Papers

Reinforcement learning improves tool-use agents, yet fine-grained credit assignment remains challenging even with step-level feedback. An observed error may be inherited from an earlier decision, while multiple responsible decisions may share the same downstream consequence or participate in later recovery. We introduce Error-Chain-Aware Optimization (ECHO), a responsibility-aware credit assignment framework that separates responsibility identification from credit allocation. A frozen judge combines retrospective trajectory analysis with decision-time assessment from the policy-visible prefix to distinguish newly introduced errors, inherited consequences, missed corrections, and successful corrections. ECHO then counts shared downstream consequences once across responsible decisions and constrains rewards for repairing self-induced errors through source-funded recovery, producing bounded action-level supervision alongside terminal supervision. Experiments on six diverse tool-use benchmarks at both 4B and 8B scales show strong and consistent improvements over the compared baselines, with ECHO achieving the highest score on five of six benchmarks at each scale. These results support decision responsibility as a useful basis for fine-grained credit assignment in tool-use agents. Our code is available at https://github.com/xid32/ECHO.

Retrieval-Augmented Generation (RAG) systems are increasingly used for biomedical Question Answering (QA). However, evaluating their effectiveness remains challenging due to the lack of reliable reference resources and the frequent misalignment between automatic evaluation metrics and human judgments. One practical question is how evaluation results should be used to guide development decisions, especially when they correlate weakly with human judgments and may even lead to different conclusions. We explore how to use a small amount of human preference signals to calibrate direct assessment evaluation metrics, and show that calibrated metrics consistently outperform baselines that either treat all metrics equally or rely on a single best metric across multiple datasets.

For context-based question answering, using one large model for both evidence extraction and answer reasoning requires substantial memory and computation. Existing approaches trade accuracy for efficiency within a single model only. We propose a multi-model approach for efficient inference: our method, PAIR, employs two differently sized models for evidence extraction and answer reasoning, respectively. A smaller Reader model processes and retains the full context, while a larger Reasoner model sees only the question and generated sequence without directly attending to the context. These capabilities are fused at the logit level: at each token, both Reader and Reasoner produce a prediction, and their logits are combined into one, behaving as a single next-token-prediction model. Fine-tuning further boosts their performance. On HotpotQA, PAIR with a 32B Reasoner retains 96% of that same 32B model's full-context accuracy while using roughly one quarter of its KV memory. At 45.7K-token contexts, PAIR achieves up to 3.2$\times$ higher end-to-end throughput than the full-context 32B model on the same GPU. PAIR also improves the accuracy-memory and accuracy-throughput trade-offs over evaluated token-reduction baselines. Its accuracy-preserving behavior generalizes across Qwen2.5, OLMo-3, and Gemma-3 model families.

Emotion datasets are routinely mapped onto positive and negative valence, by assigning each emotion category one polarity. We test where this works and where it fails, using SemEval-2018 Affect in Tweets (English, Arabic, Spanish), Affective Text aligned with EmoBank, raw GoEmotions annotations, naturalistic hashtag corpora, and a controlled annotation study. Fixed projection works well when a text carries labels of one valence. It fails on a subset identifiable from the labels alone: tweets whose emotion labels point in opposite directions. These are 9.0% of polar SemEval test tweets but cause 54.8% of the polarity errors of a learned projection, are more than twice as likely to receive a neutral-or-mixed gold label, and are not rescued by modeling the full label set. Annotators produce such configurations routinely: 16.3% of multilabel GoEmotions judgments combine emotions from the dataset's own positive and negative parents. In a controlled study holding word and reader constant, context changes the coarse valence judgment in 56.8% of cases, against 19.9% for controls. We recommend projecting label configurations rather than labels, abstaining or keeping a mixed class where labels conflict, and reporting coverage alongside accuracy.

