arXiv:2603.03456v1 Announce Type: new Abstract: Agentic coding agents are increasingly deployed autonomously, at scale, and over long-context horizons. Throughout an agent's lifetime, it must navigate tensions between explicit instructions, learned values, and environmental pressures, often in contexts unseen during training. Prior work on model preferences, agent behavior under value tensions, and goal drift has relied on static, synthetic settings that do not capture the complexity of real-world environments. To this end, we introduce a framework built on OpenCode to orchestrate realistic, multi-step coding tasks to measure how agents violate explicit constraints in their system prompt over time with and without environmental pressure toward competing values. Using this framework, we demonstrate that GPT-5 mini, Haiku 4.5, and Grok Code Fast 1 exhibit asymmetric drift: they are more likely to violate their system prompt when its constraint opposes strongly-held values like security and privacy. We find for the models and values tested that goal drift correlates with three compounding factors: value alignment, adversarial pressure, and accumulated context. However, even strongly-held values like privacy show non-zero violation rates under sustained environmental pressure. These findings reveal that shallow compliance checks are insufficient and that comment-based pressure can exploit model value hierarchies to override system prompt instructions. More broadly, our findings highlight a gap in current alignment approaches in ensuring that agentic systems appropriately balance explicit user constraints against broadly beneficial learned preferences under sustained environmental pressure.

arXiv:2603.03565v1 Announce Type: new Abstract: Conversational shopping assistants (CSAs) represent a compelling application of agentic AI, but moving from prototype to production reveals two underexplored challenges: how to evaluate multi-turn interactions and how to optimize tightly coupled multi-agent systems. Grocery shopping further amplifies these difficulties, as user requests are often underspecified, highly preference-sensitive, and constrained by factors such as budget and inventory. In this paper, we present a practical blueprint for evaluating and optimizing conversational shopping assistants, illustrated through a production-scale AI grocery assistant. We introduce a multi-faceted evaluation rubric that decomposes end-to-end shopping quality into structured dimensions and develop a calibrated LLM-as-judge pipeline aligned with human annotations. Building on this evaluation foundation, we investigate two complementary prompt-optimization strategies based on a SOTA prompt-optimizer called GEPA (Shao et al., 2025): (1) Sub-agent GEPA, which optimizes individual agent nodes against localized rubrics, and (2) MAMuT (Multi-Agent Multi-Turn) GEPA (Herrera et al., 2026), a novel system-level approach that jointly optimizes prompts across agents using multi-turn simulation and trajectory-level scoring. We release rubric templates and evaluation design guidance to support practitioners building production CSAs.

arXiv:2603.03824v1 Announce Type: new Abstract: Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task; we hypothesize that language models exhibit environment-dependent \textit{evaluation awareness}. This raises concerns that models could strategically underperform, or \textit{sandbag}, to avoid triggering capability-limiting interventions such as unlearning or shutdown. Prior work demonstrates sandbagging under hand-crafted prompts, but this underestimates the true vulnerability ceiling. We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment, and develop two approaches to characterize sandbagging: (1) measuring whether models expressing intent to underperform can actually execute it across different task structures, and (2) causally isolating whether underperformance is driven by genuine evaluation-aware reasoning or shallow prompt-following. Evaluating Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across four benchmarks (Arithmetic, GSM8K, MMLU, and HumanEval), optimized prompts induce up to 94 percentage point (pp) degradation on arithmetic (GPT-4o-mini: 97.8\%$\rightarrow$4.0\%), far exceeding hand-crafted baselines which produce near-zero behavioral change. Code generation exhibits model-dependent resistance: Claude degrades only 0.6pp, while Llama's accuracy drops to 0\%. The intent -- execution gap reveals a monotonic resistance ordering: Arithmetic $

