Effective travel planning requires handling multiple information sources, managing constraints, and tailoring recommendations to user preferences. While LLMs have introduced dynamic, context-aware decision-making, single-agent models often struggle with complex multi-step reasoning, constraint adherence, and modular execution. To overcome these limitations, we propose a hierarchical multi-agent framework where a supervising chatbot agent coordinates specialized agents for schedule validation, user query formulation, real-time data retrieval, and constraint-aware itinerary generation. By integrating LLM reasoning, external APIs, and iterative refinement mechanisms based on a deterministic, multi-dimensional scoring framework, our approach ensures adaptability, efficiency, and personalization. Evaluations on the well-known TravelPlanner benchmark demonstrate improvements in Budget Constraint planning by 10\%, wider constraints set-based planning by 25\%, and personalized planning results by 45\%. compared to the ReAct baseline. Our code is available at https://github.com/nimashaklaa/TravelPlanner_Research
Mobile agents on personal smartphones can affect private data, payments, settings, and the foreground screen. We present MATE (Mobile Agent Trustworthy Execution), an execution-layer system that controls the boundary between agent action proposals and device effects without modifying the underlying model. MATE represents each UI primitive as a structured action record that combines the instruction, target app, screen context, primitive, and action fields. Runtime Guardrail (RG) maps this record to an allow, block, or user-approval decision; Predefined Task Routing (PTR) serves recurring requests through verified routes with post-execution screen checks; and a floating interface keeps execution in a compact overlay. On MobileSafetyBench, MATE raises high-risk blocking from 29.7% to 97.0% under GPT-4o. On AndroidWorld, it reduces LLM calls by up to 6x and raises recurring-use task success to 95.7%.
We present DRVEXO-4B-Unified, a modular multimodal assistant built around a 4B language backbone. A self-contained checkpoint combines text, image, and speech inputs with reference-conditioned speech output and structured interaction. A temporal residual connector adapts acoustic sequences to the language interface; explicit text boundaries support incremental speech synthesis; and three task-conditioned scoring heads rank supplied action candidates alongside generative tool calls. These interfaces connect conversation to application-defined choices while leaving action authorization to the application. The system obtains 74.22% on all 900 MMMU validation questions, exceeding the published validation scores of Gemma 3 4B IT and InternVL3.5-4B without test-time scaling. Strict prompt-level IFEval accuracy is 82.26% and MATH-500 accuracy is 89.40%, above the original Qwen3-4B non-thinking report's 81.2% and 84.8%, respectively. On NVIDIA GB10, the model achieves a speech-synthesis first-chunk latency of 390 ms. For concurrent text generation, a growth-aware admission policy completes a 128-answer batch at 594.41 output tokens/s, including filling and draining, without preemptions. We plan to release the model weights and accompanying code under Apache License 2.0, retaining applicable third-party terms.
A voice agent that waits for a fixed silence before replying either interrupts users who pause mid-utterance or responds late to users who have finished. This study examines the decision at the onset of silence, using turn ends and turn-internal pauses extracted from the Real-TurnTurk corpus of Turkish dyadic video calls. A fine-tuned Turkish TurnGPT model, a zero-shot Turkish language model and a deployed multilingual detector are compared, together with voice-activity timing and facial features, under recording-disjoint and speaker-disjoint evaluation with paired cluster bootstrap tests. On held-out recordings, the fine-tuned model separates turn ends from pauses better than both text baselines and is the best calibrated of the three. Timing features of the current speaker are similarly informative, and their combination with the fine-tuned text score gives the best timing-based results; with logistic regression, this combination also outperforms the same combination with either text baseline. In a latency simulation, adding the text score to a silence timeout shortens the mean response delay at the same interruption rate. On automatic Whisper transcripts, the fine-tuned model retains its advantage over the zero-shot model.
