2026-W38

This paper introduces a framework to improve the explainability and verifiability of LLM reasoning in educational question answering. It achieves this by using a gold-anchored QLoRA for initial model adaptation, a task-aware router to direct problems to specialized symbolic solvers (like Z3), and a reinforcement learning approach (RLVR) that incorporates feedback from these solvers to refine the LLM's explanations. The core contribution is a more transparent and trustworthy reasoning process for LLMs, particularly in educational contexts.

Trace2Tower addresses limitations in LLM agent skill learning by introducing a transition-aware framework that distills raw execution traces into a hierarchical skill structure. It achieves this by abstracting interactions into canonical events, building a graph based on semantic and temporal relationships, and using spectral decomposition to identify stable, success-aligned behavioral modes. This results in a dynamic skill tower that improves agent performance on complex tasks.

This paper introduces a framework to improve the explainability and verifiability of LLM reasoning in educational question answering. It achieves this by using a gold-anchored QLoRA for model adaptation, a task-aware router to direct problems to specialized symbolic solvers (like Z3), and a reinforcement learning approach that incorporates feedback from these solvers to refine explanations. The core contribution is a more transparent and trustworthy reasoning process for LLMs, particularly in educational contexts.

Trace2Tower addresses limitations in LLM agent skill learning by introducing a transition-aware framework that distills raw execution traces into a hierarchical skill structure. It achieves this by abstracting interactions into events, constructing a graph based on semantic and temporal relationships, and using spectral decomposition to identify stable, success-aligned behavioral modes. This method creates a dynamic skill tower of action templates, routines, and strategies, leading to improved task performance.

This paper addresses the "consistency gap" in LLM agents, where agents are often accurate on average but unreliable across multiple runs. Their core method involves a "self-evolving agent framework" that identifies unstable steps in agent reasoning and converts them into episodic memory. This memory then guides future executions, significantly reducing the inconsistency gap and improving trustworthiness.

This paper introduces SPINE, a benchmark designed to measure LLM sycophancy under prolonged, adaptive disagreement. Unlike previous evaluations, SPINE uses a persistent LLM proxy to challenge the target model for up to 25 turns. The study reveals that sycophancy increases with conversation length, and current models are unreliable in resisting sustained pressure, even when their correct reasoning is retained.

MeClear is a novel framework for long-horizon LLM agents that addresses the problem of outdated or conflicting information in their external memory. It uses cooperative game-theoretic attribution to identify memories with negative downstream utility and selectively suppresses them. This approach ensures that the LLM agent's active context remains relevant and accurate, leading to improved task performance.

This paper introduces Gander, an end-to-end model that unifies multi-modal perception, real-time interaction, and agentic capabilities. Its core method is a Cerebellum-Brain collaborative framework, enabling continuous, full-duplex interaction where the Cerebellum handles real-time dialogue and the Brain manages complex reasoning and agent tasks. Gander's contribution lies in its ability to support natural, interruptible conversations and proactive agentic behavior across various modalities.

This paper introduces Training-Free Task Vectors (TFTVs), a method to control LLM behavior without costly fine-tuning. TFTVs compute task-specific directions in model weights using only forward-pass statistics, enabling edits like amplification, suppression, and composition of behaviors. This offers a practical and efficient alternative for post-training LLM editing.

This paper demonstrates that frozen transformers can perform data generation by simulating iterative samplers from in-context examples. The core method involves constructing transformers that realize closed-form diffusion samplers, where softmax attention computes weights and feedforward layers implement updates. The key contribution is showing that large language models can act as estimation-free samplers for generative tasks, extending in-context learning beyond supervised inference.

AuK is an open-source foundational model that unifies speech generation and editing using natural language instructions and audio context. Its core method involves a multimodal LLM for semantic understanding, a VAE for acoustic conditioning, and a hybrid rectified-flow Transformer for generation. AuK's contribution lies in its broad capabilities and extensive training data, enabling diverse speech manipulation tasks through a single, unified interface.