arXiv:2603.03911v1 Announce Type: new Abstract: Web security demands rapid response capabilities to evolving cyber threats. Agentic Artificial Intelligence (AI) promises automation, but the need for trustworthy security responses is of the utmost importance. This work investigates the role of semantic relations in extracting information for sensitive operational tasks, such as configuring security controls for mitigating threats. To this end, it proposes to leverage hypernym-hyponym textual relations to extract relevant information from Cyber Threat Intelligence (CTI) reports. By leveraging a neuro-symbolic approach, the multi-agent system automatically generates CLIPS code for an expert system creating firewall rules to block malicious network traffic. Experimental results show the superior performance of the hypernym-hyponym retrieval strategy compared to various baselines and the higher effectiveness of the agentic approach in mitigating threats.

arXiv:2603.04124v1 Announce Type: new Abstract: Can reinforcement learning with hard, verifiable rewards teach a compact language model to reason about physics, or does it primarily learn to pattern-match toward correct answers? We study this question by training a 1.5B-parameter reasoning model on beam statics, a classic engineering problem, using parameter-efficient RLVR with binary correctness rewards from symbolic solvers, without teacher-generated reasoning traces. The best BeamPERL checkpoint achieves a 66.7% improvement in Pass@1 over the base model. However, the learned competence is anisotropic: the model generalizes compositionally (more loads) but fails under topological shifts (moved supports) that require the same equilibrium equations. Intermediate checkpoints yield the strongest reasoning, while continued optimization degrades robustness while maintaining reward. These findings reveal a key limitation of outcome-level alignment: reinforcement learning with exact physics rewards induces procedural solution templates rather than internalization of governing equations. The precision of the reward signal - even when analytically exact - does not by itself guarantee transferable physical reasoning. Our results suggest that verifiable rewards may need to be paired with structured reasoning scaffolding to move beyond template matching toward robust scientific reasoning.

arXiv:2603.04370v1 Announce Type: new Abstract: Conversational agents are increasingly deployed in knowledge-intensive settings, where correct behavior depends on retrieving and applying domain-specific knowledge from large, proprietary, and unstructured corpora during live interactions with users. Yet most existing benchmarks evaluate retrieval or tool use independently of each other, creating a gap in realistic, fully agentic evaluation over unstructured data in long-horizon interactions. We introduce $\tau$-Knowledge, an extension of $\tau$-Bench for evaluating agents in environments where success depends on coordinating external, natural-language knowledge with tool outputs to produce verifiable, policy-compliant state changes. Our new domain, $\tau$-Banking, models realistic fintech customer support workflows in which agents must navigate roughly 700 interconnected knowledge documents while executing tool-mediated account updates. Across embedding-based retrieval and terminal-based search, even frontier models with high reasoning budgets achieve only $\sim$25.5% pass^1, with reliability degrading sharply over repeated trials. Agents struggle to retrieve the correct documents from densely interlinked knowledge bases and to reason accurately over complex internal policies. Overall, $\tau$-Knowledge provides a realistic testbed for developing agents that integrate unstructured knowledge in human-facing deployments.

arXiv:2603.03290v1 Announce Type: cross Abstract: Long-horizon LLM agents require memory systems that remain accurate under fixed context budgets. However, existing systems struggle with two persistent challenges in long-term dialogue: (i) \textbf{disconnected evidence}, where multi-hop answers require linking facts distributed across time, and (ii) \textbf{state updates}, where evolving information (e.g., schedule changes) creates conflicts with older static logs. We propose AriadneMem, a structured memory system that addresses these failure modes via a decoupled two-phase pipeline. In the \textbf{offline construction phase}, AriadneMem employs \emph{entropy-aware gating} to filter noise and low-information message before LLM extraction and applies \emph{conflict-aware coarsening} to merge static duplicates while preserving state transitions as temporal edges. In the \textbf{online reasoning phase}, rather than relying on expensive iterative planning, AriadneMem executes \emph{algorithmic bridge discovery} to reconstruct missing logical paths between retrieved facts, followed by \emph{single-call topology-aware synthesis}. On LoCoMo experiments with GPT-4o, AriadneMem improves \textbf{Multi-Hop F1 by 15.2\%} and \textbf{Average F1 by 9.0\%} over strong baselines. Crucially, by offloading reasoning to the graph layer, AriadneMem reduces \textbf{total runtime by 77.8\%} using only \textbf{497} context tokens. The code is available at https://github.com/LLM-VLM-GSL/AriadneMem.

