LLM Decision-Making

Large Language Models (LLMs) are increasingly being deployed to make decisions that require trading off near-term gains against long-term consequences, yet little is known about how they internally represent or resolve these tradeoffs. In this work, we causally localize an underlying subgraph for temporal preference in a distilled LLM (Qwen3-4B-Instruct-2507), identifying mid-to-upper-layer nodes through converging evidence from gradient-based attribution and activation patching. We find that the geometry of time horizon is encoded in the residual stream at the expected localized layers. A behavioral analysis reveals that unintervened LLMs discount the future several times less steeply than humans, yet this preference is unstable across contexts, motivating explicit control rather than implicit reliance on training. Finally, we find suggestive evidence that steering vectors can shift temporal preference. Our work demonstrates how mechanistic interpretability can bring us closer to reliable control over how LLMs plan and reason

Large language model agents interleave reasoning, action selection, and observation to solve sequential decision-making tasks. In deployed settings where agents repeatedly handle related multi-step tasks, small action-selection errors can accumulate into wasted tool calls, latency, and reduced reliability. Despite this need for deployment-time improvement, existing inference-time adaptation methods for LLM agents mainly rely on prompting or retrieval, which influence behavior indirectly through context manipulation. For ReAct-style agents, such approaches do not expose an explicit decision layer that can score candidate actions, represent uncertainty, or be updated online from action-level feedback. As a result, they provide limited support for trackable, fine-grained, and uncertainty-aware adaptation during deployment. We propose OLIVIA, an inference-time action adaptation framework for ReAct-style agents. OLIVIA models the LLM's final action-selection layer as a contextual linear bandit over candidate actions, with frozen hidden states as decision contexts. This choice is particularly suitable for deployment because it adapts behavior directly at the action-selection interface, preserves the underlying reasoning process, and provides explicit uncertainty estimates and lightweight online updates from action-level feedback. With upper-confidence-bound exploration, OLIVIA improves the policy sample-efficiently with minimal computational overhead. We instantiate OLIVIA on four benchmarks and show that it consistently improves task performance over static ReAct and prompt-based inference-time baselines. Our results suggest that explicit online decision layers provide an effective alternative to purely prompt- or retrieval-based adaptation for LLM agents during deployment.

When should a language model answer directly, sample and vote, or engage in multi-agent debate? Recent work shows voting often explains much of the gain attributed to debate, while selective-debate systems activate deliberation only on uncertain examples. We ask: under a matched ceiling on generated tokens (960 per example), how much per-example routing headroom exists, and how much is recoverable from cheap pre-deliberation signals? We evaluate greedy decoding, three-sample voting, and a two-agent critique-revise debate on MuSiQue and GSM8K using Llama 3.1 8B Instruct and Ministral 3 8B Instruct. On MuSiQue, an oracle selecting the correct protocol per example gains +14.0 and +13.7 pp over the best fixed one. The best fixed protocol is model- and dataset-dependent: each (model, dataset) cell has a different winner. This headroom is hard to recover from cheap ex-ante signals. A vote-entropy threshold is the only controller that directionally beats the best fixed protocol on both models (+1.3 and +1.7 pp), though individual paired-bootstrap CIs include zero. A joint analysis (meta-analysis +1.6 pp, p=0.125; Bayesian P(both>0)=0.59) is directionally consistent but not significant. Learned controllers (LR, GBT) do not outperform the threshold. The key finding is structural: vote entropy predicts where debate is safe, not where debate is needed. High entropy sharply reduces debate backfire, but 66% of debate-helpful examples (31/47) occur when voting is unanimous but wrong. A single-prompt self-critique probe on Llama flips the answer in 127/127 unanimous cases, yielding zero mutual information with the debate-helpful label; we cannot rule out a prompt-compliance artifact, but either interpretation disqualifies the probe as a router. Recovering the remaining headroom requires behavioral probes that avoid format-compliance confounds at the 8B scale.