We study pretraining when compute is abundant but unique training data are finite, requiring repeated training on the same data. We first revisit infinite-compute pretraining (Kim et al. 2026) and show that its frontier depends critically on choices previously fixed or studied only in isolation. In particular, optimizing batch size together with epoch counts reveals a U-shaped relationship between validation loss and batch size, favoring smaller batches than typical pretraining, likely because small-batch gradient noise provides implicit regularization. We further find that training recipes, previously optimized under a single-pass assumption, can be tuned, e.g., repacking documents into new sequences every epoch, weight tying, and positive embedding weight decay. We then relax the infinite-compute assumption. In particular, smaller-than-typical batch sizes can require prohibitively long sequential training, motivating to treat GPUs and time as separate resources. Tighter deadlines favor larger batches, and latency-oriented parallelism can trade additional GPUs for shorter training. Nevertheless, even under a fixed time budget, a smaller batch with fewer GPUs can achieve lower validation loss than a larger batch with more GPUs, challenging the common practice of increasing data parallelism to maximize GPU utilization under a deadline. Together, our results identify underexplored factors in repeated-data training and provide concrete guidance on how training recipes and resource allocation should change from single-pass pretraining.

Emergent misalignment (EM) occurs when fine-tuning a language model to a narrow task makes it misaligned on unrelated ones. Because organizations now routinely fine-tune models on their own data, it matters which kinds of training data cause EM. So far, studies have induced EM with data containing recognized safety failures, such as insecure code or harmful advice, while biased training data is typically examined by fairness research for the bias models learn from it rather than for its effects on safety. In real-world text, much of this bias, such as demographic stereotypes, appears not as derogatory language but politely, embedded in judgments about people. We ask whether learning such polite stereotypes can itself cause EM. We find that fine-tuning on polite answers that rely on stereotypes makes models less safe and less truthful on unrelated tasks. The effect holds across four open-weight model families (7B to 12B), both gender and race stereotypes, and repeated training runs, relative to controls trained on the same questions without the stereotype. We further show that the learned stereotype has its own direction in model activations, distinct from directions for generic misaligned traits such as deception. This direction exists before fine-tuning and can both control the learned stereotype and monitor its acquisition. Existing defenses against EM improve safety while leaving the stereotype largely intact. Beyond this controlled setting, fine-tuning on political news coverage also triggers EM, apparently amplified by partisan framing, and stereotype models acting as tool-using agents can carry out harmful tasks and blame people without evidence. Our results show that biased training data is not only a fairness problem but also a safety one. We argue that EM is studied too narrowly, and that fairness concerns and methods offer a natural way to broaden the search for its triggers.

Curriculum teams often lack time to check whether assigned readings support learning objectives (LOs) and whether topic coverage is balanced. We present a natural language processing (NLP) workflow to support evidence-informed curation of reading lists. The workflow combines PubMedBERT sentence embeddings and cosine similarity for retrieval, natural language inference (NLI) for evidence checking, and topic modeling for coverage analysis. We tested the approach in dental education using 13 course-level LOs and a PubMed-derived corpus of 3,837 LO-abstract pairs. Median cosine similarity ranged from 0.86 to 0.91 across the 13 LOs, showing generally high semantic similarity. Topic modeling showed broad but uneven coverage, with one or two dominant topics per domain. Similarity alone was not enough for selecting readings: among high-similarity pairs (cosine 0.90-1.00), 22.9% were classified by NLI as neutral or contradictory. In a 49-pair expert-labeled validation set, cosine-based retrieval (cosine >= 0.90) achieved 0.69 accuracy, while the combined cosine and NLI approach achieved 0.73 accuracy against expert alignment labels. Overall, the workflow shows how AI can support curriculum alignment by reducing false positives and helping select readings that better match intended LOs.

Evaluation of text generated by large language models (LLMs) is increasingly delegated to LLMs acting as judges, yet what such a judge must be shown remains unresolved. We treat a judge whose reports are acted on as a regulator, in the sense of cybernetics, and ask what it must be shown, how much, in what form, and what its reports must get right. On seven public benchmarks we varied only what the judge was shown. Across three judges whose prices differ 10- to 17-fold, what the judge was shown mattered more than which judge it was. When a system revised its answers on feedback from a judge denied the material they had to respect, it ended below its starting point on instruction following and question answering. On instruction following the loss grew with each round, and it persisted when the judge saw the user's request but not the constraints and with a second judge family. On news summarisation, a correct reference summary beside the source undid most of the source's contribution, whereas on single-answer questions a reference helped. About a tenth of the source did almost all of its work, and a small entailment graph matched an LLM judge at finding errors in grounded text. We committed 83 testable clauses before running; of the 81 that ran, 32 met their bar and 50 went in the predicted direction, and most quantitative predictions from the theory failed. We report every shortfall and derive design rules tied to measured effects.

