ILRe: Intermediate Layer Retrieval for Context Compression in Causal Language Models

Large Language Models (LLMs) have demonstrated success across many benchmarks. However, they still exhibit limitations in long-context scenarios, primarily due to their short effective context length, quadratic computational complexity, and high memory overhead when processing lengthy inputs. To mitigate these issues, we introduce a novel context compression pipeline, called Intermediate Layer Retrieval (ILRe), which determines one intermediate decoder layer offline, encodes context by streaming chunked prefill only up to that layer, and recalls tokens by the attention scores between the input query and full key cache in that specified layer. In particular, we propose a multi-pooling kernels allocating strategy in the token recalling process to maintain the completeness of semantics. Our approach not only reduces the prefilling complexity from $O(L^2)$ to $O(L)$, but also achieves performance comparable to or better than the full context in the long context scenarios. Without additional post training or operator development, ILRe can process a single $1M$ tokens request in less than half a minute (speedup $\approx 180\times$) and scores RULER-$1M$ benchmark of $\approx 79.8$ with model Llama-3.1-UltraLong-8B-1M-Instruct on a Huawei Ascend 910B NPU.

CausalSent: Interpretable Sentiment Classification with RieszNet

Despite the overwhelming performance improvements offered by recent natural language procesing (NLP) models, the decisions made by these models are largely a black box. Towards closing this gap, the field of causal NLP combines causal inference literature with modern NLP models to elucidate causal effects of text features. We replicate and extend Bansal et al's work on regularizing text classifiers to adhere to estimated effects, focusing instead on model interpretability. Specifically, we focus on developing a two-headed RieszNet-based neural network architecture which achieves better treatment effect estimation accuracy. Our framework, CausalSent, accurately predicts treatment effects in semi-synthetic IMDB movie reviews, reducing MAE of effect estimates by 2-3x compared to Bansal et al's MAE on synthetic Civil Comments data. With an ensemble of validated models, we perform an observational case study on the causal effect of the word "love" in IMDB movie reviews, finding that the presence of the word "love" causes a +2.9% increase in the probability of a positive sentiment.

The AI Data Scientist

Imagine decision-makers uploading data and, within minutes, receiving clear, actionable insights delivered straight to their fingertips. That is the promise of the AI Data Scientist, an autonomous Agent powered by large language models (LLMs) that closes the gap between evidence and action. Rather than simply writing code or responding to prompts, it reasons through questions, tests ideas, and delivers end-to-end insights at a pace far beyond traditional workflows. Guided by the scientific tenet of the hypothesis, this Agent uncovers explanatory patterns in data, evaluates their statistical significance, and uses them to inform predictive modeling. It then translates these results into recommendations that are both rigorous and accessible. At the core of the AI Data Scientist is a team of specialized LLM Subagents, each responsible for a distinct task such as data cleaning, statistical testing, validation, and plain-language communication. These Subagents write their own code, reason about causality, and identify when additional data is needed to support sound conclusions. Together, they achieve in minutes what might otherwise take days or weeks, enabling a new kind of interaction that makes deep data science both accessible and actionable.

Detecting and Characterizing Planning in Language Models

Modern large language models (LLMs) have demonstrated impressive performance across a wide range of multi-step reasoning tasks. Recent work suggests that LLMs may perform planning - selecting a future target token in advance and generating intermediate tokens that lead towards it - rather than merely improvising one token at a time. However, existing studies assume fixed planning horizons and often focus on single prompts or narrow domains. To distinguish planning from improvisation across models and tasks, we present formal and causally grounded criteria for detecting planning and operationalize them as a semi-automated annotation pipeline. We apply this pipeline to both base and instruction-tuned Gemma-2-2B models on the MBPP code generation benchmark and a poem generation task where Claude 3.5 Haiku was previously shown to plan. Our findings show that planning is not universal: unlike Haiku, Gemma-2-2B solves the same poem generation task through improvisation, and on MBPP it switches between planning and improvisation across similar tasks and even successive token predictions. We further show that instruction tuning refines existing planning behaviors in the base model rather than creating them from scratch. Together, these studies provide a reproducible and scalable foundation for mechanistic studies of planning in LLMs.

LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components

Large Language Models (LLMs) often display overconfidence, presenting information with unwarranted certainty in high-stakes contexts. We investigate the internal basis of this behavior via mechanistic interpretability. Using open-sourced Llama 3.2 models fine-tuned on human annotated assertiveness datasets, we extract residual activations across all layers, and compute similarity metrics to localize assertive representations. Our analysis identifies layers most sensitive to assertiveness contrasts and reveals that high-assertive representations decompose into two orthogonal sub-components of emotional and logical clusters-paralleling the dual-route Elaboration Likelihood Model in Psychology. Steering vectors derived from these sub-components show distinct causal effects: emotional vectors broadly influence prediction accuracy, while logical vectors exert more localized effects. These findings provide mechanistic evidence for the multi-component structure of LLM assertiveness and highlight avenues for mitigating overconfident behavior.