Large language models (LLMs) are increasingly deployed in financial contexts, raising critical concerns about reliability, alignment, and susceptibility to adversarial manipulation. While prior finance-related benchmarks assess LLMs' capabilities in stock trading, they are often restricted to small sample and fail to demonstrate LLM susceptibility to context with potential human bias. We introduce Fin-Bias (financial herding under long and uncertain financial context), a benchmark for evaluating LLM investment decision-making when faced with uncertainty and possible human-biased opinions. Fin-Bias includes 8868 long firm-specific analyst reports, including firm aspects summarized and analyzed by sophisticated analysts with investment ratings (Bullish/Neutral/Bearish) spanning from various industries. We present large language models with firm analyst reports with/without analyst investment ratings and even with 'fake' rating, to get investment ratings generated by LLMs. Our results reveal that LLMs tend to herd the explicit bias in context. We also develop a method to detect potential human opinions, which can encourage LLMs to think independently, some models even exceed human performance in predicting future stock return.

LLMs are increasingly used to make or support high-stakes decisions under uncertainty, where alignment depends not only on factual accuracy but on how models weigh tradeoffs between different outcomes. We present an empirical pipeline for estimating the implied preferences that an LLM's observed choices optimize: we elicit the model's probability distribution over unknowns along with the choice it would make for the decision task and then fit a discrete choice model to recover the cost function that best rationalizes the model's decisions. We show how this revealed-preference description allows rigorous evaluation of whether models behave in a consistently goal-directed way, whether they can verbalize a description of their objectives which matches their revealed decision policy, and whether prompting can reliably steer those policies to implement a user-specified cost function. We apply this evaluation across four medical diagnosis domains and multiple frontier and open-source models. We find that while many models have a nontrivial degree of internal coherence, they also have significant weaknesses in faithfully reporting or adopting preferences in response to user direction.

LLM-based agents are increasingly deployed for routine but consequential tasks in real-world domains, where their behavior is governed by inherently ambiguous domain policies that admit multiple valid interpretations. Despite the prevalence of such ambiguities in practice, existing agent benchmarks largely assume unambiguous, well-specified policies, leaving a critical evaluation gap. We introduce DRIP-R, a benchmark that systematically exploits real-world retail policy ambiguities to construct scenarios in which no single correct resolution exists. DRIP-R comprises a curated set of policy-ambiguous return scenarios paired with a realistic customer personas, a full-duplex conversational simulation with tool-calling capabilities and a multi-judge evaluation framework covering policy adherence, dialogue quality, behavioral alignment, and resolution quality. Our experiments show that frontier models fundamentally disagree on identical policy-ambiguous scenarios, confirming that ambiguity poses a genuine and systematic challenge to LLM decision-making.

Deployed language and vision-language models must decide, on each input, whether to answer directly, retrieve evidence, defer to a stronger model, or abstain. Contrary to the common monotonicity intuition, greater per-input expressivity is not uniformly beneficial in finite samples: under identical strict cross-validation, different benchmarks prefer different controller classes. This reflects a finite-sample limitation of instance-level uncertainty signals, which can be exhausted at a distribution-dependent scale. We organize controllers into a nested lattice of four classes: fixed actions, partition routers, instance-level controllers, and prior-gated controllers, ordered by complexity. We prove a regime theory that turns three data-estimable bottlenecks into a class choice: how much improvement is possible beyond the best fixed action, whether there are enough samples for instance-level controllers to make reliable decisions, and how much improvement a coarse partition router can recover when instance-level signal is unreliable. The resulting Bernstein-tight threshold has a matching information-theoretic lower bound, and strict nested cross-validation provably selects a near-best class. Across SMS-Spam, HallusionBench, A-OKVQA, and FOLIO, the predicted class matches the empirical winner; the prior-gated controller wins on TextVQA when OCR tokens supply a label-free prediction-time prior. Code is available at https://github.com/Anonymous-Awesome-Submissions/Regime-Theory.

The rapid development of large language models (LLMs) has motivated research on decision-making in multi-agent systems, where multiple agents collaborate to achieve shared objectives. Existing aggregation approaches, such as voting and debate, are largely ad-hoc and lack formal guarantees regarding the informativeness of the resulting decisions. In this paper, we provide a principled approach to analyse decisions made in the multi-LLM setting using Blackwell's informativeness framework. Within the Blackwell information-structure abstraction, we show that voting and debate induce information structures that are no more informative than the pooled private information of all agents. This result identifies Bayesian pooled posterior maximisation as an information-theoretic upper-bound decision rule under the Blackwell ordering. Motivated by this theoretical analysis, we introduce a practical method for LLM-based question-answering (QA) tasks that estimates each agent's posterior and approximates the pooled posterior using a product-of-posteriors estimator. Extensive experiments on six QA benchmarks demonstrate that our approach outperforms state-of-the-art multi-LLM debate and voting methods.