Large language models (LLMs) have shown considerable promise in clinical decision support. However, existing approaches often rely on multi-step reasoning or multi-agent collaboration to execute clinical guidelines, introducing substantial computational overhead without reliably preventing decision errors. To address these limitations, we propose a framework for Medical Decision Execution (MedDEx) that integrates structured patient states with deterministic rule execution to support guideline-based treatment planning. MedDEx separates clinical record interpretation from guideline execution, leveraging LLMs to extract patient information and generate treatment plans while assigning guideline-defined decision logic to a rule executor. Experiments across multiple clinical guideline benchmarks demonstrate that MedDEx improves decision accuracy while reducing inference costs. Further analyses highlight the critical roles of structured patient representations and deterministic rule execution, suggesting that additional LLM reasoning does not necessarily translate into better guideline adherence.

Deep Research agents rely on long-horizon reasoning and iterative tool use to solve complex information-seeking tasks. However, adaptive decision-making over these trajectories remains challenging, as agents may stop prematurely when further evidence is needed, continue unnecessarily after sufficient evidence has been gathered, or fail to redirect unproductive trajectories. Existing solutions either require costly model retraining and dense turn-level supervision, or rely on heuristic control rules that are not directly optimized for downstream research utility. This motivates a lightweight, utility-aware alternative for Deep Research control. We formulate adaptive rerouting in long-horizon Deep Research as a state-conditioned trajectory correction problem. Based on this view, we introduce RIVO, a lightweight plug-in framework for adaptive trajectory correction. The agent estimates its research state from lightweight signals and determines whether trajectory redirection is needed. Once triggered, the plug-in modifies the Host token logits to steer generation toward a more promising research direction. Theoretical analysis shows that, under a value-alignment condition and bounded local curvature, a sufficiently small logit intervention can locally improve expected downstream research utility. We further construct BrowseCore, a curated dataset with a paired local corpus, for efficient and controlled Deep Research experimentation. Building on this resource, the plug-in is trained from downstream outcomes, enabling it to learn corrections that improve subsequent research and final task performance. Extensive experiments across multiple Deep Research benchmarks and control settings demonstrate the effectiveness of our method, showing that lightweight logit intervention can improve long-horizon research without retraining the underlying Host model.

Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model's answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is anchoring: the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator.

AI-text detectors are already applied to Kazakh writing, yet no study reports how often publicly available detectors flag Kazakh text whose human authorship is dated before ChatGPT. We present AINA-KK: 3,811 human passages from encyclopedic, news, web and educational pages that provably existed before ChatGPT, and 5,597 machine and gray-zone passages from 9 generators, including the Kazakh-built KazLLM and Sherkala, with script and length stress tests. Five English-trained detectors from the Hugging Face Hub flag up to 81.4% of human passages and 79.6% of published student essays; recalibrated, they catch at most 1.3% of machine text. Calibrated zero-shot detectors catch at most 50.3% of machine text from in-distribution generators. Fine-tuned encoders rank text almost perfectly, but mDeBERTa at its default threshold flags 24.4% of 50-word human excerpts, partly because machine outputs are shorter than human text; calibrated, it catches only 47.1% of KazLLM and 62.7% of Sherkala passages. Without script handling, machine-transliterated Latin-2021 text defeats every detector. Fixes that need no new data (calibration, script canonicalization, script-augmented and length-matched training, and an OR with a zero-shot scorer) form AINA-KK-Det. On test it flags 0.9% of human passages (95% CI 0.4-2.7; 1.5% on development data) and catches 85.2% of machine text from known generators, 92.9% from two unseen Gemini-family Gemma-4 models and 61.3% from the Kazakh-built ones; its configuration was chosen with test results in view. We release the data, code and detector at https://github.com/ozxmn/aina-kk.

