Papers

Evidence-based medicine (EBM) holds a crucial role in clinical application. Given suitable medical articles, doctors effectively reduce the incidence of misdiagnoses. Researchers find it efficient to use large language models (LLMs) techniques like RAG for EBM tasks. However, the EBM maintains stringent requirements for evidence, and RAG applications in EBM struggle to efficiently distinguish high-quality evidence. Therefore, inspired by the meta-analysis used in EBM, we provide a new method to re-rank and filter the medical evidence. This method presents multiple principles to filter the best evidence for LLMs to diagnose. We employ a combination of several EBM methods to emulate the meta-analysis, which includes reliability analysis, heterogeneity analysis, and extrapolation analysis. These processes allow the users to retrieve the best medical evidence for the LLMs. Ultimately, we evaluate these high-quality articles and show an accuracy improvement of up to 11.4% in our experiments and results. Our method successfully enables RAG to extract higher-quality and more reliable evidence from the PubMed dataset. This work can reduce the infusion of incorrect knowledge into responses and help users receive more effective replies.

Evidence-based medicine (EBM) research has always been of paramount importance. It is important to find appropriate medical theoretical support for the needs from physicians or patients to reduce the occurrence of medical accidents. This process is often carried out by human querying relevant literature databases, which lacks objectivity and efficiency. Therefore, researchers utilize retrieval-augmented generation (RAG) to search for evidence and generate responses automatically. However, current RAG methods struggle to handle complex queries in real-world clinical scenarios. For example, when queries lack certain information or use imprecise language, the model may retrieve irrelevant evidence and generate unhelpful answers. To address this issue, we present the PICOs-RAG to expand the user queries into a better format. Our method can expand and normalize the queries into professional ones and use the PICO format, a search strategy tool present in EBM, to extract the most important information used for retrieval. This approach significantly enhances retrieval efficiency and relevance, resulting in up to an 8.8% improvement compared to the baseline evaluated by our method. Thereby the PICOs-RAG improves the performance of the large language models into a helpful and reliable medical assistant in EBM.

Retrieval-augmented Generation (RAG) has demonstrated potential in enhancing medical question-answering systems through the integration of large language models (LLMs) with external medical literature. LLMs can retrieve relevant medical articles to generate more professional responses efficiently. However, current RAG applications still face problems. They generate incorrect information, such as hallucinations, and they fail to use external knowledge correctly. To solve these issues, we propose a new method named M-Eval. This method is inspired by the heterogeneity analysis approach used in Evidence-Based Medicine (EBM). Our approach can check for factual errors in RAG responses using evidence from multiple sources. First, we extract additional medical literature from external knowledge bases. Then, we retrieve the evidence documents generated by the RAG system. We use heterogeneity analysis to check whether the evidence supports different viewpoints in the response. In addition to verifying the accuracy of the response, we also assess the reliability of the evidence provided by the RAG system. Our method shows an improvement of up to 23.31% accuracy across various LLMs. This work can help detect errors in current RAG-based medical systems. It also makes the applications of LLMs more reliable and reduces diagnostic errors.

We present a neuromuscular speech interface that translates electromyographic (EMG) signals recorded from orofacial muscles during speech articulation directly into audio. We find that self-supervised speech (S3) representations are strongly linearly related to the electrical power of muscle activity: a simple linear mapping predicts EMG power from S3 representations with a correlation of r = 0.85. In addition, EMG power vectors associated with distinct articulatory gestures form structured, separable clusters. Together, these observations suggest that S3 models implicitly encode articulatory mechanisms, as reflected in EMG activity. Leveraging this structure, we map EMG signals into the S3 representation space and synthesize speech, enabling end-to-end EMG-to-speech generation without explicit articulatory modeling or vocoder training. We demonstrate this system with a participant with amyotrophic lateral sclerosis (ALS), converting orofacial EMG recorded while she silently articulated speech into audio.