Recent work reports strong pragmatic performance by large language models (LLMs) in English, yet whether this performance generalises across languages and evaluation settings remains unclear. This question is increasingly important as LLMs are deployed across languages, cultures, and conversational contexts. In this paper, we first introduce PILAPS, a bilingual English/Mandarin Chinese benchmark covering eight pragmatic phenomena across purpose-written and dialogue-derived items, with both culturally adapted and independently sourced data. We then evaluate eight LLMs from four families across a range of model sizes and prompting strategies. Our results reveal substantial variation across models and pragmatic phenomena: open-source models continue to struggle particularly with humour, while cross-linguistic performance varies with model origin. Our error analysis shows that, compared with human participants, models are more prone to literal interpretations in conversational contexts. Theory-of-Mind prompting lowers the literal share of errors, whereas Chain-of-Thought prompting slightly raises it. Together, these findings demonstrate that strong pragmatic performance in one evaluation setting does not necessarily generalise across languages or conversational contexts.
Large language models demonstrate strong medical question-answering capabilities, yet reasoning performance remains uneven across languages. Process reward models (PRMs) provide step-level feedback for improving reasoning, but a fixed PRM may become increasingly misaligned with the trajectory distribution of an evolving policy during reinforcement learning, thereby reducing the reliability of its supervision. We propose a policy--reward alternating training framework that jointly adapts a medical PRM and optimizes a multilingual reasoning policy. Following supervised initialization, the current policy generates multilingual trajectories to update the PRM, which then provides process-level rewards for the next policy update. Medical step supervision maintains the PRM's medical foundation, while process, outcome, and language-related rewards guide policy learning. We develop multilingual medical PRMs (PAIR-PRM) and reasoning models (PAIR-MED) at 8B and 9B scales across 12 languages. On the medical subsets of Global-MMLU and MMLU-ProX, our models improve average accuracy by 3.01--7.47% over their corresponding Qwen backbones. On CUREMED-BENCH, PAIR-MED-9B improves the joint language consistency and logical accuracy metric by 12.88% and 7.45% over baselines for in-domain and out-of-domain languages. Further analyses examine PRM adaptation, multilingual reasoning quality, and the contributions of individual training components.
KV cache compression is critical for large reasoning models, where most of the cost comes from long generated reasoning traces rather than the input prompt. These traces often contain redundant and low-utility tokens due to looping, self-verification, and corrections. However, such tokens may still receive high attention scores, making it difficult to remove them without harming the reasoning process. We propose RoCoKV, a KV cache compression method designed for large reasoning models. RoCoKV introduces two key components. First, uncertainty-guided adaptive pruning adjusts the compression strength based on the model's confidence, allocating more memory during uncertain phases and applying stronger pruning when the model is confident. Second, redundancy-aware token selection reduces the impact of tokens that are highly similar to recent reasoning content, encouraging the model to retain more informative and diverse tokens. Beyond the algorithm, we co-design RoCoKV with the serving stack: we develop fused three-phase CUDA decode kernels and integrate them into vLLM as a drop-in plugin that preserves its paged KV layout and CUDA graph capture, turning algorithmic savings into real end-to-end speedups. Extensive experiments across three reasoning benchmarks and two reasoning models demonstrate that RoCoKV consistently improves reasoning accuracy under constrained KV cache budgets, achieving higher accuracy compared to state-of-the-art KV cache compression methods. For DeepSeek-R1-Distill-Llama-8B in vLLM, RoCoKV achieves a 1.27X speedup and 38% higher throughput on AIME25, the most challenging evaluated reasoning benchmark, and a 1.43X speedup and 45% higher throughput on the long synthetic benchmark, compared with Full-KV.