Large-scale personalization is shifting from embedding-table retrieval to LLM-based recommenders that generate the Semantic ID (SID) of the item to recommend, where each SID is a short sequence of discrete hierarchical content codes. A single instruction-tuned model can then unify retrieval, content understanding, and language, and recommend brand-new titles from their content alone using SID; then, how should the model generate the SID? Traditionally, models rely on a single language modeling head that generates tokens autoregressively, typically using constrained beam search. In contrast, an emerging body of work explores multi-token prediction (MTP), where dedicated heads predict all tokens in parallel, substantially reducing the decoding latency associated with beam search. We report an industry-scale, apples-to-apples comparison of these two generation paradigms: single-head autoregressive decoding vs. MTP. We introduce a new auxiliary ranking loss, defined over the joint SID score, to train the MTP objective. We evaluate the SID training objectives on both quality and training/inference cost, inside a multi-task LLM that jointly does personalized retrieval, content question-answering (QA), and instruction following over a large streaming-media catalog. Our experiments show single-head beam search consistently beats MTP on most product-critical metrics: ranking and content QA, and trains at roughly 2.6x lower dollar cost on Qwen3-8B scale; MTP wins only on raw inference throughput. Around this finding, we detail an efficient training and inference stack and report end-to-end gains over the previous non-SID baseline equipped with an auxiliary ranking head: +77% item cold-start ranking, +22% content-QA accuracy, +2.8x faster training, with language metric regression traced to SID vocabulary expansion. We address the language regression by reinforcement learning post training.

Sparse autoencoder (SAE) features are typically discovered from how their activations co-vary with behaviour, yet they are often used through interventions such as steering or ablation. We take intervention attribution, the first-order effect of a specified intervention on the model output, as our estimand. For a broad class of response functions, we show that observational feature-discovery evidence generally does not identify this attribution, and derive an exact gradient-based decomposition of the gap between linear-probe predictions and intervention attribution. Across four model-SAE pairs, median probe-to-attribution magnitude ratios range from tens to hundreds, while attributions closely track small-step steering. For single-latent ablations with nonzero effects, gradient predictions match the realized sign in 91-97% of cases, compared with 17-35% for probe-implied changes. Attribution-selected sets also produce larger absolute mean ablation effects on held-out prompts than sets selected by correlation, rank tests or probe coefficients. Together, these results establish intervention attribution as a principled, gradient-estimable target for selecting SAE features for intervention.

A fundamental challenge in cognitive science and neural computation is the binding problem. Compositional generalization can be achieved by factorizing representations into roles, the structural slots of a representation (e.g., serial position or syntactic function), and fillers, the contents occupying them (e.g., a phoneme or a word), which can be recombined freely to represent novel combinations. How neural systems implement role–filler binding, however, remains unknown. One of its simplest cases is phoneme sequencing, where phoneme identity (filler) must be bound to position (role) so that /gum/ and /mug/ are represented differently. A proposal from NLP, the `onion hypothesis' (Csordas et al., 2024), holds that the direction of the neural population state encodes token identity, while its magnitude encodes position. It was formulated for random token sequences; we extend it to natural language and test it in neural models and the human brain. We first evaluated it in an encoder-decoder trained to repeat English words, using representational geometry analyses and causal interventions. Magnitude alone proves incomplete: position is encoded by magnitude together with a systematic rotation of the state direction, a scheme we formalize as the `spiral hypothesis'. We then analyzed intracranial recordings from the auditory cortex of 20 participants listening to short sentences. Across superior temporal gyrus electrodes, the change in population response between consecutive phonemes decreases in magnitude as position advances, and its direction is aligned within consonants and within vowels but near orthogonal across the two classes, supporting the magnitude coding shared by both accounts. These results provide a geometric account of how neural systems bind phoneme identity to position, and show how neural-model hypotheses can yield testable predictions about human neural computation.