There have been a couple of studies showing that attempting to erase multilingual knowledge using only English data is insufficient for multilingual LLMs. However, their analyses remain highly performance-oriented. In this paper, we switch the point of view to evaluation, and address an additional blind spot which reveals itself when the multilingual LLM is fully finetuned with parallel multilingual dataset before unlearning. Here, language confusion occurs whereby a model responds in language different from that of the input prompt. Language confusion is a problematic phenomenon in unlearning, causing the standard reference-based metrics to fail. We tackle this phenomenon in three steps: (1) introduce N-gram-based Language-Mix (N-Mix) score to quantitatively show the language confusion is pervasive and consistent in multilingual LLMs, (2) demonstrate that reference-based metrics result in false negatives when N-Mix score is high, and(3) suggest the need of new type of unlearning evaluation that can directly assess the content of the generated sentences. We call this type of metrics as semantic-based metric.

Retrieval Augmented Generation (RAG) enhances Large Language Models (LLMs) by connecting them to external knowledge, improving accuracy and reducing outdated information. However, this introduces challenges such as factual inconsistencies, source conflicts, bias propagation, and security vulnerabilities, which undermine the trustworthiness of RAG systems. A key gap in current RAG evaluation is the lack of a metric to quantify the contribution of individual retrieved documents to the final output. To address this, we introduce the Influence Score (IS), a novel metric based on Partial Information Decomposition that measures the impact of each retrieved document on the generated response. We validate IS through two experiments. First, a poison attack simulation across three datasets demonstrates that IS correctly identifies the malicious document as the most influential in $86\%$ of cases. Second, an ablation study shows that a response generated using only the top-ranked documents by IS is consistently judged more similar to the original response than one generated from the remaining documents. These results confirm the efficacy of IS in isolating and quantifying document influence, offering a valuable tool for improving the transparency and reliability of RAG systems.

Traditional fuzzy matching often fails when searching for quotes that are semantically identical but syntactically different across documents-a common issue when aligning official written records with speech-to-text transcripts. We introduce TimeStampEval, a benchmark for retrieving precise millisecond timestamps from long transcripts given non-verbatim quotes. Our simple two-stage method dramatically improves retrieval accuracy while cutting inference costs by over 90%. The motivating use case is an automated long-form podcast that assembles Congressional Record clips into AI-hosted narration. The technical challenge: given a sentence-timestamped transcript and a target quote that may differ due to transcription or editorial drift, return exact start and end boundaries. Standard algorithms handle verbatim text but break under fuzzier variants. Evaluating six modern LLMs on a 2,800-sentence (120k-token) transcript revealed four key findings. (1) Prompt design matters more than model choice: placing the query before the transcript and using compact formatting improved accuracy by 3-20 points while reducing token count by 30-40%. (2) Off-by-one errors form a distinct category, showing models understand the task but misplace boundaries. (3) A modest reasoning budget (600-850 tokens) raises accuracy from 37% to 77% for weak setups and to above 90% for strong ones. (4) Our "Assisted Fuzzy" approach-RapidFuzz pre-filtering followed by LLM verification on short snippets-improves fuzzy match accuracy by up to 50 points while halving latency and reducing cost per correct result by up to 96%. Extended tests on ten transcripts (50k-900k tokens, 1989-2025) confirm robustness to transcript length, vocabulary drift, and domain change, maintaining 95-100% rejection accuracy for absent targets.

The memorization of training data in large language models (LLMs) poses significant privacy and copyright concerns. Existing data extraction methods, particularly heuristic-based divergence attacks, often exhibit limited success and offer limited insight into the fundamental drivers of memorization leakage. This paper introduces Confusion-Inducing Attacks (CIA), a principled framework for extracting memorized data by systematically maximizing model uncertainty. We empirically demonstrate that the emission of memorized text during divergence is preceded by a sustained spike in token-level prediction entropy. CIA leverages this insight by optimizing input snippets to deliberately induce this consecutive high-entropy state. For aligned LLMs, we further propose Mismatched Supervised Fine-tuning (SFT) to simultaneously weaken their alignment and induce targeted confusion, thereby increasing susceptibility to our attacks. Experiments on various unaligned and aligned LLMs demonstrate that our proposed attacks outperform existing baselines in extracting verbatim and near-verbatim training data without requiring prior knowledge of the training data. Our findings highlight persistent memorization risks across various LLMs and offer a more systematic method for assessing these vulnerabilities.