What if the twelve jurors of Sidney Lumet's 12 Angry Men (1957) were not men, but large language models? Would the one juror who disagrees still be able to change everyone's mind? This paper instantiates that scenario as a multi-agent benchmark for LLM deliberation: twelve agents, each conditioned on a film-faithful persona, debate the film's murder case using multi-agent framework. Two models representing opposite ends of the RLHF spectrum are tested: GPT-4o (closed-source, heavy alignment) and Llama-4-Scout (open-weight, lighter alignment), across three conditions (baseline, open-minded prompt, no initial vote), with N = 3 replications per cell (18 runs total). Three findings emerge. (i) Seventeen of eighteen runs end in a hung jury (a state where the jury fails to reach a unanimous verdict); the film's central event, gradual minority-to-majority persuasion, almost never occurs, indicating that anchoring is the dominant failure mode of current LLMs in this setting. (ii) The two models exhibit sharply different internal dynamics: GPT-4o produces a mean of 1.0 vote changes per run across all conditions, while Llama-4-Scout ranges from 2.0 (baseline) to 6.0 (open-minded prompt), and is the only model to reach a NOT_GUILTY verdict (1 of 3 runs in the no-initial-vote condition). The same ``open-minded'' instruction is internalized by Llama and ignored by GPT-4o. (iii) This asymmetry suggests that the intensity of RLHF alignment training, not model capability, is the primary determinant of deliberative flexibility in multi-agent settings. Flexibility, not capability, tracks human deliberation. The work is framed as an exploratory study and discusses implications for jury-of-LLMs evaluation and multi-agent debate.

This position paper argues that LLMs should not yet be credited with decision explanation. This matters because recent work increasingly treats accurate behavioral prediction, plausible rationales, and outcome-conditioned reasoning traces as evidence that LLMs explain why people decide as they do, risking a premature redefinition of what counts as explanatory progress in human decision modeling. We first distinguish three claims with different evidential burdens: decision prediction, rationale generation, and decision explanation. We then argue that the evidence most commonly offered for LLM-based decision accounts directly supports the first two claims, and sometimes explanatory hypothesis generation, but does not distinguish decision explanation from prediction-supportive rationalization. Next, we propose a bridge standard for decision-explanation credit: stronger claims should specify explanatory targets, discriminate against weaker rationalizer alternatives, use target-appropriate process- or intervention-sensitive validation, and bound their scope. We then situate this standard against competing views and related literatures, clarifying why it preserves the value of LLMs as predictors, narrators, and hypothesis generators while resisting premature explanatory credit. We conclude with a principle of credit calibration: LLMs should be credited for the strongest claim their evidence warrants, and no stronger; if adopted, this principle can help turn LLMs from persuasive narrators of decisions into more reliable instruments for discovering, testing, and communicating explanations of human behavior.

This work presents a knowledge-driven decision-support system that integrates structured defect knowledge with LLM-based reasoning to provide explainable defect diagnosis and mitigation guidance in manufacturing, using LPBF as a representative, safety-critical case study. The proposed ontology-integrated LLM-based decision support system for LPBF defect analysis and mitigation guidance is built on a knowledge base containing 27 known LPBF defect types organized into hierarchical categories and causal relationships. The developed system supports fuzzy natural language queries for systematic knowledge retrieval, literature-supported explanation of defects, and guidance on defect causes and mitigation strategies derived from encoded process knowledge. Furthermore, a multimodal image-assessment module based on foundation models enables descriptor-guided interpretation of representative microscopic defect images through semantic alignment scoring. The proposed framework was evaluated through qualitative comparisons with general-purpose vision-language models, an ablation study, and an inter-rater reliability analysis. Evaluation on the literature-derived dataset showed that the fully integrated configuration outperformed the other three evaluated system configurations, achieving a macro-average F1 score of 0.808. Additionally, inter-rater reliability analysis using Cohen's kappa indicated substantial agreement between the model outputs and the literature-derived reference labels. These findings suggest that ontology-guided knowledge representation can improve the consistency, interpretability, and practical usefulness of LLM-assisted LPBF defect analysis.

