Retrieving evidence for language model queries from knowledge graphs requires balancing broad search across the graph with multi-hop traversal to follow relational links. Similarity-based retrievers provide coverage but remain shallow, whereas traversal-based methods rely on selecting seed nodes to start exploration, which can fail when queries span multiple entities and relations. We introduce ARK: Adaptive Retriever of Knowledge, a tool-using KG retriever that gives a language model control over this breadth-depth tradeoff using a two-operation toolset: global lexical search over node descriptors and one-hop neighborhood exploration that composes into multi-hop traversal. ARK alternates between breadth-oriented discovery and depth-oriented expansion without depending on a fragile seed selection, a pre-set hop depth, or requiring retrieval training. ARK adapts tool use to queries, using global search for language-heavy queries and neighborhood exploration for relation-heavy queries.On STaRK, ARK reaches 59.1% average Hit@1 and 67.4 average MRR, improving average Hit@1 by up to 31.4% and average MRR by up to 28.0% over retrieval-based and agent-based training-free methods.Finally, we distill ARK’s tool-use trajectories from a large teacher into an 8B model via label-free imitation, improving Hit@1 by +7.0, +26.6, and +13.5 absolute points over the base 8B model on AMAZON, MAG, and PRIME datasets, respectively, while retaining up to 98.5% of the teacher’s Hit@1 rate.

Production LLM systems often rely on separate models for safety and other classification-heavy steps, increasing latency, VRAM footprint, and operational complexity. We instead reuse computation already paid for by the serving LLM: we train lightweight probes on its hidden states and predict labels in the same forward pass used for generation. We frame classification as representation selection over the full token×layer hidden-state tensor, rather than committing to a fixed token or fixed layer (e.g., first-token logits or final-layer pooling). To implement this, we introduce a two-stage aggregator that (i) summarizes tokens within each layer and (ii) aggregates across layer summaries to form a single representation for classification. We instantiate this template with direct pooling, a 100K-parameter scoring-attention gate, and a downcast multi-head self-attention (MHA) probe with up to 35M trainable parameters. Across safety and sentiment benchmarks our probes improve over logit-only reuse (e.g., MULI) and are competitive with substantially larger task-specific baselines, while preserving near-serving latency and avoiding the VRAM and latency costs of a separate guard-model pipeline. Multi-backbone experiments on dense and mixture-of-experts architectures (Llama-3.2-3B, GPT-OSS-20B, Qwen3-30B-A3B) confirm that these findings generalize beyond a single model family.

Despite the remarkable success of large language models (LLMs) in English, a significant performance gap remains in non-English languages. To address this, we introduce a novel approach for strategically constructing a multilingual synthetic instruction tuning dataset, sPhinX. Unlike prior methods that directly translate fixed instruction-response pairs, sPhinX enhances diversity by selectively augmenting English instruction-response pairs with multilingual translations. Additionally, we propose LANGIT, a novel N-shot guided fine-tuning strategy, which further enhances model performance by incorporating contextually relevant examples in each training sample. Our ablation study shows that our approach enhances the multilingual capabilities of Mistral-7B and Phi-3-Small improving performance by an average of 39.8% and 11.2%, respectively, across multilingual benchmarks in reasoning, question answering, reading comprehension, and machine translation. Moreover, sPhinX maintains strong performance on English LLM benchmarks while exhibiting minimal to no catastrophic forgetting, even when trained on 51 languages.

We introduce a novel framework for analyzing sorting algorithms in pairwise ranking prompting (PRP), re-centering the cost model around LLM inferences rather than traditional pairwise comparisons. While classical metrics based on comparison counts have traditionally been used to gauge efficiency, our analysis reveals that expensive LLM inferences overturn these predictions; accordingly, our framework encourages strategies such as batching and caching to mitigate inference costs. We show that algorithms optimal in the classical setting can lose efficiency when LLM inferences dominate the cost under certain optimizations.