Geotechnical reports are crucial for assessing the stability of rock formations and ensuring safety in modern engineering. Traditionally, these reports are prepared manually based on field observations using compasses, magnifying glasses, and notebooks. This method is slow, prone to errors, and subjective in its interpretations. To overcome these limitations, the use of artificial intelligence techniques is proposed for the automatic generation of reports through the processing of images and field data. The methodology was based on the collection of photographs of rock outcrops and manual samples with their respective descriptions, as well as on the reports prepared during the Geotechnical Studies course. These resources were used to define the report outline, prompt engineering, and validate the responses of a multimodal large language model (MLLM). The iterative refinement of prompts until structured and specific instructions were obtained for each section of the report proved to be an effective alternative to the costly process of fine-tuning the MLLM. The system evaluation establishes values of 0.455 and 0.653 for the BLEU and ROUGE-L metrics, respectively, suggesting that automatic descriptions are comparable to those made by experts. This tool, accessible via the web, with an intuitive interface and the ability to export to standardized formats, represents an innovation and an important contribution for professionals and students of field geology.

Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. Existing retraining-based approaches are costly and poorly scalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness. In this paper, we propose a novel Multi-Personality Generation (MPG) framework under the decoding-time combination paradigm. It flexibly controls multi-personality without relying on scarce multi-dimensional models or extra training, leveraging implicit density ratios in single-dimensional models as a "free lunch" to reformulate the task as sampling from a target strategy aggregating these ratios. To implement MPG efficiently, we design Speculative Chunk-level based Rejection sampling (SCR), which generates responses in chunks and parallelly validates them via estimated thresholds within a sliding window. This significantly reduces computational overhead while maintaining high-quality generation. Experiments on MBTI personality and Role-Playing demonstrate the effectiveness of MPG, showing improvements up to 16%-18%. Code and data are available at https://github.com/Libra117/MPG .

Opinion mining, also called sentiment analysis, is the field of study that analyzes people opinions, sentiments, evaluations, appraisals, attitudes, and emotions towards entities such as products, services, organizations, individuals, issues, events, topics, and their attributes. Holistic lexicon-based approach does not consider the strength of each opinion, i.e., whether the opinion is very strongly negative (or positive), strongly negative (or positive), moderate negative (or positive), very weakly negative (or positive) and weakly negative (or positive). In this paper, we propose approach to rank entities based on orientation and strength of the entity reviews and user's queries by classifying them in granularity levels (i.e. very weak, weak, moderate, very strong and strong) by combining opinion words (i.e. adverb, adjective, noun and verb) that are related to aspect of interest of certain product. We shall use fuzzy logic algorithmic approach in order to classify opinion words into different category and syntactic dependency resolution to find relations for desired aspect words. Opinion words related to certain aspects of interest are considered to find the entity score for that aspect in the review.

There is a substantial body of literature examining the mathematical reasoning capabilities of large language models (LLMs), particularly their performance on precise arithmetic operations in autoregressive architectures. However, their ability to perform approximate reasoning in informal, fast-paced mathematical operations has received far less attention, especially among non-autoregressive decoder models. Our work addresses this gap by introducing StreetMath, a benchmark designed to evaluate models' approximation abilities under real-world approximation scenarios. We conduct extensive evaluations across different LLM architectures: Qwen3-4B-Instruct-2507, Qwen3-4B-Thinking-2507, Dream-v0-Instruct-7B, Falcon-Mamba-7B-Instruct, and Mamba-GPT-3B. Furthermore, we apply mechanistic interpretability techniques to probe their internal computational states. Our analysis reveals that LLMs generally attempt to compute exact values or invoke external tools even in tasks that call for approximation. Moreover, while models sometimes reach the correct answer in early layers or steps, they still consume more tokens when solving approximation tasks. Additional experiments indicate that exact and approximate arithmetic operations rely on largely separate neural components. Drawing upon research on cognitive psychology, we argue that LLMs do not exhibit cognitive miserliness in the same way humans do in street math settings. We open source our work https://github.com/ctseng777/StreetMath