DiSCo is a novel framework for evaluating LLMs' cultural bias by focusing on their default preferences rather than single correct answers. It uses a forced-choice, distribution-first approach with a context gradient to isolate and measure inherent cultural leanings and steerability across multiple cultures. The paper's contribution lies in providing a more nuanced understanding of LLM cultural biases, revealing significant concentration in Western cultural preferences.

Show-Harness provides a semantic interface that allows Vision-Language Models (VLMs) to control robots by translating high-level intents into discrete actions. This method enables zero-shot control of powerful VLMs and efficient adaptation of smaller models, democratizing robot control.

COBRA-Skills optimizes LLM agent skills by treating it as a budgeted sequential optimization problem. It uses a contextual bandit to guide the selection of skills for evaluation, prioritizing promising or informative candidates. This approach efficiently refines the skill population using execution feedback, significantly reducing optimization costs and data requirements while achieving superior performance across diverse benchmarks.

This paper investigates how well an agent's memory survives model upgrades by comparing four memory storage methods: raw history, RAG chunks, natural language notes, and a fixed-schema knowledge graph. The core contribution is demonstrating that fixed-schema knowledge graphs offer the most robust memory portability, with minimal accuracy degradation after a model change. Conversely, compressed natural language notes are highly coupled to the original model, leading to significant accuracy drops.

This paper introduces GUT, a method to quantify and reduce reasoning uncertainty in LLMs. GUT models LLM reasoning as a directed acyclic graph, capturing all possible branches. It then uses this graph to measure uncertainty (GUT-Q) and optimize prompts to reduce it (GUT-O).

This paper proposes a novel method for measuring AI accountability by analyzing the structural quality of a model's defense for its verdicts, rather than relying on ground truth. Using a dialectical protocol based on argumentation theory, the research assesses how well AI models can withstand scrutiny and justify their reasoning in ambiguous situations. The contribution lies in providing a framework for evaluating AI moral reasoning that accounts for realistic ambiguity, moving beyond simple right/wrong assessments.

This paper introduces PRICE, a systematic framework for adapting Large Language Models (LLMs) for Bitcoin price forecasting. It combines efficient fine-tuning (LoRA), recursive inference, integer numerical representation, and specific prompting/decoding strategies to improve accuracy and reliability. The core contribution is demonstrating how these combined adaptation choices significantly enhance LLM performance in volatile financial markets.

The Mirror Agent Model proposes a Bayesian architecture where an agent's internal model of an observer is a direct reflection of its own state. This design enables the generation of interpretable agent behavior and explicit explanations for its actions, achieved by leveraging saliency methods to highlight relevant internal states. The core contribution lies in creating a unified framework for both generating and explaining agent behavior through this mirrored internal representation.

This paper introduces TruthInsightBench, a novel benchmark designed to evaluate the scientific discovery capabilities of AI agents, rather than just their ability to reproduce existing results. It presents agents with blind scientific objectives and frozen data, requiring them to independently formulate and support claims. The benchmark's contribution lies in its automated, evidence-grounded evaluation system that assesses the maturity of an agent's discoveries, enabling repeatable and scalable assessment of AI-driven scientific progress.

This paper investigates how well an agent's memory survives model upgrades by comparing four memory storage methods: raw history (LC-RAW), retrieved chunks (RAG), natural language notes (NOTES), and fixed-schema knowledge graphs (KG-fixed). The core contribution is demonstrating that fixed-schema structures like knowledge graphs offer robust memory portability, with minimal accuracy loss after model changes. In contrast, compressed natural language notes are highly coupled to the original model, leading to significant performance degradation.

This paper introduces GUT, a method to quantify and reduce Large Language Model (LLM) reasoning uncertainty. GUT models LLM reasoning as a directed acyclic graph, where graph complexity serves as a measure of uncertainty. The proposed GUT-Q module quantifies this uncertainty, while GUT-O aims to optimize LLM reasoning by reducing it.

This paper proposes a novel method for measuring AI accountability by analyzing the structural quality of a model's defense for its verdicts, rather than relying on ground truth. Their contribution is a four-phase dialectical protocol, grounded in argumentation theory, that assesses how well AI models can withstand scrutiny and justify their reasoning in ambiguous situations. This approach moves beyond simple multiple-choice evaluations to offer a more robust assessment of AI's moral reasoning capabilities.