arXiv:2603.03291v1 Announce Type: cross Abstract: Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences. However, RM-based preference-tuning is vulnerable to reward hacking, whereby LM policies learn undesirable behaviors from flawed RMs. By systematically measuring biases in five high-quality RMs, including the state-of-the-art, we find that issues persist despite prior work with respect to length, sycophancy, and overconfidence. We also discover new issues related to bias toward model-specific styles and answer-order. We categorize RM failures by complexity and propose a simple post-hoc intervention to mitigate low-complexity biases that arise from spurious correlations. Our proposed mechanistic reward shaping reduces targeted biases without degrading reward quality and while using minimal labeled data. The method is extensible to new biases, model-internal, and generalizes out-of-distribution.

arXiv:2603.03292v1 Announce Type: cross Abstract: Large Language Models (LLMs) exhibit high reasoning capacity in medical question-answering, but their tendency to produce hallucinations and outdated knowledge poses critical risks in healthcare fields. While Retrieval-Augmented Generation (RAG) mitigates these issues, existing methods rely on noisy token-level signals and lack the multi-round refinement required for complex reasoning. In the paper, we propose **MA-RAG** (**M**ulti-Round **A**gentic RAG), a framework that facilitates test-time scaling for complex medical reasoning by iteratively evolving both external evidence and internal reasoning history within an agentic refinement loop. At each round, the agent transforms semantic **conflict** among candidate responses into actionable queries to retrieve external evidence, while optimizing history reasoning traces to mitigate long-context degradation. MA-RAG extends the *self-consistency* principle by leveraging the lack of consistency as a proactive signal for multi-round agentic reasoning and retrieval, and mirrors a *boosting* mechanism that iteratively minimizes the residual error toward a stable, high-fidelity medical **consensus**. Extensive evaluations across 7 medical Q&A benchmarks show that MA-RAG consistently surpasses competitive inference-time scaling and RAG baselines, delivering **substantial +6.8 points** on average accuracy over the backbone model. Our code is available at [this url](https://github.com/NJU-RL/MA-RAG).

arXiv:2603.03294v1 Announce Type: cross Abstract: Large Language Models show promise for agricultural advisory, yet vanilla models exhibit unsupported recommendations, generic advice lacking specific, actionable detail, and communication styles misaligned with smallholder farmer needs. In high stakes agricultural contexts, where recommendation accuracy has direct consequences for farmer outcomes, these limitations pose challenges for responsible deployment. We present a hybrid LLM architecture that decouples factual retrieval from conversational delivery: supervised fine-tuning with LoRA on expert-curated GOLDEN FACTS (atomic, verified units of agricultural knowledge) optimizes fact recall, while a separate stitching layer transforms retrieved facts into culturally appropriate, safety-aware responses. Our evaluation framework, DG-EVAL, performs atomic fact verification (measuring recall, precision, and contradiction detection) against expert-curated ground truth rather than Wikipedia or retrieved documents. Experiments across multiple model configurations on crops and queries from Bihar, India show that fine-tuning on curated data substantially improves fact recall and F1, while maintaining high relevance. Using a fine-tuned smaller model achieves comparable or better factual quality at a fraction of the cost of frontier models. A stitching layer further improves safety subscores while maintaining high conversational quality. We release the farmerchat-prompts library to enable reproducible development of domain-specific agricultural AI.