Agent skills package procedural knowledge as natural-language instructions, and a skill helps only when its instructions are correct and properly followed by the agent. Self-evolving agents revise their own skills from the runs they make, but existing methods judge each run as a whole, by its outcome or by where it went wrong, and cannot tell whether a failure comes from a wrong instruction or from one the agent did not follow. Instruction-level coverage was introduced to test skills and labels each instruction of a run as follow, not-follow, or not covered, which separates these causes. We leverage it for skill self-evolution with SkillPlumber, a framework in which an agent evolves its own skills where coverage shows they fall short. A persistent root cause analysis agent turns each label into a revision question and checks each revision against the next run, while a maintainer distills investigations across tasks into a shared strategy library. Across three benchmarks and three executors, SkillPlumber solves 64.7% to 70.5% of the tasks, 25.6 to 30.4 percentage points higher than the initial skill and 8.7 to 14.0 percentage points higher than the strongest baseline, and removing coverage reduces recovered failures by 20.8% on SkillsBench with GPT-5.5.
Long-horizon agents rely on stored records to recover a person's state as facts change. Reliable answers depend on how facts are recorded, retrieved, and presented to the reader. We treat a long-form persona dossier as a memory representation of a known latent state, an invariant identity plus a timeline of facts, routines, and decision rules tied to life phases. We present ChronoDossier, a benchmark of 311 synthetic dossiers of 1k-10k words written by seven generators from dated timelines. Intrinsic metrics evaluate the records against their intended states; extrinsic metrics test recovery by five readers on a seeded 32-persona GPT-5.1 panel balanced across lengths (520 rated turns per condition). The linked states, records, evidence packets, and answers support tests of construction and evidence delivery. In a contextual-stress control, tag-scoped writer inputs remove detected out-of-scope anchor copying. Frozen questions authored on staged dossiers transfer less reliably to one-pass dossiers of the same states. In matched evidence controls, whole sections repair 29 of 32 GPT-5.1 and 25 of 30 Sonnet 4.6 failures labeled as packaging, compared with 2 of 8 and 1 of 8 answer-formation failures. Attribution uses separate key-based and key-free support measures, with model disagreements retained. Length remains confounded with persona and question composition; the observed between-bucket widening depends on excerpt construction. The benchmark supports offline evaluation against known facts. Memory acquisition from interaction and validation against real users remain outside its scope.
Argumentation frameworks (AFs) provide explicit, formal mechanisms for representing and resolving conflicts and thus have been incorporated into large language model (LLM) decision-making. The benefits vary and depend on how arguments and their relations are constructed, evaluated, and incorporated into the decision process. {We abstract a workflow consisting of four stages based on existing argumentation-augmented methods} and investigate their design choices over tasks of debate-based decision-making and retrieval-based fact verification. Our evaluation with recent LLMs shows that existing methods do not consistently improve performance across the considered tasks. Motivated by this observation, we develop two task-specific methods. The first is called Qnt-Arg, which organizes arguments around the original speakers in debate materials and aggregates model-estimated argument strengths and candidate-answer priors under gradual semantics. The second is Qlt-Arg that uses credibility estimates to establish preferences between conflicting evidence items and applies grounded semantics to select acceptable evidence for subsequent LLM inference. We examine the roles of generative and decision models within the two methods. In the debate setting, Qnt-Arg consistently outperforms other argumentation-augmented methods across matched configurations; Qlt-Arg improves retrieval-based fact-verification performance compared with the baselines. These findings demonstrate the value of combining formal argumentation and LLM while highlighting the importance of aligning with tasks in argumentation construction, model selection, and the role of formal semantic reasoning.
Large language model (LLM) agents can now autonomously conduct AI research by writing code and running experiments. They typically do so by iteratively optimizing a codebase, while ignoring the problem structure behind the task. This formulation is effective on simpler tasks, but on difficult problems it often gets stuck in the wrong solution region because the agent never explicitly represents its understanding of the problem. We introduce Pro$^2$ Agent, which shifts the primary state of autonomous AI research from a codebase state to a Problem State: an explicit structure of what problems prevent success and which solutions are likely to address them. Pro$^2$ Agent initializes the Problem State through comprehensive analysis of available resources and constraints. It then performs targeted research to relate the identified problems to candidate solutions. Through controlled hypothesis testing, the agent updates the importance of a suspected problem or the benefit brought by a candidate solution. On the complete MLE-Bench, Pro$^2$ Agent matches the best reported medal rate (70.7%) under the standard 24-hour budget and reaches 76.0% with a 48-hour budget, the highest medal rate reported on the complete benchmark. It also achieves competitive performance on PostTrainBench. In addition, it produces research notes for each task that make the research process transparent and reusable by human researchers.