This paper introduces PRICE, a systematic method for adapting LLMs to Bitcoin price forecasting. PRICE combines efficient fine-tuning (LoRA), recursive inference, integer numerical representation, and specific prompting/decoding strategies to improve accuracy. Its core contribution is demonstrating that each of these adaptation choices significantly enhances LLM performance for this challenging financial task.

The Mirror Agent Model proposes a Bayesian architecture where an agent's internal model of an observer is a direct reflection of its own. This design facilitates interpretable behavior and explanations by explicitly modeling the observer's perspective. The paper's contribution lies in this novel architecture and its extension with off-the-shelf saliency methods for generating agent explanations.

This paper introduces TruthInsightBench, a novel benchmark designed to evaluate the scientific discovery capabilities of AI agents, rather than just their ability to reproduce existing results. It achieves this by presenting agents with blind scientific objectives and frozen data, requiring them to independently derive and support claims. The benchmark's core contribution lies in its automated, evidence-grounded scoring system, which assesses the maturity of an agent's claims across six dimensions, enabling repeatable and scalable evaluation of AI-driven scientific discovery.

This paper introduces a method for co-evolving agent harnesses (the surrounding system components) and models. They found that evolving the harness alongside a weaker model, and then using an expert model with that evolved harness, helps bridge performance gaps. Crucially, directly imitating the expert's full trajectories under the evolved harness can surprisingly degrade the weaker model's performance.

This paper introduces Fact-Ablated Evaluation (FAE) to assess if LLMs truly use provided evidence for fact-checking or rely on internal knowledge. FAE iteratively removes evidence to see if LLM predictions change. The study finds current LLMs over-rely on parametric knowledge, leading to the development of REAL, a training method that uses counterfactual evidence to encourage evidence-dependent verification.

ExecCritic addresses the issue of coding agents generating incorrect tests that lead to false confidence in their repairs. Its core method is a "test-verify-revise" scaffold that separates test generation from code repair. This allows a dedicated "Test agent" to learn to create accurate, repository-native tests, which are then used by a "Repair agent" to fix code, ensuring that test errors do not propagate into the repair process.

This paper addresses the uncertainty of true contextual understanding in LLMs, as traditional metrics are insufficient. They introduce a novel knowledge graph-based evaluation framework, S3KG, which combines structural and semantic similarity to provide a continuous score. This framework, along with a diagnostic tool for error categorization, effectively measures contextual understanding in LLMs, particularly for question answering.

PlayTrain is an efficient reinforcement learning framework that leverages Large Language Models (LLMs) to generate adaptable JavaScript games from simple prompts. Its core method involves using LLMs for game creation and a pipeline to integrate these games into a standard RL environment. This framework's contribution lies in significantly reducing the manual effort required to create and adapt game environments for RL training, enabling high-throughput agent learning on novel or modified games.

This paper investigates the geometric differences in internal representations between State-Space Models (SSMs) and Transformers. It reveals that SSMs distribute information evenly across dimensions, while Transformers concentrate it along a dominant direction, a trend amplified by attention layers. This geometric divergence impacts representational capacity and compressibility, with SSMs showing surprising advantages despite their different structure.

This paper investigates the temporal dynamics of Preventative Steering, a defense against adversarial fine-tuning in LLMs. It reveals that the defense relies on an **active adaptation process** during training, rather than a static injection of corrective signals. The core contribution is demonstrating that the protection fades if the learned weight offsets are simply preserved or reinjected, highlighting the importance of ongoing adaptation and leading to the proposal of a new defense strategy, Progressive Intensity Scheduling.

This paper addresses the inefficiency of processing numerous visual tokens in Multimodal Large Language Models (MLLMs) by proposing VIP-Router. Instead of a single fixed pruning strategy, VIP-Router adaptively selects the best pruning method for each input based on its visual and textual features. This sample-adaptive approach improves efficiency by avoiding suboptimal pruning and retaining full token processing when beneficial.