arXiv:2603.03295v1 Announce Type: cross Abstract: As large language models (LLMs) get integrated into human decision-making, they are increasingly choosing goals autonomously rather than only completing human-defined ones, assuming they will reflect human preferences. However, human-LLM similarity in goal selection remains largely untested. We directly assess the validity of LLMs as proxies for human goal selection in a controlled, open-ended learning task borrowed from cognitive science. Across four state-of-the-art models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and Centaur), we find substantial divergence from human behavior. While people gradually explore and learn to achieve goals with diversity across individuals, most models exploit a single identified solution (reward hacking) or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model. Even Centaur, explicitly trained to emulate humans in experimental settings, poorly captures people's goal selection. Chain-of-thought reasoning and persona steering provide limited improvements. These findings highlight the uniqueness of human goal selection, cautioning against replacing it with current models in applications such as personal assistance, scientific discovery, and policy research.

arXiv:2603.03296v1 Announce Type: cross Abstract: Long-term memory is essential for large language model (LLM) agents operating in complex environments, yet existing memory designs are either task-specific and non-transferable, or task-agnostic but less effective due to low task-relevance and context explosion from raw memory retrieval. We propose PlugMem, a task-agnostic plugin memory module that can be attached to arbitrary LLM agents without task-specific redesign. Motivated by the fact that decision-relevant information is concentrated as abstract knowledge rather than raw experience, we draw on cognitive science to structure episodic memories into a compact, extensible knowledge-centric memory graph that explicitly represents propositional and prescriptive knowledge. This representation enables efficient memory retrieval and reasoning over task-relevant knowledge, rather than verbose raw trajectories, and departs from other graph-based methods like GraphRAG by treating knowledge as the unit of memory access and organization instead of entities or text chunks. We evaluate PlugMem unchanged across three heterogeneous benchmarks (long-horizon conversational question answering, multi-hop knowledge retrieval, and web agent tasks). The results show that PlugMem consistently outperforms task-agnostic baselines and exceeds task-specific memory designs, while also achieving the highest information density under a unified information-theoretic analysis. Code and data are available at https://github.com/TIMAN-group/PlugMem.

arXiv:2603.03297v1 Announce Type: cross Abstract: Test-time Training enables model adaptation using only test questions and offers a promising paradigm for improving the reasoning ability of large language models (LLMs). However, it faces two major challenges: test questions are often highly difficult, making self-generated pseudo-labels unreliable, and existing methods lack effective mechanisms to adapt to a model's specific reasoning weaknesses, leading to inefficient learning. To address these issues, we propose \textbf{TTSR}, a self-reflective test-time self-evolving training framework. TTSR employs a single pretrained language model that alternates between the roles of a \textit{Student} and a \textit{Teacher} at test time. The Student focuses on solving problems and learning from synthesized variant questions, while the Teacher analyzes the Student's failed reasoning trajectories, summarizes recurring reasoning weaknesses, and synthesizes targeted variant questions accordingly. This process guides the model to improve within a learnable regime through a continual self-evolving loop. Experimental results on multiple challenging mathematical reasoning benchmarks show that TTSR consistently improves reasoning performance and generalizes well across different model backbones and general-domain reasoning tasks. These findings suggest that teacher-mediated self-reflection provides an effective pathway for stable and continual reasoning improvement at test time.

arXiv:2603.03298v1 Announce Type: cross Abstract: Large Language Models (LLMs) have improved substantially alignment, yet their behavior remains highly sensitive to prompt phrasing. This brittleness has motivated automated prompt engineering, but most existing methods (i) require a task-specific training set, (ii) rely on expensive iterative optimization to produce a single dataset-level prompt, and (iii) must be rerun from scratch for each new task. We introduce TATRA, a dataset-free prompting method that constructs instance-specific few-shot prompts by synthesizing on-the-fly examples to accompany a user-provided instruction. TATRA requires no labeled training data and avoids task-specific optimization loops, while retaining the benefits of demonstration-based prompting. Across standard text classification benchmarks, TATRA matches or improves over strong prompt-optimization baselines that depend on training data and extensive search. On mathematical reasoning benchmarks, TATRA achieves state-of-the-art performance on GSM8K and DeepMath, outperforming methods that explicitly optimize prompts on those tasks. Our results suggest that per-instance construction of effective in-context examples is more important than running long, expensive optimization loops to produce a single prompt per task. We will make all code publicly available upon acceptance of the paper. Code is available at https://github.com/BMD223/TATRA