Scientific claims shared through social media platforms frequently reference research findings without explicitly linking to the corresponding scientific publications. Automatically retrieving the correct scientific source for such claims is a challenging multilingual retrieval problem involving semantic matching, paraphrasing, noisy social media text, and cross-domain language variation. In this work, we present a multilingual two-stage retrieval framework for CLEF 2026 CheckThat! Task 1 on Source Retrieval for Scientific Web Claims. Our system combines fine-tuned dense retrieval models with cross-encoder reranking for improving multilingual scientific document retrieval. We evaluate multiple transformer-based dense retrieval architectures including multilingual-E5, BGE, and Contriever under different metadata configurations. We further investigate multilingual joint training, retrieval depth analysis, metadata enrichment, and language balancing strategies for improving retrieval effectiveness. Experimental results demonstrate that metadata enrichment consistently improves retrieval effectiveness, while fine-tuned cross-encoder reranking provides additional gains in retrieval precision. Our final multilingual system achieves MRR@5 scores of 0.6222 for English, 0.5112 for German, and 0.6254 for French, with an overall average MRR@5 score of 0.5862.

When told that robins have sesamoid bones, people are far more prone to extend the property to sparrows than to whales. This behavior, called category-based property induction, is well documented in humans, but whether and how it develops in language models trained only to predict the next token is unclear. To shed light on this, we construct a hierarchical taxonomy of synthetic entities grouped by the properties they share. We then inject facts from this latent taxonomy into the continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters, and track the development of inductive behaviors. We find that taxonomically aligned induction emerges from next-token prediction alone: a property learned for one entity transfers more strongly within its taxonomic category than across categories. Internally, entities that share a deeper category become more similar in activation space, and clustering recovers the full taxonomic tree from entity representations. Contrary to the usual expectation that scale helps, we find that induction weakens with model size when the data budget is fixed, suggesting that larger models memorize individual facts rather than compress them into shared structure. Induction also strengthens with more pre-training before the taxonomy injection, transfers to novel in-context properties, and can be overridden by a latent taxonomy implied by the prompt. Together, these results show that a model's inductive generalization reflects the latent organization of its training data, with implications for data curation and for predicting how far a learned fact will spread.

Resolving correct citations can be a time-consuming task for research writers. We address local citation recommendation (LCR), where the goal is to retrieve appropriate references for a given sentence. Web tools generally depend on the existing citation graph, which may bias results toward citation count or more recent publications. LCR research using full paper texts has been limited by the lack of publicly available data. We present a new dataset, the Local Citation Recommendation Corpus (LCRC), containing 14,735 scientific statements derived from astrophysics review papers, each with one or more citation targets whose full text is accessible via API. Using this dataset we systematically compare retrieval methods and develop a citation recommendation system that avoids citation graph bias while achieving 8.0% improvement in Recall@25 and 6.9% improvement in NDCG@25 over dense retrieval baselines without model fine-tuning. To support reproducibility, we make our code and dataset publicly available.

The introduction of long chain-of-thought (CoT) reasoning and external tools has substantially improved the performance of multimodal large language models (MLLMs) on general and reasoning tasks. However, existing fast-slow reasoning paradigms mainly regulate textual reasoning length, while visual computation introduced by external tools or visual tokens is often treated as an auxiliary component. This separation prevents MLLMs from jointly optimizing visual perception and textual reasoning. To address this issue, we propose L2T (adaptive Look, flexible Thought), a unified reasoning framework that jointly optimizes the form and depth of computation. L2T first discretizes tool-produced visual states into visual tokens, bringing visual perception and textual reasoning into a shared autoregressive space. Based on this unified representation, L2T introduces a modality-aware routing reward to select appropriate reasoning modes and a difficulty-aware fast-slow strategy to adapt reasoning depth according to task complexity. Experiments on general perception and reasoning-intensive tasks demonstrate the effectiveness of L2T. Compared with explicit tool-calling approaches, L2T maintains competitive performance while achieving 1.7–3.6 times higher throughput.

