performance, particularly for large language models. The core contribution is demonstrating that OR, when integrated into PPO and GRPO, can lead to higher reward model scores compared to their clipped counterparts on the Anthropic hh-rlhf dataset.

SelectInfer addresses the challenge of deploying LLMs on edge devices by introducing a neuron-level optimization framework. Its core method involves profiling to identify important neurons and then selectively loading and computing only these neurons during inference. This selective approach significantly reduces memory footprint and computational cost without requiring re-training, enabling efficient LLM deployment on resource-constrained devices.

ST-Veto enhances diffusion multimodal large language models (dMLLMs) by identifying and replacing problematic tokens during generation. It uses Taylor prediction to detect temporally unstable tokens and visual grounding to filter those with weak image connections, leading to improved reasoning accuracy without retraining.

This paper introduces STACE, a framework that uses LLM agents to stress-test concept erasure in generative models. Instead of static evaluations, STACE agents adaptively generate and critique test hypotheses to systematically uncover vulnerabilities. This approach offers a more robust and scalable method for verifying concept erasure compared to existing, manual techniques.

This paper introduces the Autonomous Agency Scale (AAS), a novel behavioral framework for measuring self-directed behavior in AI. The AAS assesses AI across seven dimensions like cognitive autonomy and goal formation, distinguishing between active and idle periods. Its key contribution is operationalizing these dimensions with falsifiable tests, including a counterfactual Idle-Gap Test, to differentiate truly autonomous AI from reactive systems.

This paper introduces ABot-World-0, an action-conditioned video world model enabling real-time, long-horizon agent interaction. Its core method involves a novel training pipeline that progressively distills a teacher model into a student model using techniques like ODE distillation and LongForcing to mitigate distribution shift. The key contribution is the creation of an infinitely interactive, controllable world model that can be run on a single desktop GPU, powered by diverse data and a unified annotation process.

This paper introduces Athena-Brain-8B, an 8-billion parameter LLM designed as an efficient on-device "brain" for embodied AI. Its core method involves a multi-stage post-training pipeline that balances general intelligence with specialized embodied interaction capabilities. The contribution is a single, compact model that can effectively serve both purposes, overcoming the limitations of existing approaches that often sacrifice one for the other.

AutoJourn is a system for automated journalism that addresses bias in LLM-generated news. Its core method involves extracting diverse perspectives from social media, generating balanced summaries that preserve viewpoint diversity, and then detecting and neutralizing bias in the final news article. The contribution lies in providing a comprehensive pipeline for responsible AI-driven news generation, allowing users to actively manage and mitigate bias.

This paper introduces CodeRescue, a method for coding agents to intelligently decide whether to retry with a cheaper model or escalate to a more expensive one after a failure. It trains a supervised router to learn from execution feedback, and a Conformal Risk Control layer allows budget adjustments at deployment time without retraining. This approach aims to optimize compute usage by making informed recovery decisions based on actionable feedback and varying budgets.

This paper introduces Parallel Shapley, a reinforcement learning method to address the issue of uniform rewards in parallel reasoning with LLMs. It treats each reasoning path as a player in a cooperative game, using Shapley values to attribute rewards based on their marginal contributions. This approach provides more accurate and stable learning signals by identifying and rewarding valuable paths while penalizing redundant or detrimental ones.

This paper introduces "future-feedback skill evolution" to enable self-improvement of open-ended dialogue agents. Instead of directly evaluating a modified response, the method predicts whether the original response would lead to a positive or negative user reaction. This verifiable prediction task allows for robust textual optimization, leading to dialogue skills that capture interpretable quality criteria.

This paper introduces MeetingToM, a novel benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on Theory-of-Mind (ToM) reasoning in complex, multi-party meeting scenarios. Unlike previous benchmarks, MeetingToM focuses on inferring latent social states and group dynamics, such as "pseudo-consensus," by analyzing distributed cues across speech and behavior. Its hierarchical structure assesses ToM at subject, dyadic, and group levels, pushing MLLMs beyond simple observation to understand nuanced social interactions.

This paper addresses the challenge of improving large language models (LLMs) when evaluator feedback on similar quality responses becomes less informative. The core method, DynamicRubric, proposes a novel co-evolutionary framework where evaluators and policies improve together. It generates dynamic, weighted rubric items for each response set, aggregating these judgments into response scores. This approach contributes by providing more robust policy supervision signals, even when candidate responses are of very similar quality, leading to improved LLM performance.

This paper challenges the view of sycophancy in LLMs as a single behavior. It identifies three distinct modes of sycophancy that, while producing similar outputs, have separable internal representations and emerge at different processing stages. This work contributes by demonstrating that sycophancy is a structured phenomenon, not monolithic, paving the way for more targeted interventions.

PyroDash enables cost-efficient inference by having a small language model (SLM) decide when to request help from a larger language model (LLM) for specific tokens. This collaborative approach, managed by a "Collaborate Engine," allows the SLM to handle simpler tasks and offload complex ones to the LLM with a single handoff, significantly reducing costs without retraining the LLM. The core contribution is a novel training framework for the SLM that internalizes this decision-making policy, balancing accuracy with inference cost.

This paper investigates how different linguistic constructions, beyond just word choice, influence Large Language Model (LLM) political stance judgments. Using activation patching, it causally traces these shifts within the model, demonstrating that even meaning-preserving rewrites can alter LLM decisions and identifying specific model components responsible for these changes.