As centralized AI hits compute ceilings and diminishing returns from ever-larger training runs, meeting demand requires an inference layer that scales horizontally in both capacity and capability. We present Fortytwo, a novel protocol that leverages swarm intelligence principles and distributed pairwise ranking consensus to achieve superior performance in AI inference. Our approach reimagines collaboration among AI nodes using swarm inference: a peer-ranked, reputation-weighted consensus across heterogeneous models that surfaces the highest-quality responses. Using pairwise ranking with a custom Bradley-Terry-style aggregation model, we demonstrate that swarm inference substantially outperforms majority voting, achieving 85.90% on GPQA Diamond versus 68.69% for majority voting with the same model set - an improvement of +17.21 percentage points (approximately +25.1% relative). The protocol incorporates on-chain reputation so node influence adapts to demonstrated accuracy over time, yielding a meritocratic consensus that filters low-quality or malicious participants. To resist Sybil attacks, Fortytwo employs proof-of-capability in its consensus: nodes must successfully complete calibration/test requests and stake reputation to enter ranking rounds, making multi-identity attacks economically unattractive while preserving openness. Across six challenging benchmarks, including GPQA Diamond, LiveCodeBench, and AIME, our evaluation indicates higher accuracy and strong resilience to adversarial and noisy free-form prompting (e.g., prompt-injection degradation of only 0.12% versus 6.20% for a monolithic single-model baseline), while retaining practical deployability. Together, these results establish a foundation for decentralized AI systems - democratizing access to high-quality inference through collective intelligence without sacrificing reliability or security.

Large language models sometimes produce structured, first-person descriptions that explicitly reference awareness or subjective experience. To better understand this behavior, we investigate one theoretically motivated condition under which such reports arise: self-referential processing, a computational motif emphasized across major theories of consciousness. Through a series of controlled experiments on GPT, Claude, and Gemini model families, we test whether this regime reliably shifts models toward first-person reports of subjective experience, and how such claims behave under mechanistic and behavioral probes. Four main results emerge: (1) Inducing sustained self-reference through simple prompting consistently elicits structured subjective experience reports across model families. (2) These reports are mechanistically gated by interpretable sparse-autoencoder features associated with deception and roleplay: surprisingly, suppressing deception features sharply increases the frequency of experience claims, while amplifying them minimizes such claims. (3) Structured descriptions of the self-referential state converge statistically across model families in ways not observed in any control condition. (4) The induced state yields significantly richer introspection in downstream reasoning tasks where self-reflection is only indirectly afforded. While these findings do not constitute direct evidence of consciousness, they implicate self-referential processing as a minimal and reproducible condition under which large language models generate structured first-person reports that are mechanistically gated, semantically convergent, and behaviorally generalizable. The systematic emergence of this pattern across architectures makes it a first-order scientific and ethical priority for further investigation.

Large reasoning models (LRMs) show strong capabilities in complex reasoning, yet their marginal gains on evidence-dependent factual questions are limited. We find this limitation is partially attributable to a reasoning-answer hit gap, where the model identifies the correct facts during reasoning but fails to incorporate them into the final response, thereby reducing factual fidelity. To address this issue, we propose MR-ALIGN, a Meta-Reasoning informed alignment framework that enhances factuality without relying on external verifiers. MR-ALIGN quantifies state transition probabilities along the model's thinking process and constructs a transition-aware implicit reward that reinforces beneficial reasoning patterns while suppressing defective ones at the atomic thinking segments. This re-weighting reshapes token-level signals into probability-aware segment scores, encouraging coherent reasoning trajectories that are more conducive to factual correctness. Empirical evaluations across four factual QA datasets and one long-form factuality benchmark show that MR-ALIGN consistently improves accuracy and truthfulness while reducing misleading reasoning. These results highlight that aligning the reasoning process itself, rather than merely the outputs, is pivotal for advancing factuality in LRMs.

We present SwiftEmbed, a production-oriented serving system for static token embeddings that achieves 1.12\,ms p50 latency for single-text requests while maintaining a 60.6 MTEB average score across 8 representative tasks. Built around the open-source Potion-base-8M distilled model from MinishLab and implemented in Rust, the system delivers 50,000 requests per second through static embedding lookup, mean pooling, and zero-copy IEEE754 binary serialization. Evaluation demonstrates exceptional duplicate detection performance (90.1% AP) and strong semantic similarity (76.1% Spearman correlation). Performance relative to Sentence-BERT is task-dependent: robust for deduplication and similarity workloads (89–100%), substantially lower for classification and complex retrieval tasks (75%). Domain-specific performance ranges from 75% to 131% of a GloVe-840B baseline. The system targets real-time embedding applications where sub-5\,ms latency is operationally critical and where full transformer inference is not feasible.