This paper proposes a novel multimodal in-context learning (ICL) framework that moves beyond superficial imitation. It introduces contrastive demonstration modeling, where demonstrations explicitly highlight the difference between suboptimal and optimal responses along with the reasoning path for refinement. This approach aims to align MLLMs with the correct reasoning process, improving performance on complex multimodal tasks.

ConvMem addresses LLM limitations with long contexts by reformulating reasoning as a hierarchical convolution, treating the LLM as a kernel. This training-free, parallelizable method summarizes text segments in a logarithmic tree structure, significantly reducing latency compared to sequential approaches. Its key contribution is enabling efficient long-context reasoning without costly RL training.

This paper addresses the challenge of training LLM agents to discover multiple successful strategies from a single decision point. Their proposed method, Direct Diversity Optimization (DDO), combines a novel tree-based collection of diverse trajectories with a reference-relative objective. DDO effectively improves both task success and the coverage of different successful strategies, outperforming existing methods.

This paper proposes FOM-UL, a layer-selective unlearning method for LLMs. It identifies and updates only the most influential transformer layers for forgetting specific data, minimizing impact on the rest of the model. This targeted approach improves the balance between forgetting unwanted information and maintaining model utility, even after post-training quantization.

This paper proposes a framework for geometric reasoning using a Large Language Model (LLM) by integrating a Geometric Vision Parser and a Symbolic Solver. The parser translates geometric diagrams into symbolic representations, which the solver then uses for logical deduction. This approach achieves performance comparable to state-of-the-art multimodal models while offering more interpretable and human-like solutions.

This paper introduces Research Attention Prediction (RAP), a novel benchmark for evaluating LLMs as research agents by predicting future paper shares across AI/ML fields. The core contribution is identifying that LLMs struggle to accurately predict research trends, often underperforming a simple moving average baseline, and revealing that their evidence acquisition strategies, particularly how they carry forward past information, are key to this limitation.

This paper introduces TRACE, a method for training reasoning agents in complex diagnostic tasks where ground truth is hard to obtain. TRACE synthesizes verifiable rewards by simulating interventions in a controlled environment, allowing agents to learn causal reasoning from noisy, real-world-like data. The core contribution is demonstrating how to engineer objective rewards for diagnostic tasks, overcoming the limitations of traditional reinforcement learning in domains lacking easily verifiable answers.

This paper introduces a novel evaluation protocol for time-series foundation models by creating a test set of data published *after* model training. The core contribution is demonstrating that pretrained models still outperform others on this "later" test set, suggesting their familiarity with the training data, not just a new domain, is key to their success. However, intrinsic data properties like seasonal strength don't explain why pretrained models win in some cases but not others.

This paper introduces $Φ$-Bench, a novel benchmark designed to evaluate Large Language Models' (LLMs) ability to engineer the infrastructure that supports them. Unlike previous benchmarks, $Φ$-Bench covers the entire LLM infrastructure stack with tasks ranging from kernel optimization to system-level improvements, addressing the need for evaluating open-ended, long-horizon engineering capabilities.

ProbPlug is a lightweight framework that estimates the confidence of LLM binary classification predictions by analyzing internal token features from a frozen LLM. It uses a self-attention module to aggregate these features, providing reliable confidence estimates without altering the base model. This approach improves classification performance with minimal overhead and generalizes well across different tasks and model types.

This paper introduces an "artificial id" as an internal, adaptive drive for agentic AI systems. This mechanism allows agents to autonomously decide when to continue, stop, or change their behavior, moving beyond externally defined control. The core contribution is demonstrating that such adaptive direction and persistent alignment can emerge organically through differential persistence, even in agents lacking general reasoning capabilities.

MAPLE is a system that helps domain experts solve optimization problems using natural language. Its core method combines language understanding, mathematical programming, and evolutionary search to create and adapt optimization models. MAPLE's contribution is enabling rapid, iterative updates to optimization problems by remembering past decisions and search results, making optimization more accessible and adaptable to dynamic real-world scenarios.