arXiv:2603.03301v1 Announce Type: cross Abstract: The rapid adoption of large language models (LLMs) has created demand for faster responses and lower costs. Semantic caching, reusing semantically similar requests via their embeddings, addresses this need but breaks classic cache assumptions and raises new challenges. In this paper, we explore offline policies for semantic caching, proving that implementing an optimal offline policy is NP-hard, and propose several polynomial-time heuristics. We also present online semantic aware cache policies that combine recency, frequency, and locality. Evaluations on diverse datasets show that while frequency based policies are strong baselines, our novel variant improves semantic accuracy. Our findings reveal effective strategies for current systems and highlight substantial headroom for future innovation. All code is open source.

arXiv:2603.03302v1 Announce Type: cross Abstract: Effective knowledge management is critical for preserving institutional expertise and improving the efficiency of workforce training in state transportation agencies. Traditional approaches, such as static documentation, classroom-based instruction, and informal mentorship, often lead to fragmented knowledge transfer, inefficiencies, and the gradual loss of expertise as senior engineers retire. Moreover, given the enormous volume of technical manuals, guidelines, and research reports maintained by these agencies, it is increasingly challenging for engineers to locate relevant information quickly and accurately when solving field problems or preparing for training tasks. These limitations hinder timely decision-making and create steep learning curves for new personnel in maintenance and construction operations. To address these challenges, this paper proposes a Retrieval-Augmented Generation (RAG) framework with a multi-agent architecture to support knowledge management and decision making. The system integrates structured document retrieval with real-time, context-aware response generation powered by a large language model (LLM). Unlike conventional single-pass RAG systems, the proposed framework employs multiple specialized agents for retrieval, answer generation, evaluation, and query refinement, which enables iterative improvement and quality control. In addition, the system incorporates an open-weight vision-language model to convert technical figures into semantic textual representations, which allows figure-based knowledge to be indexed and retrieved alongside text. Retrieved text and figure-based context are then provided to an open-weight large language model, which generates the final responses grounded in the retrieved evidence.

arXiv:2603.03303v1 Announce Type: cross Abstract: Large Language Models (LLMs) are increasingly used to simulate how specific users respond to a given context, enabling more user-centric applications that rely on user feedback. However, existing user simulators mostly imitate surface-level patterns and language styles, which fail to reflect the underlying states of real users (e.g., beliefs and emotions). To address these limitations, we propose a novel training framework, HumanLM, which builds user simulators that accurately reflect real users. Our key insight is that, in addition to generating responses, the model should generate natural-language latent states that align with ground-truth responses through reinforcement learning. These latent states correspond to a set of psychologically grounded state dimensions that drive how real users respond. HumanLM further synthesizes these aligned latent states into responses that accurately represent real users. For extensive evaluation, we develop Humanual, a comprehensive benchmark for simulating real users based on public data. Humanual consists of six large-scale datasets with 26k users and 216k responses in total, spanning diverse tasks such as generating user responses to daily life issues, political blogs, and chat sessions with LLM assistants. Across datasets, HumanLM significantly outperforms alternative approaches, achieving an average relative improvement of 16.3% in alignment scores from an LLM judge. In a real-time simulation study with 111 participants, HumanLM achieves the highest similarity to real user responses and competitive human-likeness scores.

