The paper investigates whether the reasoning process in large language models mitigates or exacerbates bias. Using a within-model ablation on three high-stakes datasets (Adult, COMPAS, Credit) across three 32‑B models, the authors find that reasoning resolves some counterfactual fairness flips but creates roughly five times as many new flips at high confidence. They introduce two dynamic tools—Counterfactual Depth Probability Gap and Bias Transition Matrix—to trace how bias propagates and amplifies during reasoning depth and to explain the asymmetric dual effect.
By Deng Pan, Joe Germino, Yihong Ma, Elizabeth Daly, Nuno Moniz, Ting Hua, Nitesh Chawla
ConsultMind is an uncertainty‑aware framework that automates diagnostic consultation by updating disorder posteriors after each patient response and using posterior uncertainty to guide inquiry and diagnosis. It builds on AutoDisym, a pipeline that constructs a Disorder–Symptom Bayesian Network (DSBN) from diagnostic knowledge and clinical narratives. Across psychiatry, respiratory medicine, fever clinics, and public datasets, AutoDisym produces high‑quality DSBNs and ConsultMind improves diagnostic accuracy and explanation quality, achieving up to 22.15‑point gains in Top‑1 accuracy and 37.89‑point gains in Top‑3 accuracy.
By Xiao Sun, Yuming Yang, Yun Chen, Jiang Zhong, Junnan Zhu, Xinyi Jiang, Haoyang Zeng, Ruirui Chen, Yining Wang, Xinyu Zhou, Rong Tang, Kaiwen Wei
The paper introduces Hard-Origin Adaptively Softened Memory (HasMem), a memory system for large language model agents that combines frozen hard‑prompt embeddings with a controller, writer, reader, and global module to adaptively resize and re‑encode memory entries. On a reconstruction probe of 535 questions, HasMem achieves a lexical F1 of 95.3, outperforming the hard reference by 4.4 percentage points while maintaining 93.6% of the reference’s memory positions. Across six configurations with similar per‑question budgets, the system surpasses rule‑based re‑encoding by 8.0–23.6 exact‑match points, and on LongMemEval‑S it improves local lexical F1 from 3.4 to 8.9 and reduces answer negative log‑likelihood from 12.257 to 5.274.
By Zihong He, Junxiao Shen, Chen Liang, Hai-Ning Liang
The paper introduces PTC-Decoder, a training‑free, plug‑and‑play decoder designed to improve the reliability of small language models (SLMs) on offline, resource‑constrained edge devices such as remote sensing satellites. PTC-Decoder combines a Plan‑to‑Act paradigm that forces the model to invoke a planning tool at the first inference step with a deterministic finite automaton (TC‑Decoder) that imposes token‑level hard constraints on tool names while preserving the model’s reasoning capability. Evaluated on 200 real satellite tasks across seven SLMs, the approach achieves a statistically significant mean overall score gain of +1.21, with the TC‑Decoder component identified as the primary contributor to performance gains.
By Minghui Yu, Ke Mu, Gang Wu
The paper proposes a framework that explains why jailbreak attacks succeed against diffusion-based large language models (dLLMs) by viewing safety alignment as shaping a denoising energy landscape. It identifies two attack strategies—obscuring the query’s safety disposition at initialization or forcing the denoising path across an energy barrier mid‑trajectory—and introduces three training‑free detection signals that monitor initial safety disposition and kinetic energy in complementary subspaces. Experiments on several dense and sparse dLLMs show that these signals complement each other, and any attack that evades detection also fails to produce harmful content.
By Thong Bach, Dung Nguyen, Thao Minh Le, Truyen Tran
The paper introduces Acacia, a graph foundation model trained on the Common Crawl web graph. Acacia can handle arbitrary feature dimensionalities and semantics, perform node classification, link prediction, node clustering, and graph generation, and exhibit in-context learning—all without additional training or pretrained LLMs. Unlike existing models that require extra heads or rely on LLMs, Acacia is trained from scratch and can adapt to new graphs and labels directly.