Language models are increasingly used to support human judgement in complex reasoning tasks, yet their predictions often remain difficult to interpret and validate. Explainability techniques provide insight into model decision-making, helping expose errors, failure modes, and potential bias so users can decide whether to rely on model outputs. Explanations are simplified approximations of a language model's internal workings. The goal is to obtain explanations that reflect the underlying computation as closely as possible (faithfulness), are understandable to users (plausibility), and help them achieve their intended goals (utility). Evaluation protocols are essential for assessing these dimensions. This thesis studies highlight explanations, natural language explanations, and chain-of-thought reasoning. These forms offer complementary perspectives on model behaviour and exhibit distinct strengths and limitations across tasks. Producing explanations that satisfy multiple quality dimensions remains challenging, as does evaluating them reliably across forms and settings. This thesis aims to improve people's understanding of models' decision-making processes in complex reasoning tasks. First, it proposes two evaluation frameworks for assessing highlight explanations across multiple properties and explanation types, together with a meta-evaluation framework for examining the reliability of existing faithfulness metrics for chain-of-thought explanations. Second, it proposes two techniques that generate natural language explanations designed to be faithful to the model's prediction process while remaining concise, coherent, and helpful for understanding model decisions. Finally, a study of human-AI interaction explores the roles of evidence and explanations in helping users validate model verdicts in a fact-checking setting.
Zero-shot large language models (LLMs) are increasingly used as text classifiers, but their labels are only useful when we know to what extent we can trust them. We compare black-box methods (verbalised confidence and sampling) with white-box methods that read token probabilities (P(True), and multiple logit-scoring methods using constrained generation). We evaluate eleven open-weight LLMs (1B-32B parameters) on five classification datasets with 3 to 151 labels, measuring classification performance (accuracy, F1-macro), discrimination (AUROC), and calibration (ECE and Brier score). The choice of confidence method rarely impacts accuracy significantly, but logit-based scoring methods lead to much better discrimination (AUROC 0.74-0.77) than black-box methods (AUROC 0.57-0.66). Replacing labels with single-token identifiers preserves discrimination but costs 3-6 accuracy points on average. We also introduce decision-only scoring, which only scores positions where labels diverge in the prefix tree (trie). It reduces decode steps by 88-96% on the three largest label sets and performs comparably to full-label scoring. The best open-weight model-method configurations (Gemma 4 31B and Qwen 3.5 9B) are competitive with the commercial Jev 1.13.0 classifier.
Post-hoc rewriting is used to reduce demographic disparities in LLM-generated text when the source model is inaccessible. A rewrite can look successful merely because it improved its own target or satisfied its selector. We study this as a transfer problem on 660 gender pairs and 1,980 age contrasts of demographic-conditioned climate messages. A rewriter proposes edits, a selection model chooses among constraint-satisfying candidates, and an independent evaluator uninvolved in either rates pairwise framing disparity. A fixed policy that always targets the message with the lower Persuasion Bias Index (PBI) transfers inconsistently across axes and outcome aggregations. An adaptive policy that proposes edits on both sides but accepts at most one significantly improves on that rule, applied to the same candidates, in seven of eight comparisons across two evaluator families, mostly by editing more pairs, not by guided direction choice, at higher preservation risk. A policy editing both messages gives the largest point reductions on fewer pairs. These two policies reduce disparity across a preservation-threshold sweep; absolute ratings vary across evaluator families and prompts. Shrinking the PBI gap increases dimension-mean disparity under both evaluator families. Mitigation should be reported as transfer across objectives, evaluators and their prompts, and editing policies.