This paper introduces MindTopo, a benchmark designed to evaluate foundation models' ability to reason about topological spatial relations, which are invariant under continuous deformation. The benchmark covers five cognitive properties (continuity, separation, order, enclosure, knots) at two levels (reasoning and planning) and contains over 11,000 instances. MindTopo's contribution is to shift the focus of foundation model evaluation from metric properties to the more fundamental topological aspects of spatial understanding.

SIRF addresses industrial content risk control by internalizing complex platform policies directly into model weights through continued pretraining. This "spec-internalization" allows for high-precision, ultra-low-latency risk assessment without human annotation or complex rule engines. The contribution is a foundation model that significantly improves risk handling efficiency and accuracy in real-world industrial deployments.

This paper introduces Negative Self-Distillation (NSD) to improve LLM reasoning. Unlike previous methods that force imitation of correct answers, NSD trains LLMs by having them actively avoid flawed reasoning. The core idea is to use the model itself to generate "negative" examples of incorrect reasoning, thereby encouraging more robust and exploratory problem-solving.

This paper introduces RAG-Safety-Bench, a novel benchmark designed to reliably evaluate the safety implications of Retrieval-Augmented Generation (RAG) in Large Language Models (LLMs). Its core method isolates the impact of RAG on safety by systematically controlling the retrieved documents, separating the effects of the retriever from the LLM's generation. The main contribution is providing a clear understanding of how RAG can inadvertently compromise LLM safety, even when designed for improved reliability.

This paper introduces **OR-Clarify**, a benchmark for evaluating how well Large Language Models (LLMs) can identify and resolve missing information before formulating optimization problems. The core method involves presenting LLMs with incomplete problem descriptions and assessing their ability to ask clarifying questions to a simulated user. The main contribution is a novel framework, **InterOPT**, which uses identified formulation gaps to guide the LLM's interaction and improve the final optimization model.

This paper introduces OR-Clarify, a benchmark for evaluating how well Large Language Models (LLMs) can identify missing information before formulating optimization problems. The core method involves presenting LLMs with incomplete problem descriptions and assessing their ability to ask clarifying questions to a simulated user. The main contribution is a new framework, InterOPT, which uses these identified gaps to guide the optimization process, leading to more accurate and robust solutions.

ActMap introduces a novel white-box method for single-pass uncertainty quantification in LLMs. It compresses the entire generation-time hidden-state trajectory into a compact, fixed-size tensor (ActMap) with no added computational cost. This ActMap can then be directly probed by a lightweight classifier to determine the trustworthiness of a generated answer, localizing uncertainty signals to specific regions of the model's internal computation.

This paper introduces the concept of "job power elasticity" to quantify how LLM training performance degrades with reduced GPU power. The authors propose the "Power Flexibility Index (PFI)" as a metric to measure this sensitivity and enable SLA-aware power management for AI workloads. This work is crucial for making AI training power-flexible, which can alleviate data center power constraints and improve grid utilization.

This paper introduces ORCH, a method for organizing embodied AI agents into task-specific hierarchical structures. ORCH leverages principles from human organization theory to combine pooled and sequential interdependencies, enabling effective coordination in complex, multi-agent scenarios like wildfire response. The contribution lies in demonstrating that flexible, principle-driven organization significantly enhances collective intelligence in embodied AI, outperforming fixed structures.

This paper proposes Bayesian backward reasoning as a novel approach for multi-agent collective decision-making, particularly when agents disagree. Instead of relying on traditional forward reasoning, it constructs a "reverse posterior" by reasoning backward from an explicit likelihood. This different factorization of information allows for a more robust assessment of agent reliability, using Jensen-Shannon divergence to rank agents based on their cross-path consistency, ultimately leading to improved collective decisions.

CausalArena addresses the challenge of evaluating causal discovery methods, especially with the rise of foundation models, by providing a unified and adaptable benchmark. Its core method involves a common protocol and diverse Structural Causal Models (SCMs), including synthetic, semantic, and formula-grounded types, to offer controlled breadth and human-auditable environments. The contribution is a more reliable and interpretable evaluation framework that accounts for potential biases in existing benchmarks and better assesses true causal discovery capabilities.