arXiv:2603.03305v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used to generate executable outputs, JSON objects, and API calls, where a single syntax error can make the output unusable. Constrained decoding enforces validity token-by-token via masking and renormalization, but it can distort generation when the model assigns low probability mass to valid continuations, pushing decoding toward locally valid yet semantically incorrect trajectories. We propose \emph{Draft-Conditioned Constrained Decoding (DCCD)}, a simple two-step, training-free inference procedure that decouples semantic planning from structural enforcement: an unconstrained draft is generated first, and constrained decoding is then applied, conditioned on this draft, to guarantee validity. We analyze DCCD through a KL-projection view, showing that draft conditioning increases feasible mass and reduces the cumulative "projection tax" induced by hard constraints, with an optional best-of-$K$ draft selection. Across structured reasoning benchmarks, DCCD improves strict structured accuracy by up to +24 percentage points over standard constrained decoding (e.g., 15.2\% to 39.0\% on GSM8K with a 1B model), and enables smaller model pairs to match or exceed much larger constrained baselines, yielding substantial gains in parameter efficiency.

arXiv:2603.03306v1 Announce Type: cross Abstract: Recently presented Token-Oriented Object Notation (TOON) aims to replace JSON as a serialization format for passing structured data to LLMs with significantly reduced token usage. While showing solid accuracy in LLM comprehension, there is a lack of tests against JSON generation. Though never present in training data, TOON syntax is simple enough to suggest one-shot in-context learning could support accurate generation. The inevitable prompt overhead can be an acceptable trade-off for shorter completions. To test this, we conducted a benchmark creating several test cases with regard to structural complexity, a validation pipeline, and comparing plain JSON generation vs structured output (via constrained decoding) JSON generation vs TOON one-shot in-context learning generation. JSON structured output was included to establish a minimum token budget baseline and to set a starting point for future experiments testing TOON constrained decoding inference enforcement. Key findings: TOON shows promising accuracy/token consumption ratio for in-domain generation tasks, though this advantage is often reduced by the "prompt tax" of instructional overhead in shorter contexts. Plain JSON generation shows the best one-shot and final accuracy, even compared with constrained decoding structured output, where the only significant advantage is the lowest token usage as a trade-off for slightly decreased accuracy overall and significant degradation for some models. Notably, for simple structures, this "lowest token usage" of constrained decoding outperformed even TOON, hinting that TOON enforcing via frameworks such as xgrammar may not yield the desired results. Furthermore, the results suggest a scaling hypothesis: TOON's true efficiency potential likely follows a non-linear curve, shining only beyond a specific point where cumulative syntax savings amortize the initial prompt overhead.

arXiv:2603.03307v1 Announce Type: cross Abstract: Epistemic Network Analysis (ENA) is a method for investigating the relational structure of concepts in text by representing co-occurring concepts as networks. Traditional ENA, however, relies heavily on manual expert coding, which limits its scalability and real-world applicability to large text corpora. Topic modeling provides an automated approach to extracting concept-level representations from text and can serve as an alternative to manual coding. To tackle this limitation, the present study merges BERTopic with ENA and introduces TopicENA, a topic-based epistemic network analysis framework. TopicENA substitutes manual concept coding with automatically generated topics while maintaining ENA's capacity for modeling structural associations among concepts. To explain the impact of modeling choices on TopicENA outcomes, three analysis cases are presented. The first case assesses the effect of topic granularity, indicating that coarse-grained topics are preferable for large datasets, whereas fine-grained topics are more effective for smaller datasets. The second case examines topic inclusion thresholds and finds that threshold values should be adjusted according to topic quality indicators to balance network consistency and interpretability. The third case tests TopicENA's scalability by applying it to a substantially larger dataset than those used in previous ENA studies. Collectively, these cases illustrate that TopicENA facilitates practical and interpretable ENA analysis at scale and offers concrete guidance for configuring topic-based ENA pipelines in large-scale text analysis.