The correspondence of Heinrich Bullinger (1504-1575), Swiss Reformer and Antistes of Zurich, documents the political and religious debates of the Reformation, including extensive appeals to the Church Fathers as authoritative sources. Detecting such references computationally remains largely unexplored. We investigate dense and sparse retrieval for patristic reference detection, and metadata-informed topic modelling as a proxy for the topics and ideas discussed. Prior work on Latin intertextuality has fine-tuned dense retrievers for their respective target domain, and we test how well these models transfer to Bullinger's correspondence. We compare lexical (BM25), dense (bi-encoders), and hybrid methods on explicit and implicit references. A Latin encoder fine-tuned on biblical reuse performs best among the bi-encoders (82%, Recall@20 on explicit and 16% on implicit references), while BM25 outperforms dense approaches on explicit citations (94%, R@20). Because BM25 and dense retrieval recover disjoint sets of implicit references, fusing their rankings raises implicit R@20 to 22%, close to cross-encoder reranking of the same candidates (26%) but without its inference cost. For topic modelling, we extend FASTopic with a seed-guided loss function and editorial abstract embeddings. Measurable coherence gains over the baseline remain minimal, but qualitative and temporal analysis shows that some topics track documented historical events. However, dominant topic weights barely exceed a uniform distribution, making topics unusable to inform retrieval. A pilot application of LLM-based concept induction (LLooM) further points towards a more human-interpretable alternative. Together, these results present a first approach to retrieving patristic references and modelling fine-grained topics in Bullinger's correspondence.
Scientific discovery requires revisiting the hypotheses in light of new evidence and identifying promising directions beyond those already explored. However, research agents often make claims that are not supported by their observations, explore a narrow hypothesis space, and retain beliefs despite recognizing conflicting evidence. These failures motivate the search for an explicit epistemic state that captures both beliefs about existing hypotheses and directions for further exploration. To this end, we propose Epistemic Compass (EpiCompass), a plug-and-play scaffold that maintains this state through an epistemic tree linking hypotheses, evidence, beliefs, and revisions. As an investigation progresses, EpiCompass reassesses the hypotheses explored against accumulated evidence and uses unresolved conflicts and unexplored directions to guide hypothesis generation. It selects candidates by anticipated epistemic gain, accounting for both refinement of existing hypotheses and expansion of the hypothesis space, and verifies the final conclusion against the accumulated evidence. Experimental results across three scientific discovery benchmarks and three LLM backbones demonstrate that EpiCompass consistently outperforms baselines, with further analyses showing reductions in errors of unsupported claims and failures to revise beliefs.
Long event streams require identifying each change and applying the changes in order. We train a selector on a frozen language model to read each transition. With API-only access, we ask the model to list the transitions. An external integer register composes them in order. A confidence rule can flag uncertain updates or retain several candidates. We study products of permutations of five items, a task that fixed-depth, log-precision transformers cannot solve on all input lengths under the standard conjecture that $\mathrm{TC}^0 \ne \mathrm{NC}^1$. On two-event windows, a 27B model answered 3 of 100 correctly without visible reasoning and 96 of 100 when it generated intermediate steps. A frozen 0.8B model with a trained selector read every event in 299 of 299 windows of 43,687 events each. The same package passed 298 of 299 new windows on a fresh machine, missing the target of 0.99 for the one-sided 95% lower confidence bound. A separate follow-up sized to allow one failure yielded 472 of 473. On 1,000-event room-swap documents, three of ten models solved all five documents by reasoning alone. All seven models tested both ways at this length solved all five with extraction and the register, with no chain of thought requested. With concurrent extraction, median wall-clock time fell from 136.1 to 10.4 s for Sonnet 5.5 and from 313.5 to 28.0 s for GPT-6.1 Sol. At fixed per-token rates, mean cost per document fell by 55% for Sonnet 5.5 and 44% for GPT-6.1 Sol. A regular expression with an executor also solved all 15 fixed-format documents. On 500 free-form issue-tracker comments, a separately adapted 4B reader identified the recorded value in 68.8% of cases, compared with 61.0% for a frequency baseline fitted on the same training projects.