Unbiased Reasoning for Knowledge-Intensive Tasks in Large Language Models via Conditional Front-Door Adjustment

Large Language Models (LLMs) have shown impressive capabilities in natural language processing but still struggle to perform well on knowledge-intensive tasks that require deep reasoning and the integration of external knowledge. Although methods such as Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) have been proposed to enhance LLMs with external knowledge, they still suffer from internal bias in LLMs, which often leads to incorrect answers. In this paper, we propose a novel causal prompting framework, Conditional Front-Door Prompting (CFD-Prompting), which enables the unbiased estimation of the causal effect between the query and the answer, conditional on external knowledge, while mitigating internal bias. By constructing counterfactual external knowledge, our framework simulates how the query behaves under varying contexts, addressing the challenge that the query is fixed and is not amenable to direct causal intervention. Compared to the standard front-door adjustment, the conditional variant operates under weaker assumptions, enhancing both robustness and generalisability of the reasoning process. Extensive experiments across multiple LLMs and benchmark datasets demonstrate that CFD-Prompting significantly outperforms existing baselines in both accuracy and robustness.

MissionHD: Data-Driven Refinement of Reasoning Graph Structure through Hyperdimensional Causal Path Encoding and Decoding

Reasoning graphs from Large Language Models (LLMs) are often misaligned with downstream visual tasks such as video anomaly detection (VAD). Existing Graph Structure Refinement (GSR) methods are ill-suited for these novel, dataset-less graphs. We introduce Data-driven GSR (D-GSR), a new paradigm that directly optimizes graph structure using downstream task data, and propose MissionHD, a hyperdimensional computing (HDC) framework to operationalize it. MissionHD uses an efficient encode-decode process to refine the graph, guided by the downstream task signal. Experiments on challenging VAD and VAR benchmarks show significant performance improvements when using our refined graphs, validating our approach as an effective pre-processing step.

DEPTH: Hallucination-Free Relation Extraction via Dependency-Aware Sentence Simplification and Two-tiered Hierarchical Refinement

Relation extraction enables the construction of structured knowledge for many downstream applications. While large language models (LLMs) have shown great promise in this domain, most existing methods concentrate on relation classification, which predicts the semantic relation type between a related entity pair. However, we observe that LLMs often struggle to reliably determine whether a relation exists, especially in cases involving complex sentence structures or intricate semantics, which leads to spurious predictions. Such hallucinations can introduce noisy edges in knowledge graphs, compromising the integrity of structured knowledge and downstream reliability. To address these challenges, we propose DEPTH, a framework that integrates Dependency-aware sEntence simPlification and Two-tiered Hierarchical refinement into the relation extraction pipeline. Given a sentence and its candidate entity pairs, DEPTH operates in two stages: (1) the Grounding module extracts relations for each pair by leveraging their shortest dependency path, distilling the sentence into a minimal yet coherent relational context that reduces syntactic noise while preserving key semantics; (2) the Refinement module aggregates all local predictions and revises them based on a holistic understanding of the sentence, correcting omissions and inconsistencies. We further introduce a causality-driven reward model that mitigates reward hacking by disentangling spurious correlations, enabling robust fine-tuning via reinforcement learning with human feedback. Experiments on six benchmarks demonstrate that DEPTH reduces the average hallucination rate to 7.0\% while achieving a 17.2\% improvement in average F1 score over state-of-the-art baselines.

GALA: Can Graph-Augmented Large Language Model Agentic Workflows Elevate Root Cause Analysis?

Root cause analysis (RCA) in microservice systems is challenging, requiring on-call engineers to rapidly diagnose failures across heterogeneous telemetry such as metrics, logs, and traces. Traditional RCA methods often focus on single modalities or merely rank suspect services, falling short of providing actionable diagnostic insights with remediation guidance. This paper introduces GALA, a novel multi-modal framework that combines statistical causal inference with LLM-driven iterative reasoning for enhanced RCA. Evaluated on an open-source benchmark, GALA achieves substantial improvements over state-of-the-art methods of up to 42.22% accuracy. Our novel human-guided LLM evaluation score shows GALA generates significantly more causally sound and actionable diagnostic outputs than existing methods. Through comprehensive experiments and a case study, we show that GALA bridges the gap between automated failure diagnosis and practical incident resolution by providing both accurate root cause identification and human-interpretable remediation guidance.

LoraxBench: A Multitask, Multilingual Benchmark Suite for 20 Indonesian Languages

As one of the world's most populous countries, with 700 languages spoken, Indonesia is behind in terms of NLP progress. We introduce LoraxBench, a benchmark that focuses on low-resource languages of Indonesia and covers 6 diverse tasks: reading comprehension, open-domain QA, language inference, causal reasoning, translation, and cultural QA. Our dataset covers 20 languages, with the addition of two formality registers for three languages. We evaluate a diverse set of multilingual and region-focused LLMs and found that this benchmark is challenging. We note a visible discrepancy between performance in Indonesian and other languages, especially the low-resource ones. There is no clear lead when using a region-specific model as opposed to the general multilingual model. Lastly, we show that a change in register affects model performance, especially with registers not commonly found in social media, such as high-level politeness `Krama' Javanese.