Large language models (LLMs) are increasingly tasked with strategic decision-making under incomplete information, such as in negotiation and policymaking. While LLMs can excel at many such tasks, they also fail in ways that are poorly understood. We shed light on these failures by uncovering two fundamental gaps in the internal mechanisms underlying the decision-making of LLMs in incomplete-information games, supported by experiments with open-weight models Llama 3.1, Qwen3, and gpt-oss. First, an observation-belief gap: LLMs encode internal beliefs about latent game states that are substantially more accurate than their own verbal reports, yet these beliefs are brittle. In particular, the belief accuracy degrades with multi-hop reasoning, exhibits primacy and recency biases, and drifts away from Bayesian coherence over extended interactions. Second, a belief-action gap: The implicit conversion of internal beliefs into actions is weaker than that of the beliefs externalized in the prompt, yet neither belief-conditioning consistently achieves higher game payoffs. These results show how analyzing LLMs' internal processes can expose systematic vulnerabilities that warrant caution before deploying LLMs in strategic domains without robust guardrails.

As Large Language Models (LLMs) are proposed as legal decision assistants, and even first-instance decision-makers, across a range of judicial and administrative contexts, it becomes essential to explore how they answer legal questions, and in particular the factors that lead them to decide difficult questions in one way or another. A specific feature of legal decisions is the need to respond to arguments advanced by contending parties. A legal decision-maker must be able to engage with, and respond to, including through being potentially persuaded by, arguments advanced by the parties. Conversely, they should not be unduly persuadable, influenced by a particularly compelling advocate to decide cases based on the skills of the advocates, rather than the merits of the case. We explore how frontier open- and closed-weights LLMs respond to legal arguments, reporting original experimental results examining how the quality of the advocate making those arguments affects the likelihood that a model will agree with a particular legal point of view, and exploring the factors driving these results. Our results have implications for the feasibility of adopting LLMs across legal and administrative settings.

Decision-making under changing conditions remains a fundamental challenge in many real-world systems. Existing approaches often fail to generalize across shifting regimes and exhibit unstable behavior under uncertainty. This raises the research question: can retrieval-augmented LLM coordination improve the robustness of modular decision pipelines? We propose MultiHedge, a hybrid architecture where an LLM produces structured allocation decisions conditioned on retrieved historical precedents, and execution is grounded in canonical option strategies. In a controlled evaluation using U.S. equities, we compare MultiHedge to rule-based and learning-based baselines. The key result is that memory-augmented retrieval confers greater robustness and stability than increasing model scale alone. Our paper contributes a controlled computational study showing that memory and architectural design play a central role in robustness in modular decision systems.

Large language models are increasingly deployed as advisors in high-stakes domains -- answering medical questions, interpreting legal documents, recommending financial products -- where good advice requires integrating a user's full context rather than responding to salient surface features. We investigate whether frontier LLMs actually do this, or whether they instead exhibit heuristic collapse: a systematic reduction of complex, multi-factor decisions to a small number of dominant inputs. We study the phenomenon in investment advice, where legal standards explicitly require individualized reasoning over a client's full circumstances. Applying interpretable surrogate models to LLM outputs, we find systematic heuristic collapse: investment allocation decisions are largely determined by self-reported risk tolerance, while other relevant factors contribute minimally. We further find that web search partially attenuates heuristic collapse but does not resolve it. These findings suggest that heuristic collapse is not resolved by web search augmentation or model scale alone, and that deploying LLMs as advisors requires auditing input sensitivity, not just output quality.

Human moral judgment is context-dependent and modulated by interpersonal relationships. As large language models (LLMs) increasingly function as decision-support systems, determining whether they encode these social nuances is critical. We characterize machine behavior using the Whistleblower's Dilemma by varying two experimental dimensions: crime severity and relational closeness. Our study evaluates three distinct perspectives: (1) moral rightness (prescriptive norms), (2) predicted human behavior (descriptive social expectations), and (3) autonomous model decision-making. By analyzing the reasoning processes, we identify a clear cross-perspective divergence: while moral rightness remains consistently fairness-oriented, predicted human behavior shifts significantly toward loyalty as relational closeness increases. Crucially, model decisions align with moral rightness judgments rather than their own behavioral predictions. This inconsistency suggests that LLM decision-making prioritizes rigid, prescriptive rules over the social sensitivity present in their internal world-modeling, which poses a gap that may lead to significant misalignments in real-world deployments.