LLM judges serve as rewards where no answer key exists, in the hope that their knowledge passes to the trained model. In 814 RL runs that teach 7-8B models to answer closed-book questions or abstain, we ask what passes. What a grader knows about single answers stays home. A wrong answer that a judge rewards is memorised for that one question. A model trained with the answer key, or with a far stronger judge, chooses which new questions to answer no better than the untrained model does by sampling 32 answers and abstaining where they disagree; the grader's knowledge adds about one corrected answer per hundred questions. What a grader believes about a kind of question travels. A judge that rejects eleven of the twelve correct training answers about capitals of former states teaches the model to refuse 98% of such capitals on new questions, a loss that accuracy and abstention rates do not reveal. Changing only those twelve verdicts by hand moves the refusal from 5% to 90%, even for a grader that accepts as many answers as the answer key; at that leniency, two of four judges already cost 38-42% of these capitals. The capitals are not forgotten: with "I don't know" blocked, the model answers 94% of them correctly. The grader therefore needs no knowledge of its own: rewarding the model for giving the answer that most of its own untrained samples agree on, with no labels and no judge, earns the answer key's reward and keeps the category.
Can the allocation used to train experts also predict which expert will perform best on a prompt? Such a connection would make the intended division of labor inspectable before training and allow the allocation itself to guide inference. We introduce InflEns, a mixture of LoRA experts whose training weights and inference routes come from the same location game in a fixed prompt-trait space. Players choose positions to maximize their expected shares of the prompt distribution, without using expert losses. We utilize the game's kernel-general theoretical criterion for when co-located positions become locally unstable to separation. We evaluate InflEns on a corpus assembled from seven safety-related benchmarks, which demonstrates that the highest-share expert has the lowest held-out response negative log-likelihood (NLL) on approximately 90% of prompts across four kernels under a sharpened simplex representation, compared with approximately 15% after permuting allocations across prompts. We demonstrate that the method performs comparably to published learned gates without having to train an explicit gate on the experts. Because the router is fixed, we can trace harmful behavior to a single expert and retrain only that expert, which lowers HarmBench attack success from 0.51 to 0.02 while keeping instruction following. The game also gives us several controls over how the experts specialize, including the kernel choice and its parameters, the trait-space geometry, and the number of experts. Importantly, InflEns is a general framework that can be extended to any encoder and data distribution.
Reinforcement learning (RL) can only teach a language model what its reward can recognize. Answer checkers are exact but need a reference answer; learned reward models and LLM judges work everywhere, but their scores are not probabilities. We study RLCDR (Reinforcement Learning from Calibrated Decision Rewards). An LLM converted into a decision model answers a typed question about a response ("is the final answer correct?", "which step is the first error?") with a probability distribution, read in one forward pass, debiased over option orders and calibrated with about a hundred labels. Training Qwen3-1.7B with GRPO on MATH, this reference-free reward raises MATH500 accuracy from 72.2% to 76.8% (+4.65 points, p < 1e-4), matching training on the answer key (76.9%), while a random reward gains only +0.65. The gain grows with difficulty, from 44.8% to 51.7% on the hardest level. On the policy's own changing samples the reward stays calibrated: its expected calibration error is half that of a softmax judge on the same 27B backbone (0.074 vs. 0.141), and it shows no sign of reward hacking. Without step training, the same model finds the first wrong step better than a dedicated 7B process reward model (ProcessBench F1 0.772 vs. 0.722, 20/20 splits). We also map where the approach stops being enough: trained only on problems the base model mostly fails, the reference-free reward still improves the policy (+1.9 MATH500 points, p = 0.03) but trails the answer key, because its probabilities become overconfident on that distribution. We state six hypotheses in advance and test each with a control designed to falsify it.