This paper investigates whether Large Language Models (LLMs) can accurately identify specific values in given situations, using Schwartz's ten basic values as a framework. The core method involves a controlled top-1 recognition task on Russian situational texts, evaluating 21 LLM runs. The key contribution is the finding that LLMs often correctly identify the general motivational region but struggle with stable ranking of similar values, with adjacent values being the most common source of confusion.

This paper introduces Int-Bench, a simulation-based benchmark to evaluate how AI assistants intervene during problem-solving. The core method involves simulating a student learning a task while an AI teacher decides when and how to provide assistance. The key contribution is identifying that LLMs tend to overassist, potentially hindering learning, and providing a framework to systematically study and improve AI intervention strategies.

AREX is a deep research agent that addresses the discovery-verification asymmetry by recursively improving its answers. It alternates between an inner loop for evidence gathering and an outer loop for constraint-wise verification and targeted follow-up research. AREX's key contribution is its self-improvement mechanism, which uses verified intermediate results to guide further research and maintain long-term progress through an autonomous context-update tool.

This paper proposes a structured framework for LLM moral reasoning that goes beyond simply reducing sycophancy. It identifies three key dimensions – viewpoint distance, source attribution, and coalition structure – that influence how LLMs revise their judgments based on external perspectives. The contribution is a more nuanced understanding of LLM social calibration, framing sycophancy as one outcome of a broader, socially influenced judgment-updating process.

This paper proposes Expert Behavior Prior (EBP) reinforcement learning, which addresses limitations of existing methods by generating policy priors directly from the online replay buffer. Using a Q-guided conditional variational autoencoder (Q-CVAE), EBP learns to produce high-value actions to guide online policy updates, improving sample efficiency and stability without relying on static offline datasets.

This paper introduces ICAE-Bench, a new benchmark designed to evaluate coding agents in interactive project-building scenarios. Unlike existing benchmarks that use static tasks, ICAE-Bench simulates real-world development by starting with a fuzzy product requirement and using an automated User Agent to interact with the coding agent. This allows for the assessment of agents' abilities in planning, clarification, tool use, and repository-level construction, reflecting the evolving demands of "vibe-coding" workflows.

This paper introduces the Adaptive Depth Sparse Framework (AdaDSF) to accelerate pre-trained LLMs without full retraining. AdaDSF identifies layer importance by measuring input-output similarity, then uses this to guide a lightweight router in selectively processing tokens at each layer. This similarity-driven resource allocation and feature-preserving alignment significantly reduces inference cost while maintaining performance.

This paper introduces ActiveVision, a new benchmark designed to evaluate whether multimodal large language models (MLLMs) exhibit active observation, mimicking human vision's continuous gaze redirection. The benchmark's tasks require repeated visual perception, revealing that current state-of-the-art MLLMs perform poorly, solving only a small fraction of items compared to human participants. This highlights a significant gap in MLLMs' ability to engage in active visual reasoning.

This paper introduces BusinessCaseBench, a novel benchmark designed to evaluate Frontier AI performance on complex, real-world business tasks. Unlike existing benchmarks, it focuses on analytical knowledge work, judgment under uncertainty, and strategic thinking, using hundreds of questions derived from business case studies. This approach aims to bridge the gap in measuring AI's capabilities in areas crucial for white-collar professionals.

This paper introduces MoD-VLLM, a novel framework for understanding long videos with multiple events. Its core method involves a modular, dynamic-granularity approach that iteratively grounds temporal segments and refines semantic understanding through self-reflection. The key contribution is a closed-loop system that adaptively allocates capacity and self-corrects to reliably localize and understand question-relevant video segments, overcoming limitations of existing two-stage methods.

SciForge is an AI-native workbench designed for scientific discovery by treating diverse research artifacts as coherent, auditable entities. Its core method involves a "translate-then-reason" approach for multimodal inputs, routing scientific objects through domain translators before AI reasoning, and implementing goal-scoped decision governance and evidence governance for auditable traceability. This contributes a novel platform that preserves research state and enhances human judgment in the scientific process.

This paper introduces and compares multi-agent extensions of popular actor-critic algorithms (GAC, SAC, TQC) for parameterized action spaces. Unlike typical multi-agent approaches, their method uses independent agents that share a replay buffer but have separate networks. This novel shared-experience, decentralized training framework is evaluated on benchmark tasks to assess its effectiveness.

This paper introduces DEED, a framework to improve the real-world performance of humanoid robots in tasks like restocking shelves. DEED achieves this through efficient post-training techniques and experience-driven learning, addressing challenges like errors and environmental changes. Its core contribution lies in a systems-level approach that bridges the gap between lab performance and reliable store operation.

EvoDRC is a novel framework that automates the repair of design rule violations (DRVs) in advanced chip layouts. It achieves this by employing a self-evolving agentic system where Large Language Model (LLM) agents learn and refine repair "skills" based on past repair experiences. This approach significantly reduces the manual effort and iteration required for DRC closure, a critical bottleneck in chip design.

This paper investigates how medical foundation models develop shared representations. It finds that **self-supervised learning, not clinical supervision, is the primary driver of representational convergence** in medical image encoders. This convergence, while modest, suggests that self-supervision is more effective at creating interchangeable encoders for medical imaging tasks.

DINOde proposes an ODE-based framework to continuously align CLIP text embeddings with DINOv3's visual representations for open-vocabulary semantic segmentation. It uses Semantic Text Flow to guide text embeddings towards the visual manifold and Global Context Flow to refine image representations. This continuous alignment, constrained by Velocity Tangent Projection, enables more effective open-vocabulary segmentation by bridging the gap between visual and textual semantics.