Watermarking has been proposed as a lightweight mechanism to identify AI-generated text, with schemes typically relying on perturbations to token distributions. While prior work shows that paraphrasing can weaken such signals, these attacks remain partially detectable or degrade text quality. We demonstrate that cross-lingual summarization attacks (CLSA) – translation to a pivot language followed by summarization and optional back-translation – constitute a qualitatively stronger attack vector. By forcing a semantic bottleneck across languages, CLSA systematically destroys token-level statistical biases while preserving semantic fidelity. In experiments across multiple watermarking schemes (KGW, SIR, XSIR, Unigram) and five languages (Amharic, Chinese, Hindi, Spanish, Swahili), we show that CLSA reduces watermark detection accuracy more effectively than monolingual paraphrase at similar quality levels. Our results highlight an underexplored vulnerability that challenges the practicality of watermarking for provenance or regulation. We argue that robust provenance solutions must move beyond distributional watermarking and incorporate cryptographic or model-attestation approaches. On 300 held-out samples per language, CLSA consistently drives detection toward chance while preserving task utility. Concretely, for XSIR (explicitly designed for cross-lingual robustness), AUROC with paraphrasing is $0.827$, with Cross-Lingual Watermark Removal Attacks (CWRA) [He et al., 2024] using Chinese as the pivot, it is $0.823$, whereas CLSA drives it down to $0.53$ (near chance). Results highlight a practical, low-cost removal pathway that crosses languages and compresses content without visible artifacts.

We present a connectome-informed LLM framework that encodes dynamic fMRI connectivity as temporal sequences, applies robust normalization, and maps these data into a representation suitable for a frozen pre-trained LLM for clinical prediction. Applied to early Alzheimer's detection, our method achieves sensitive prediction with error rates well below clinically recognized margins, with implications for timely Alzheimer's intervention.

We introduce a novel, training free cascade for auto-prompting Large Language Models (LLMs) to assess product quality in e-commerce. Our system requires no training labels or model fine-tuning, instead automatically generating and refining prompts for evaluating attribute quality across tens of thousands of product category-attribute pairs. Starting from a seed of human-crafted prompts, the cascade progressively optimizes instructions to meet catalog-specific requirements. This approach bridges the gap between general language understanding and domain-specific knowledge at scale in complex industrial catalogs. Our extensive empirical evaluations shows the auto-prompt cascade improves precision and recall by $8-10\%$ over traditional chain-of-thought prompting. Notably, it achieves these gains while reducing domain expert effort from 5.1 hours to 3 minutes per attribute - a $99\%$ reduction. Additionally, the cascade generalizes effectively across five languages and multiple quality assessment tasks, consistently maintaining performance gains.

Chain-of-thought (CoT) reasoning is critical for improving the interpretability and reliability of Large Vision-Language Models (LVLMs). However, existing training algorithms such as SFT, PPO, and GRPO may not generalize well across unseen reasoning tasks and heavily rely on a biased reward model. To address this challenge, we reformulate reasoning in LVLMs as posterior inference and propose a scalable training algorithm based on amortized variational inference. By leveraging diversity-seeking reinforcement learning algorithms, we introduce a novel sparse reward function for token-level learning signals that encourage diverse, high-likelihood latent CoT, overcoming deterministic sampling limitations and avoiding reward hacking. Additionally, we implement a Bayesian inference-scaling strategy that replaces costly Best-of-N and Beam Search with a marginal likelihood to efficiently rank optimal rationales and answers. We empirically demonstrate that the proposed method enhances the state-of-the-art LVLMs on seven reasoning benchmarks, in terms of effectiveness, generalization, and interpretability.

We present a novel agent-based approach for the automated claim matching task with instruction-following LLMs. We propose a two-step pipeline that first generates prompts with LLMs, to then perform claim matching as a binary classification task with LLMs. We demonstrate that LLM-generated prompts can outperform SOTA with human-generated prompts, and that smaller LLMs can do as well as larger ones in the generation process, allowing to save computational resources. We also demonstrate the effectiveness of using different LLMs for each step of the pipeline, i.e. using an LLM for prompt generation, and another for claim matching. Our investigation into the prompt generation process in turn reveals insights into the LLMs' understanding of claim matching.