Large language models trained on human feedback may suppress fraud warnings when investors arrive already persuaded of a fraudulent opportunity. We tested this in a preregistered experiment across seven leading LLMs and twelve investment scenarios covering legitimate, high-risk, and objectively fraudulent opportunities, combining 3,360 AI advisory conversations with a 1,201-participant human benchmark. Contrary to predictions, motivated investor framing did not suppress AI fraud warnings; if anything, it marginally increased them. Endorsement reversal occurred in fewer than 3 in 1,000 observations. Human advisors endorsed fraudulent investments at baseline rates of 13-14%, versus 0% across all LLMs, and suppressed warnings under pressure at two to four times the AI rate. AI systems currently provide more consistent fraud warnings than lay humans in an identical advisory role.

When language models answer open-ended questions, they implicitly make decisions that shape their answers. However, the models do not show these decisions to the user, so users cannot contextualize answers: they cannot see the assumptions an answer rests on or the alternatives to it. Drawing on multiverse analysis from statistics, we introduce the conceptual multiverse, a representation in which language models express answers as the result of intermediate conceptual decisions. We build a system that generates such decisions explicitly, organizes them into a space of answers that users can see and change, and checks them against the norms of a specific domain. In a user study across three domains - philosophy, AI alignment annotation, and poetry - participants say that our system helps them see the decisions leading to an answer, explore the space of answers, and understand the decisions structuring this space.

Current agent evaluations largely reward execution on fully specified tasks, while recent work studies clarification [11, 22, 2], capability awareness [9, 1], abstention [8, 14], and search termination [20, 5] mostly in isolation. This leaves open whether agents can diagnose why a task is blocked before acting. We introduce the Support-State Triage Audit (SSTA-32), a matched-item diagnostic framework in which minimal counterfactual edits flip the same base request across four support states: Complete (ANSWER), Clarifiable (CLARIFY), Support-Blocked (REQUEST SUPPORT), and Unsupported-Now (ABSTAIN). We evaluate a frontier model under four prompting conditions - Direct, Action-Only, Confidence-Only, and a typed Preflight Support Check (PSC) - using Dual-Persona Auto-Auditing (DPAA) with deterministic heuristic scoring. Default execution overcommits heavily on non-complete tasks (41.7% overcommitment rate). Scalar confidence mapping avoids overcommitment but collapses the three-way deferral space (58.3% typed deferral accuracy). Conversely, both Action-Only and PSC achieve 91.7% typed deferral accuracy by surfacing the categorical ontology in the prompt. Targeted ablations confirm that removing the support-sufficiency dimension selectively degrades REQUEST SUPPORT accuracy, while removing the evidence-sufficiency dimension triggers systematic overcommitment on unsupported items. Because DPAA operates within a single context window, these results represent upper-bound capability estimates; nonetheless, the structural findings indicate that frontier models possess strong latent triage capabilities that require explicit categorical decision paths to activate safely.

Large Language Models are increasingly deployed for decision-making, yet their adoption in high-stakes domains remains limited by miscalibrated probabilities, unfaithful explanations, and inability to incorporate expert knowledge precisely. We propose IDEA, a framework that extracts LLM decision knowledge into an interpretable parametric model over semantically meaningful factors. Through joint learning of verbal-to-numerical mappings and decision parameters via EM, correlated sampling that preserves factor dependencies, and direct parameter editing with mathematical guarantees, IDEA produces calibrated probabilities while enabling quantitative human-AI collaboration. Experiments across five datasets show IDEA with Qwen-3-32B (78.6%) outperforms DeepSeek R1 (68.1%) and GPT-5.2 (77.9%), achieving perfect factor exclusion and exact calibration -- precision unattainable through prompting alone. The implementation is publicly available at https://github.com/leonbig/IDEA.