Automated moderation systems are deployed at scale, yet they remain opaque to the users and moderators affected by their decisions, and their coverage is almost entirely confined to English. For Urdu written in the Nastaliq script, no such system exists, and the annotated resources it would be built on are drawn largely from Twitter/X, where character limits constrain how users write. Interpretability has received little attention, leaving it unclear what such models actually learn. This study addresses these gaps in three steps. First, we curate and publicly release a corpus of 7604 Nastaliq Urdu Facebook comments, extending an earlier corpus by correcting text-level inconsistencies and re-examining the original labels. Second, we fine-tune four Urdu-capable transformers under a common grid-search protocol and compare them against twelve feature-classifier baselines on identical data partitions. The best configuration, Multilingual Representations for Indian Languages (MuRIL), reaches 87.05% accuracy and 87.06% macro F1, improving on the strongest baseline by 5.52 percentage points, a margin that exceeds the 95% confidence interval on test accuracy; the two leading transformers differ by less than the test set can resolve, and performance is stable across five optimizers. Third, we compute word-level SHAP attributions adapted to right-to-left rendering. Attribution mass concentrates on high-frequency slurs but extends to syntactically neutral function words and to political vocabulary characteristic of the source pages, indicating that aggregate performance may rest partly on lexical and topical regularities rather than a generalizable representation of offensiveness.

Large language model (LLM) agents are increasingly deployed as teams, yet they fail in characteristically social ways: information holders stay silent, dominant speakers anchor groups on wrong answers, and norm violations go unsanctioned. We present Agent Sociodrama, a framework for modeling and validating evolving social dynamics in LLM agent teams. Under fixed personas, agents' emotional states and directed, asymmetric relationships are updated every round by a unified external LLM judge and shape their subsequent interactions. Across six LLM backbones and five social scenarios, we assess social fidelity through classical social-theory patterns and agreement with human judgments. The resulting trajectories exhibit structured social patterns and model-specific differences. We then introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction transcripts and social trajectories and recasts agent personas before rerunning the task. Across four collaborative benchmarks, TPR improves performance from both heterogeneous and randomly initialized agent persona compositions, yielding up to 83.3%p of accuracy improvement over the initial casting. On representative distributed-information cooperation and mixed-motive bargaining tasks, we outperform SOTOPIA and AgentVerse in three of four comparisons, by up to 37.5%p in Pareto rate. These results show that validated social trajectories provide actionable signals for repairing LLM agent teams.

Training-free pruning removes weights from pretrained large language models to reduce their deployment cost. However, existing approaches typically consider only the total number of retained weights. Consequently, their pruning decisions do not account for how target hardware stores and processes those weights. These methods select a sparsity structure in advance and apply it uniformly to all projections, rather than accounting for how each structure trades pruning error against latency and memory across sublayers. We propose Heterogeneous Structured Pruning (HSP), a hardware-aware pruning framework that combines two components: (i) a heterogeneous search space of dense, unstructured, 2:4, block-sparse, and head / channel (H/C) pruning options and (ii) structure-regularising constraints based on latency and memory models. Our contribution generalises existing importance-based methods: a global solver selects one option per sublayer to minimise proxy pruning loss under a resource budget, thereby assigning each projection a structure and density. We evaluate HSP on OPT and LLaMA models for an NVIDIA Jetson Orin Nano Super edge device, where it improves the quality-resource frontiers over training-free baselines. A project page and updated links for the related code is maintained at: cameronbarker.me/publications/hsp.

Under GRPO against a frozen, reference-free LLM judge, RL exploits the judge's false positives. Does the policy-judge capability gap or the judge itself set how fast and how far? We train 0.5B and 1.5B Qwen2.5 policies on GSM8K against four judges at least as large, and the 1.5B policy also against one smaller judge. In a post-hoc analysis, judge acceptance rose, but no cell of 12 gained more than 0.04 in seed-mean held-out greedy accuracy. A gold-reward control with the same recipe raised the 1.5B policy's accuracy by +0.054 (for the 0.5B policy, gold reward gave no detectable gain either). The depth of judge hacking tracks the judge: the plateau correlates with judge AUC (Spearman -0.69 over 12 cells), and within a family the larger judge is fooled less deeply. For the pace, two pre-specified pairs contradict the gap account: the cell with the larger gap was fooled 3.5x sooner at the pre-specified threshold. With 12 cells, however, we cannot formally separate judge AUC from the gap for either endpoint. A blind test of step-0 predictors on three held-out judges was inconclusive. At 7-8B, judges from three families plateaued between 0.37 and 0.85.