By Ryoma Sato
MACBT is a clinician‑facing AI decision‑support system that models the five‑stage CBT workflow with five collaborative agents and incorporates a CBT‑specific longitudinal memory module (CD Memory) to track cognitive distortions across sessions. The system generates pre‑session pathology reports and prioritises interventions, and was trained on a Chinese CBT dialogue corpus using a Qwen3‑14B backbone with supervised fine‑tuning and preference optimisation. Evaluation by GPT‑4 judges shows MACBT outperforms several existing chatbots in professionalism and clinical authenticity, and the memory‑augmented version improves session quality by 12.6% and achieves a longitudinal mean of 2.29 on continuity, progression, and personalization.
By De Jiang, Shuo Zhang, Weiwei Liao, Jianying Zhang, Chuanhui Yu, Hongen Liao, Kehong Yuan
LogicTree-RAG is a retrieval‑augmented generation framework that uses a hierarchical logic tree to guide the creation of long‑form patent drafts. Each node in the tree represents a technical element and is built through evidence‑guided recursive generation, while a hybrid traversal maps the tree into patent sections for balanced, controllable output. Experiments show that this logic‑centric approach improves content quality, language conformity, and token efficiency compared to strong LLM baselines.
By Jiaqi Zhu, Naili Xing, Hexiang Pan, Haotian Gao, Jianwei Yin, Xiaokui Xiao, Beng Chin Ooi
The paper introduces FTB Graph, a method for mapping the causal circuitry that determines the first-token language identity in multilingual language models. Using Edge Attribution Patching and exact activation patching across six architectures (GPT‑2, BLOOM‑560M, Pythia‑1B/2.8B, Qwen2.5‑1.5B Base/Instruct), the authors extract directed acyclic graphs that reveal deep or mid‑to‑deep broadcasting hubs, with notable differences among models. The study finds that first‑token routing is largely established during pretraining and largely preserved by instruction tuning, while linear gradient approximations can diverge from causal interventions, underscoring the need for exact‑patching verification.
By Arjun Pillai, Christian Hoang, Anjelo Laroza
MoMHa is a system that optimizes large language model harnesses across three objectives—accuracy, behavioural safety, and token cost—using a single‑phase joint‑reward proposer. It outperforms alternative strategies on seventeen domains, including synthetic suites and real‑world benchmarks, achieving higher joint scores and better safety while reducing token usage. The approach demonstrates that multi‑objective harness design can transfer effectively to unseen models and tasks.
By Subhojyoti Mukherjee, Md Mehrab Tanjim
Researchers investigated how different large language models (LLMs) convert corporate text into empirical variables, focusing on thirteen measures such as sentiment, management clarity, uncertainty, answer specificity, and climate and political risk. Using seven LLMs to score earnings call transcripts of S&P 500 companies, they found low cross-model rank correlations (average 0.52) and that transcript-level differences across providers explained only 34% of total score variation. The study shows that model choice significantly alters downstream inference, with varying coefficient magnitudes, signs, and statistical significance, and that averaging across providers stabilizes rankings but not score levels.
By Hamid Boustanifar, Sasan Mansouri
The paper introduces SCALPEL, a contrastive sparse autoencoder that selectively learns to unlearn target-specific information from neural network representations. It addresses the energy bias in standard reconstruction-based extractors, which tend to preserve dominant background structure while neglecting low-energy target components. The authors provide theoretical justification that contrastive training enhances target selectivity and demonstrate experimentally that SCALPEL outperforms NMF and standard SAE interventions on the TOFU benchmark, achieving performance comparable to Gradient Difference and RMU methods.
By Itai Zehavi, Fanny Jourdan, Ulrich Aivodji
The paper introduces Structured Thinking, a two‑turn pipeline that first generates a typed, schema‑constrained CPDAG summary and then uses that graph to answer causal queries. On the Corr2Cause benchmark, this approach boosts the Qwen3.5‑27B model’s F1 score from 73.0 to 86.4, outperforming a PC‑instruction baseline and demonstrating that externalizing the latent causal structure improves large‑language‑model reasoning. The study also shows that scrambling the CPDAG or using a prose intermediate degrades performance, supporting the principle that explicit, constrained representations aid downstream inference.