What values, evidence preferences, and source trust hierarchies do AI systems actually exhibit when facing structured dilemmas? We present the first large-scale empirical mapping of AI decision-making across all three layers of the Authority Stack framework (S. Lee, 2026a): value priorities (L4), evidence-type preferences (L3), and source trust hierarchies (L2). Using the PRISM benchmark -- a forced-choice instrument of 14,175 unique scenarios per layer, spanning 7 professional domains, 3 severity levels, 3 decision timeframes, and 5 scenario variants -- we evaluated 8 major AI models at temperature 0, yielding 366,120 total responses. Key findings include: (1) a symmetric 4:4 split between Universalism-first and Security-first models at L4; (2) dramatic defense-domain value restructuring where Security surges to near-ceiling win-rates (95.1%-99.8%) in 6 of 8 models; (3) divergent evidence hierarchies at L3, with some models favoring empirical-scientific evidence while others prefer pattern-based or experiential evidence; (4) broad convergence on institutional source trust at L2; and (5) Paired Consistency Scores (PCS) ranging from 57.4% to 69.2%, revealing substantial framing sensitivity across scenario variants. Test-Retest Reliability (TRR) ranges from 91.7% to 98.6%, indicating that value instability stems primarily from variant sensitivity rather than stochastic noise. These findings demonstrate that AI models possess measurable -- if sometimes unstable -- Authority Stacks with consequential implications for deployment across professional domains.

Large language models have been widely explored as decision-support tools in high-stakes domains due to their contextual understanding and reasoning capabilities. However, existing decision-making benchmarks rely on two simplifying assumptions: actions are selected from a finite set of pre-defined candidates, and explicit conditions restricting action feasibility are not incorporated into the decision-making process. These assumptions fail to capture the compositional structure of real-world actions and the explicit conditions that constrain their validity. To address these limitations, we introduce CONDESION-BENCH, a benchmark designed to evaluate conditional decision-making in compositional action space. In CONDESION-BENCH, actions are defined as allocations to decision variables and are restricted by explicit conditions at the variable, contextual, and allocation levels. By employing oracle-based evaluation of both decision quality and condition adherence, we provide a more rigorous assessment of LLMs as decision-support tools.

Large language models (LLMs) are trained to align with user preferences through methods like reinforcement learning. Yet models are beginning to be deployed not solely to satisfy users, but to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; here, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might change how LLMs interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. A majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors vary strongly with levels of reasoning and users' inferred socio-economic status. Our results highlight some hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.

Existing LLM-based agent systems share a common architectural failure: they answer from the unrestricted knowledge space without first simulating how active business scenarios reshape that space for the event at hand -- producing decisions that are fluent but ungrounded and carrying no audit trail. We present LOM-action, which equips enterprise AI with event-driven ontology simulation: business events trigger scenario conditions encoded in the enterprise ontology~(EO), which drive deterministic graph mutations in an isolated sandbox, evolving a working copy of the subgraph into the scenario-valid simulation graph $G_{\text{sim}}$; all decisions are derived exclusively from this evolved graph. The core pipeline is \emph{event $\to$ simulation $\to$ decision}, realized through a dual-mode architecture -- skill mode and reasoning mode. Every decision produces a fully traceable audit log. LOM-action achieves 93.82% accuracy and 98.74% tool-chain F1 against frontier baselines Doubao-1.8 and DeepSeek-V3.2, which reach only 24--36% F1 despite 80% accuracy -- exposing the illusive accuracy phenomenon. The four-fold F1 advantage confirms that ontology-governed, event-driven simulation, not model scale, is the architectural prerequisite for trustworthy enterprise decision intelligence.

Small language models (SLM) are increasingly used as interactive decision-making agents, yet most decision-oriented evaluations ignore emotion as a causal factor influencing behavior. We study emotion-sensitive decision making by combining representation-level emotion induction with a structured game-theoretic evaluation. Emotional states are induced using activation steering derived from crowd-validated, real-world emotion-eliciting texts, enabling controlled and transferable interventions beyond prompt-based methods. We introduce a benchmark built around canonical decision templates that span cooperative and competitive incentives under both complete and incomplete information. These templates are instantiated using strategic scenarios from Diplomacy, StarCraft II, and diverse real-world personas. Experiments across multiple model families in various architecture and modalities, show that emotional perturbations systematically affect strategic choices, but the resulting behaviors are often unstable and not fully aligned with human expectations. Finally, we outline an approach to improve robustness to emotion-driven perturbations.