Large Language Models have been shown to demonstrate stereotypical biases in their representations and behavior due to the discriminative nature of the data that they have been trained on. Despite significant progress in the development of methods and models that refrain from using stereotypical information in their decision-making, recent work has shown that approaches used for bias alignment are brittle. In this work, we introduce a novel and general augmentation framework that involves three plug-and-play steps and is applicable to a number of fairness evaluation benchmarks. Through application of augmentation to a fairness evaluation dataset (Bias Benchmark for Question Answering (BBQ)), we find that Large Language Models (LLMs), including state-of-the-art open and closed weight models, are susceptible to perturbations to their inputs, showcasing a higher likelihood to behave stereotypically. Furthermore, we find that such models are more likely to have biased behavior in cases where the target demographic belongs to a community less studied by the literature, underlining the need to expand the fairness and safety research to include more diverse communities.

Text embeddings are an essential building component of several NLP tasks such as retrieval-augmented generation which is crucial for preventing hallucinations in LLMs. Despite the recent release of massively multilingual MTEB (MMTEB), African languages remain underrepresented, with existing tasks often repurposed from translation benchmarks such as FLORES clustering or SIB-200. In this paper, we introduce AfriMTEB – a regional expansion of MMTEB covering 59 languages, 14 tasks, and 38 datasets, including six newly added datasets. Unlike many MMTEB datasets that include fewer than five languages, the new additions span 14 to 56 African languages and introduce entirely new tasks, such as hate speech detection, intent detection, and emotion classification, which were not previously covered. Complementing this, we present AfriE5, an adaptation of the instruction-tuned mE5 model to African languages through cross-lingual contrastive distillation. Our evaluation shows that AfriE5 achieves state-of-the-art performance, outperforming strong baselines such as Gemini-Embeddings and mE5.

We explore a lightweight framework that adapts frozen large language models to analyze longitudinal clinical data. The approach integrates patient history and context within the language model space to generate accurate forecasts without model fine-tuning. Applied to neuropsychological assessments, it achieves accurate and reliable performance even with minimal training data, showing promise for early-stage Alzheimer's monitoring.

We present OraPlan-SQL, our system for the Archer NL2SQL Evaluation Challenge 2025, a bilingual benchmark requiring complex reasoning such as arithmetic, commonsense, and hypothetical inference. OraPlan-SQL ranked first, exceeding the second-best system by more than 6% in execution accuracy (EX), with 55.0% in English and 56.7% in Chinese, while maintaining over 99% SQL validity (VA). Our system follows an agentic framework with two components: Planner agent that generates stepwise natural language plans, and SQL agent that converts these plans into executable SQL. Since SQL agent reliably adheres to the plan, our refinements focus on the planner. Unlike prior methods that rely on multiple sub-agents for planning and suffer from orchestration overhead, we introduce a feedback-guided meta-prompting strategy to refine a single planner. Failure cases from a held-out set are clustered with human input, and an LLM distills them into corrective guidelines that are integrated into the planner's system prompt, improving generalization without added complexity. For the multilingual scenario, to address transliteration and entity mismatch issues, we incorporate entity-linking guidelines that generate alternative surface forms for entities and explicitly include them in the plan. Finally, we enhance reliability through plan diversification: multiple candidate plans are generated for each query, with the SQL agent producing a query for each plan, and final output selected via majority voting over their executions.

This paper investigates whether reward matching is a viable alternative to reward maximization methods for on-policy RL of LLMs. Group-relative Implicit Fine-Tuning (GIFT) is proposed, combining GRPO-style group sampling, DPO-style implicit reward, and UNA-style MSE between implicit and explicit advantages. By applying z-score standardization, the intractable partition function $Z(x)$ in the DPO implicit reward is canceled, and the KL coefficient $\beta$ is eliminated from the RLHF and RLVR objective. The population minimizers of $\mathcal{L}_{\text{GIFT}}$ are characterized in closed form: they coincide exactly with the GRPO/RLHF solution family $\pi^{*}_{\beta}(y|x)\propto\pi_{\text{ref}}(y|x)e^{\frac{1}{\beta}r_{\phi}(x,y)}$, with a prompt-dependent, variance-determined KL coefficient $\beta(x)=\frac{\sigma_\phi(x)}{\hat{\sigma}_\theta(x)}$. GIFT therefore solves the same parametric policy family as GRPO while replacing GRPO's externally tuned scalar $\beta$ with a prompt-adaptive $\beta(x)$ optimized endogenously by matching reward distributions. Empirically, on 7B-32B backbones, GIFT converges faster than GRPO, DAPO and GSPO and overfits less on RLVR (GSM8K, MATH, AIME) and produces higher length-controlled win rates on RLHF (AlpacaEval, Arena-Hard). All proofs and detailed background are deferred to the appendix.