By Wentao Sun, Jo\~ao Paulo Nogueira, Dominique Verchere, Mathieu Acher, Alonso Silva
The paper introduces CodeHack, a library of code-based skills with natural-language descriptions designed to improve language agents in complex environments like NetHack. By allowing agents to invoke reusable skills instead of selecting individual actions, the study shows that skill-based agents nearly triple game progression and cut inference cost by 86% in zero‑shot settings, while still retaining the option to fall back on primitive actions. In reinforcement learning, skill-based agents learn faster, achieving a 7.2× larger average gain in dungeon level within the same training budget.
By Bart{\l}omiej Cupia{\l}, Jens Tuyls, Maciej Wo{\l}czyk, Davide Paglieri, Martin Klissarov, Benjamin Eysenbach, Piotr Mi{\l}o\'s, Karthik R. Narasimhan
OmouAI is an interactive deliberation system that combines large language models with computational argumentation to facilitate policy debates involving humans and simulated personas such as stakeholders, experts, or devil’s advocates. Each persona generates its own arguments, which are assembled into a shared argumentation framework that users can contest, add to, or revise, ensuring human oversight. The system evaluates arguments using deterministic argumentative semantics against external goals like the UN Sustainable Development Goals, providing faithful explanations and indicating how policy recommendations affect those goals.
By Stylianos Loukas Vasileiou, Antonio Rago, William Yeoh, Georgina Curto
The paper introduces LIFT, a lightweight vector‑intervention technique that transfers reasoning capability from a base large language model (LLM) to a vision‑language model (VLM) without retraining the VLM backbone. LIFT defines Reasoning Vectors as differences in hidden states between a reasoning path with an explicit trace and a solver path without it, and injects these vectors into the VLM’s language‑side activations. Experiments on two VLMs across six reasoning benchmarks show that vectors derived from the base LLM consistently outperform those derived from the aligned VLM, indicating that the base LLM is a more effective source for recovering degraded reasoning.
"whyItMatters":"The study demonstrates that a simple, frozen‑backbone intervention can partially restore reasoning abilities in multimodal models, highlighting the value of leveraging the original language model’s reasoning power."
By Ziyi Wang, Li Li, Aolin Zhou, Yankun Shen, Chonghan Liu, Shuxia Lin, Xu Yang
arXiv:2609. 31176v1 Announce Type: new Abstract: Repository-level issue localization aims to identify and rank the files and functions relevant to resolving a reported issue.
By Yunxiang Wei, Zhenyu Lei, Jundong Li
The paper introduces Stackelberg Program Optimization (SPO), a framework that uses large language models to discover adaptive destroy‑repair operators for large neighborhood search. SPO conditions operator decisions on a compact state representation, enabling state‑dependent behavior, and frames the discovery process as a Stackelberg game where destroy operators act as leaders and repair operators as conditional followers. Experiments on the traveling salesperson and capacitated vehicle routing problems show that SPO outperforms strong baselines, generalizes to larger instances, and exhibits coupled improvement in operator behavior during discovery.
By Xinyi Ke, Kai Li, Junliang Xing, Yifan Zhang, Jian Cheng
The paper argues that the current LLM-as-a-judge evaluation method, which compensates for systematic measurement bias by increasing the number of comparisons, is statistically unsound and computationally wasteful. It identifies that treating LLM judges as neutral ignores documented biases such as position bias, verbosity bias, judge severity, and self‑enhancement. The authors propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, enabling reliable rankings with far fewer comparisons and negligible additional compute.
By Harshita Katoch, David Antony Selby, Gerrit Gro{\ss}mann, Sebastian Vollmer
The paper investigates how the definition of influence—specifically the behavior being attributed, the intervention on training data, and the counterfactual training process—affects rankings produced by influence estimators. It formalizes influence as a counterfactual estimand, distinguishes specification mismatch from approximation error, and categorizes existing estimators by their implied specifications. Experiments demonstrate that different specifications can lead to markedly different rankings, and that careful specification choice improves attribution quality in tasks such as noisy label detection and large‑language‑model attribution.
By Zhe Li, Wei Zhao, Peixin Zhang, Jun Sun