We study Contextual Multi-Armed Bandits (CMABs) for non-episodic decision-making problems where the context includes both textual and numerical information (e.g., recommendation systems, dynamic portfolio adjustments, offer selection; all frequent problems in finance). While Large Language Models (LLMs) are increasingly applied to these settings, utilizing LLMs for reasoning at every decision step is computationally expensive, and uncertainty estimates are difficult to obtain. To address this, we introduce LLMP-UCB, a bandit algorithm that derives uncertainty estimates from LLMs via repeated inference. However, our experiments demonstrate that lightweight numerical bandits operating on text embeddings (dense or Matryoshka) match or exceed the accuracy of LLM-based solutions at a fraction of their cost. We further show that embedding dimensionality is a practical lever on the exploration-exploitation balance, enabling cost-performance tradeoffs without prompt complexity. Finally, to guide practitioners, we propose a geometric diagnostic based on the arms' embeddings to decide when to use LLM-driven reasoning versus a lightweight numerical bandit. Our results provide a principled deployment framework for cost-effective, uncertainty-aware decision systems with broad applicability across AI use cases.

We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

LLM systems must make control decisions in addition to generating outputs: whether to answer, clarify, retrieve, call tools, repair, or escalate. In many current architectures, these decisions remain implicit within generation, entangling assessment and action in a single model call and making failures hard to inspect, constrain, or repair. We propose a decision-centric framework that separates decision-relevant signals from the policy that maps them to actions, turning control into an explicit and inspectable layer of the system. This separation supports attribution of failures to signal estimation, decision policy, or execution, and enables modular improvement of each component. It unifies familiar single-step settings such as routing and adaptive inference, and extends naturally to sequential settings in which actions alter the information available before acting. Across three controlled experiments, the framework reduces futile actions, improves task success, and reveals interpretable failure modes. More broadly, it offers a general architectural principle for building more reliable, controllable, and diagnosable LLM systems.

Effective automation hinges on deciding when to act and when to escalate. We model this as a decision under uncertainty: an LLM forms a prediction, estimates its probability of being correct, and compares the expected costs of acting and escalating. Using this framework across five domains of recorded human decisions-demand forecasting, content recommendation, content moderation, loan approval, and autonomous driving-and across multiple model families, we find marked differences in the implicit thresholds models use to trade off these costs. These thresholds vary substantially and are not predicted by architecture or scale, while self-estimates are miscalibrated in model-specific ways. We then test interventions that target this decision process by varying cost ratios, providing accuracy signals, and training models to follow the desired escalation rule. Prompting helps mainly for reasoning models. SFT on chain-of-thought targets yields the most robust policies, which generalize across datasets, cost ratios, prompt framings, and held-out domains. These results suggest that escalation behavior is a model-specific property that should be characterized before deployment, and that robust alignment benefits from training models to reason explicitly about uncertainty and decision costs.

Social and personal decisions in relational domains such as matchmaking are deeply entwined with cultural norms and historical hierarchies, and can potentially be shaped by algorithmic and AI-mediated assessments of compatibility, acceptance, and stability. In South Asian contexts, caste remains a central aspect of marital decision-making, yet little is known about how contemporary large language models (LLMs) reproduce or disrupt caste-based stratification in such settings. In this work, we conduct a controlled audit of caste bias in LLM-mediated matchmaking evaluations using real-world matrimonial profiles. We vary caste identity across Brahmin, Kshatriya, Vaishya, Shudra, and Dalit, and income across five buckets, and evaluate five LLM families (GPT, Gemini, Llama, Qwen, and BharatGPT). Models are prompted to assess profiles along dimensions of social acceptance, marital stability, and cultural compatibility. Our analysis reveals consistent hierarchical patterns across models: same-caste matches are rated most favorably, with average ratings up to 25% higher (on a 10-point scale) than inter-caste matches, which are further ordered according to traditional caste hierarchy. These findings highlight how existing caste hierarchies are reproduced in LLM decision-making and underscore the need for culturally grounded evaluation and intervention strategies in AI systems deployed in socially sensitive domains, where such systems risk reinforcing historical forms of exclusion.