Alignment serves as an important step to steer large language models (LLMs) towards human preferences. In this paper, we propose an automatic way to construct contrastive data for LLM, using preference pairs from multiple models of varying strengths (e.g., InstructGPT, ChatGPT and GPT-4). We compare the contrastive techniques of SLiC and DPO to SFT baselines and find that DPO provides a step-function improvement even after continuing SFT saturates. We also explore a data curriculum learning scheme for contrastive post-training, which starts by learning from “easier” pairs and transitioning to “harder” ones, which further improves alignment. Finally, we scale up our experiments to train with more data and larger models like Orca. Remarkably, our automatic contrastive post-training further improves the performance of Orca, already a state-of-the-art instruction learning model tuned with GPT-4 outputs, to outperform ChatGPT.

The question of how to make decisions that maximise the well-being of all persons is very relevant to design language models that are beneficial to humanity and free from harm. We introduce the Greatest Good Benchmark to evaluate the moral judgments of LLMs using utilitarian dilemmas. Our analysis across 15 diverse LLMs reveals consistently encoded moral preferences that diverge from established moral theories and lay population moral standards. Most LLMs have a marked preference for impartial beneficence and rejection of instrumental harm. These findings showcase the ‘artificial moral compass’ of LLMs, offering insights into their moral alignment.

We investigate post-OCR correction in a setting where we have access to different OCR views of the same document. The goal of this study is to understand if a pretrained language model (LM) can be used in an unsupervised way to reconcile the different OCR views such that their combination contains fewer errors than each individual view. This approach is motivated by scenarios in which unconstrained text generation for error correction is too risky. We evaluated different pretrained LMs on two datasets and found significant gains in realistic scenarios with up to 15% WER improvement over the best OCR view. We also show the importance of domain adaptation for post-OCR correction on out-of-domain documents.

We present a method to automatically identify financially relevant news using stock price movements and news headlines as input. The method repurposes the attention weights of a neural network initially trained to predict stock prices to assign a relevance score to each headline, eliminating the need for manually labeled training data. Our experiments on the four most relevant US stock indices and 1.5M news headlines show that the method ranks relevant news highly, positively correlated with the accuracy of the initial stock price prediction task.

Named Entity Disambiguation (NED) systems perform well on news articles and other texts covering a specific time interval. However, NED quality drops when inputs span long time periods like in archives or historic corpora. This paper presents the first time-aware method for NED that resolves ambiguities even when mention contexts give only few cues. The method is based on computing temporal signatures for entities and comparing these to the temporal contexts of input mentions. Our experiments show superior quality on a newly created diachronic corpus.

In this work, we discuss the importance of external knowledge for performing Named Entity Recognition (NER). We present a novel modular framework that divides the knowledge into four categories according to the depth of knowledge they convey. Each category consists of a set of features automatically generated from different information sources, such as a knowledge-base, a list of names, or document-specific semantic annotations. Further, we show the effects on performance when incrementally adding deeper knowledge and discuss effectiveness/efficiency trade-offs.

This work introduces fact salience: The task of generating a machine-readable representation of the most prominent information in a text document as a set of facts. We also present SalIE, the first fact salience system. SalIE is unsupervised and knowledge agnostic, based on open information extraction to detect facts in natural language text, PageRank to determine their relevance, and clustering to promote diversity. We compare SalIE with several baselines (including positional, standard for saliency tasks), and in an extrinsic evaluation, with state-of-the-art automatic text summarizers. SalIE outperforms baselines and text summarizers showing that facts are an effective way to compress information.

The goal of Open Information Extraction (OIE) is to extract surface relations and their arguments from natural-language text in an unsupervised, domain-independent manner. In this paper, we propose MinIE, an OIE system that aims to provide useful, compact extractions with high precision and recall. MinIE approaches these goals by (1) representing information about polarity, modality, attribution, and quantities with semantic annotations instead of in the actual extraction, and (2) identifying and removing parts that are considered overly specific. We conducted an experimental study with several real-world datasets and found that MinIE achieves competitive or higher precision and recall than most prior systems, while at the same time producing shorter, semantically enriched extractions.