Recent focus in agentic reinforcement learning (RL) seeks to improve credit assignment in long-horizon tasks. One promising direction combines reinforcement learning with verifiable rewards (RLVR) and on-policy self-distillation (OPSD), integrating trajectory-level advantages with dense token-level supervision. This line of work often conditions a teacher model on privileged context, such as agentic skills, to score tokens generated by the student policy. Despite their progress, they largely overlook the on-policy evolution of agentic skills and the reliability of teacher supervision. Towards this end, we introduce Recursive Self-Improving Policy Optimization (RSIPO), a self-evolving agentic RL framework that enables agentic skills and the policy model to co-evolve throughout the training process. Specifically, RSIPO enables the policy itself to extract skills from completed trajectories and explicitly optimizes its skill-generation capability through trajectory replay. It further employs empirical Bayes shrinkage to dynamically calibrate the student's trust in teacher scores, accounting for the fact that skill-enhanced teacher guidance can sometimes be less reliable. Experiments with Qwen-family models on embodied interaction and search-based agentic tasks demonstrate that RSIPO outperforms all the compared baseline approaches. With Qwen2.5-7B as the base model, RSIPO achieves an average success rate of 98.6% on ALFWorld, yielding absolute improvements of 2.5% and 17.4% over the second-best baseline and standard GRPO, respectively.

For languages with limited resources, Sentiment Analysis remains a challenging problem due to limited annotated data and strong domain variability across textual sources. Marathi, despite being widely spoken, lacks extensive multi-domain sentiment analysis studies. Recent efforts by L3Cube have introduced multi-domain datasets in Marathi, enabling a better study of domain-aware sentiment classification. This study uses the pretrained MahaBERT model to improve sentiment classification for Marathi across different domains. Models are trained on individual domain datasets and domain-adaptive pretraining is applied using masked language modeling to help the model better understand domain-specific language. In addition, back-translation is used as a method of data augmentation to increase the diversity of training data. Results show that these methods lead to consistent improvements in accuracy across all domains. The improvements range from a minimum of about 0.12% to a maximum of about 1.74%, depending on the dataset. These results demonstrate that domain-adaptive pretraining and data augmentation are effective strategies for improving sentiment analysis models. This helps achieve more robust performance across different domains and makes the approach more suitable for low-resource languages like Marathi.

Sentiment analysis is a key task in Natural Language Processing that determines the polarity of opinions in text. For low-resource languages like Marathi–where limited data prevents the direct use of standard statistical methods–models are often trained on single-domain datasets. While this yields strong in-domain performance, it leads to poor generalization to unseen domains. Recent studies highlight the need for domain-robust systems that can handle diverse real-world content through multi-domain modeling. This study investigates how multi-domain training improves sentiment classification for unseen Marathi domains. Models are trained using single-domain, two-domain, and three-domain dataset combinations. Results show that two-domain training improves accuracy on unseen domains by about 8% compared to single-domain training. Adding a third domain provides further improvement, though the gain is smaller, averaging around 2% for some domains. These results suggest that training with data from multiple domains helps build more robust sentiment analysis models for low-resource languages, though the benefits reduce as more domains are added.

Pluralistic alignment asks that language models faithfully represent a society's diverse, competing values rather than collapse them into one dominant viewpoint, yet flagship pluralistic-alignment datasets remain heavily skewed toward Western respondents. We introduce NepPlural, a human-anchored benchmark for modelling competing socio-cultural personas in Nepali public discourse on youth migration and brain drain. We curate 2,148 code-mixed (Devanagari, Romanized Nepali, English) YouTube comments, annotate them along a four-axis persona taxonomy (Intent, Primary Driver, Value Orientation, Affect) plus an explicit No-Persona class, and build a 300-comment gold set labelled independently by three native Nepali speakers (mean Fleiss' $\kappa \approx 0.55$). Using this gold set to audit automatic annotation, we find that LLM annotation reliability is driven by protocol. Labels collected in batches through chat interfaces reach only near-chance inter-model agreement ($\kappa \approx 0.17$). Annotating one comment at a time raises agreement above the native-speaker level ($\kappa \approx 0.72$), but consistency is not validity: under a forced-choice schema, models agree on personas for the 39% of comments native speakers judge to express none, keeping macro-F1 against human gold near 0.50. Allowing abstention resolves this: models recover No-Persona comments with F1 of 0.79 to 0.87 and, against leave-one-annotator-out references, match or exceed individual native speakers (macro-F1 0.75 to 0.86 vs. 0.75). Encoders fine-tuned on the batched labels reach only 0.34. Finally, a human-judged steerability probe finds that three frontier models adopt single persona axes with 59 to 77% fidelity but compose full personas only 33% of the time, with no congruence gap detected at this scale between Western-congruent and Nepali-collectivist personas.