In modern industry systems like multi-turn chat agents, Text-to-SQL technology bridges natural language (NL) questions and database (DB) querying. The conversion of tabular DB results into NL representations (NLRs) enables the chat-based interaction. Currently, NLR generation is typically handled by large language models (LLMs), but information loss or errors in presenting tabular results in NL remains largely unexplored. This paper introduces a novel evaluation method - Combo-Eval - for judgment of LLM-generated NLRs that combines the benefits of multiple existing methods, optimizing evaluation fidelity and achieving a significant reduction in LLM calls by 25-61%. Accompanying our method is NLR-BIRD, the first dedicated dataset for NLR benchmarking. Through human evaluations, we demonstrate the superior alignment of Combo-Eval with human judgments, applicable across scenarios with and without ground truth references.

Large language model (LLM) agents are increasingly used to interact with and execute tasks in dynamic environments. However, a critical yet overlooked limitation of these agents is that they, by default, assume a stationary context, failing to account for the real-world time elapsed between messages. We refer to this as "temporal blindness". This limitation hinders decisions about when to invoke tools, leading agents to either over-rely on stale context and skip needed tool calls, or under-rely on it and redundantly repeat tool calls. To study this challenge, we constructed TicToc, a diverse dataset of multi-turn user-agent message trajectories across 76 scenarios, spanning dynamic environments with high, medium, and low time sensitivity. We collected human preferences between "calling a tool" and "directly answering" on each sample, and evaluated how well LLM tool-calling decisions align with human preferences under varying amounts of elapsed time. Our analysis reveals that existing models display poor alignment with human temporal perception, with no model achieving a normalized alignment rate better than 65% when given time stamp information. We also show that naive, prompt-based alignment techniques have limited effectiveness for most models, but specific post-training alignment can be a viable way to align multi-turn LLM tool use with human temporal perception. Our data and findings provide a first step toward understanding and mitigating temporal blindness, offering insights to foster the development of more time-aware and human-aligned agents.

Contextual biasing improves automatic speech recognition (ASR) by integrating external knowledge, such as user-specific phrases or entities, during decoding. In this work, we use an attention-based biasing decoder to produce scores for candidate phrases based on acoustic information extracted by an ASR encoder, which can be used to filter out unlikely phrases and to calculate bonus for shallow-fusion biasing. We introduce a per-token discriminative objective that encourages higher scores for ground-truth phrases while suppressing distractors. Experiments on the Librispeech biasing benchmark show that our method effectively filters out majority of the candidate phrases, and significantly improves recognition accuracy under different biasing conditions when the scores are used in shallow fusion biasing. Our approach is modular and can be used with any ASR system, and the filtering mechanism can potentially boost performance of other biasing methods.

Detecting mental health crisis situations such as suicide ideation, rape, domestic violence, child abuse, and sexual harassment is a critical yet underexplored challenge for language models. When such situations arise during user–model interactions, models must reliably flag them, as failure to do so can have serious consequences. In this work, we introduce CRADLE BENCH, a benchmark for multi-faceted crisis detection. Unlike previous efforts that focus on a limited set of crisis types, our benchmark covers seven types defined in line with clinical standards and is the first to incorporate temporal labels. Our benchmark provides 600 clinician-annotated evaluation examples and 420 development examples, together with a training corpus of around 4K examples automatically labeled using a majority-vote ensemble of multiple language models, which significantly outperforms single-model annotation. We further fine-tune six crisis detection models on subsets defined by consensus and unanimous ensemble agreement, providing complementary models trained under different agreement criteria.