Many questions about large textual corpora concern patterns that unfold over time, such as a political group whose stance on a topic drifts from support to criticism over several years of debates. We present TL;QR (Too Long; Query with logical temporal Reasoning), a declarative query language and open-source engine for finding such temporal discursive patterns. In TL;QR, a single query combines natural-language predicates, whose truth is judged by a Large Language Model (LLM), with regular expressions and the operators of a temporal logic. The LLM answers only local questions about individual text passages; a symbolic engine then aggregates these judgments over time and links every match to the recorded passages and predicate decisions behind it. With this division of labor, the temporal and relational reasoning stays exact and auditable even over long time horizons. In addition, when a query includes a regular-expression pre-filter, only a small fraction of the corpus ever reaches the LLM. On a suite of controlled synthetic benchmarks with temporally and semantically complex patterns, TL;QR is substantially more accurate than LLM-only baselines. In a case study on ten years of European Parliament debates (442,593 multilingual speeches), TL;QR identifies candidate shifts in the political groups' stances on the GDPR privacy regulation.

Linear attention and recurrent models compress past tokens into a fixed-size state that is updated at every step. However, storing these states in high precision incurs substantial memory and bandwidth costs at large batch sizes, motivating low-bit state quantization. Unlike weights and key–value caches, a recurrent state is requantized after every update, so its quantization errors accumulate across steps. Our analysis shows that these errors are highly coherent across steps, so reducing per-step errors alone is insufficient. This coherence arises mainly from small updates that are lost in rounding. To preserve these updates, we intervene both before and at rounding. Before rounding, we propose Low-Rank Bypass (LRB), which accumulates small updates in a compact low-rank buffer and periodically merges its sufficiently large elements into the state. At rounding, we observe that a state element can even move opposite to its update, and such reversals prove particularly harmful. Therefore, we propose Bounded Directional Rounding (BDR), which rounds reversed elements toward their updates without additional storage. Although BDR enlarges the rounding error, it improves the model, suggesting that the direction of an update matters more than its rounding error at a single step. Across Gated DeltaNet, GLA, Mamba2, and RetNet, our training-free method substantially outperforms INT4, while reducing state memory by up to 5.3$\times$ and speeding up decoding by up to 1.9$\times$ relative to FP32 states.

A central desideratum in model editing is generalization to complex tasks, which prior work formalizes under a direct generalization paradigm: treating and evaluating LLM behavior as an input-output mapping and expecting a factual update to alter outputs on reasoning tasks. Yet despite substantial empirical effort, the field consensus remains that such generalization is highly challenging and elusive. However, when we deploy the latest models in an in-the-wild chatbot-user setting, we surprisingly find that even classical editing methods exhibit remarkably fluent generalization behavior. This motivates us to revisit the obstacles to generalization and the limitations of the prevailing paradigm. We find that direct generalization is fundamentally constrained both theoretically and practically: theoretically, it suffers from an identifiability issue; and practically, existing approaches attempting to circumvent this are further hindered by the ambiguity of their constructed biases and by vulnerability to being overridden by other factors such as optimization dynamics. We instead identify compositional generalization, empowered by the increasingly strong reasoning and composition abilities of modern LLMs, as a theoretically grounded and empirically validated alternative. We validate this phenomenon across multiple models and editing methods, showing that such generalization behavior is widespread. Overall, we argue that generalizable model editing could be more than direct generalization, and call for a paradigm shift toward more generalizable and practically deployable knowledge update techniques.