TRACE is a deployable framework that enhances oncology language models by separating offline structure learning from online inference. It organizes oncology concepts and relations into an updatable tree‑relational structure, refines it with LM‑loss evidence, and retrieves compact prompt evidence during inference. The approach improves performance on ten classification tasks and a QA benchmark, outperforms vanilla RAG and generic GraphRAG, and provides interpretable evidence paths aligned with clinical reasoning.
By Jizheng Lai, Yingyun Li, Ying Qin, Haiyang Qian
The paper investigates whether compact surrogate models can reduce the cost of simulating large language model (LLM) societies while still reproducing their collective behavior. Using 9,455 published trajectories and new opinion‑dynamics experiments, it finds that incorporating neighbor information improves individual predictions across all 16 public‑data settings and enhances pooled collective forecasts on held‑out questions, though the collective gains vary with transfer conditions. Additional tests on 24 new statements do not confirm earlier contrasting history effects, and Qwen shows benefit from history only after three observed rounds, underscoring the need for direct collective validation, explicit limits on available observations, and comparisons with simple baselines.
By Igor Itkin
The paper introduces Less Uniform Diffusion (LUDI), a framework that improves uniform diffusion language models by using a less uniform loss and per-token time embeddings to guide reverse transitions and enable confidence-based few-step sampling. Experiments demonstrate that LUDI provides cleaner supervision, enhances few-step generation, and scales to a 7B model (LUDI-7B) that achieves a 3-token-per-step speedup over autoregressive decoding while matching masked diffusion baselines. The work suggests that UDLMs still have untapped potential for complex generation tasks.
By Kaibo Wang, Ding Ding, Fangyu Ding, Zijin Feng, Han Shi, Haili Bai, Jiacheng Sun, Yang Xiang
The paper investigates intrinsic self‑correction, where a language model revises its own answer without new evidence. Across 29 open‑weight LLMs on BoolQ, GSM8K, and Corr2Cause, the study tracks how revisions change correctness, revealing that while some models improve significantly, others lose a notable fraction of correct answers. The authors compare three runtime strategies—keeping the initial answer, always accepting the revision, and selectively gating revisions—and find that the best approach depends on the model and task, suggesting that self‑correction should be treated as a revision policy rather than a uniformly beneficial second pass.
By Tianzhu Zhang
The paper introduces StepLearn, a nonparametric framework for prequential test‑time learning in large language model agents. StepLearn separates immediate use of informative transitions from persistent trust, turning each transition into a hypothesis that guides the next step and only reusing it after prospective validation across episodes. Experiments on WebArena‑Lite and ALFWorld show StepLearn improves success rates by 2.2–12.7 percentage points over the strongest baseline, with benefits evident from the first task attempts.
By Tong Zhao, Reed Li, Yuyang Hu, Yutao Zhu, Haijin Liang, Haibo Shi, Yu Lu, Zhicheng Dou
The paper introduces RedHerring, a defense mechanism that inserts safe decoy vulnerabilities into code repositories to divert autonomous LLM agents’ verification efforts away from real security flaws. By embedding CVE-derived vulnerability chains with false bridges and providing a private certificate for quick verification, RedHerring forces agents to spend a significant portion of their limited resources on decoys. Experiments on 33 OSS‑Fuzz projects show a 38.7‑60.4% reduction in discovered real vulnerabilities, even when agents are aware of decoys.
By Kaikai Zhang, Zihan Zhang, Yuchong Xie, Zesen Liu, Shuangjie Yao, Zhixiang Zhang, Dongdong She
UNBIND is a code unlearning framework that selectively removes memorized code from large language models at inference time while keeping the model weights unchanged. It constructs separate directional steering for hidden states that correspond to target code, enabling high forgetting rates (97.3–99.1% reduction in target code reproduction) with minimal loss in programming utility. Across multiple baselines, corpora, and evaluation metrics—including F‑BLEU, HumanEval+, and MBPP+—UNBIND consistently achieves the best joint forgetting and utility scores, and it effectively eliminates long exact code spans in repeated extraction tests.
By Zhengyang Shan, Jiayun Xin, Yanjun Lin, Xu Qian, Zhiang Liu, Minghui Xu, Yue Zhang, Qin Hu, Kun Li, Xiuzhen Cheng
Mnemon is a memory agent that stores conversations as raw, dated records and uses a fast System 1 decision model (Jev) to quickly judge the relevance of records, while a slow System 2 LLM plans searches and composes answers. The agent consolidates records into topic timelines and value histories in the background, enabling efficient retrieval without rewriting conversations into structured formats. Experiments show Mnemon achieving high scores on LoCoMo and LongMemEval‑S with low context length and cost, and Jev outperforming LLMs in evidence separation and speed.
By Guangren Wang
The study examines how the release of ChatGPT-3.5 affected collective knowledge on Stack Overflow, analyzing over two million questions from 2020 to 2025. It finds that easy questions declined sharply while difficult ones increased, with rising code complexity. Data-rich topics lost share of questions, whereas data-scarce ones gained, and these patterns held across multiple programming languages.
By Myokyung Han, Taegyoon Kim, Jinhyuk Yun, Lanu Kim
PADM'E is a method for synthesizing preference‑aligned data to meta‑evaluate language‑model (LM) evaluators of agentic behaviors. It reframes meta‑evaluation as a preference judgment problem, generating criterion‑based data with small LMs and no human involvement. In a prototype, PADM'E produced 1,000 samples across four domains and three criteria, and human validation showed agreement with human judgment rising from 73% to 85% compared to a naive baseline.
By Cheng Chang, Yining Mao, Peng Qi
The paper investigates how the modality gap— the separation between image and text representations in contrastive vision‑language models—affects different downstream tasks. By showing that a single dominant direction accounts for most of the image‑text mean separation, the authors explain why reducing or removing this gap can improve zero‑shot classification, degrade retrieval, or restore performance depending on the task. The study provides a geometric framework that clarifies when and why gap interventions should be applied in vision‑language systems.
By Aditya Sharma, Divya Saxena
The paper investigates how multimodal large language models are more susceptible to prompt injection when adversarial instructions are presented as text rather than as non-textual inputs like images. It proposes a training‑free defense that renders untrusted payloads into typographic images (or audio) before they reach the model, a method called Pictionary. Experiments on ten models and two benchmarks show that this approach significantly lowers attack success rates while maintaining normal functionality, and that fine‑tuning on image‑rendered instructions can further reduce the modality gap.
By Jie Zhang, Andrei Baroian, Jan N. van Rijn, Avital Shafran, Florian Tram\`{e}r
ThinQuant introduces efficient rotation learning for low‑bit weight and activation quantization of large language models by reducing calibration data through a geometric selection of activations and solving a lower‑dimensional optimization problem via an ADMM algorithm. The method achieves comparable or better quantization performance with dramatically fewer calibration points, completing rotation calibration for Llama‑3‑70B in under 12 minutes and for Llama‑3.1‑405B in just over 2 hours on a single GPU. ThinQuant outperforms existing gradient‑free approaches such as DartQuant and gradient‑based SpinQuant in both speed and perplexity metrics on WikiText‑2.
By Mehdi Makni, Ryan Lucas, Rahul Mazumder
The paper investigates how large language models activate a latent symbolic reasoning circuit—comprising abstraction, induction, and retrieval—when presented with in-context examples. By tracking this circuit across different shot counts and model families, the authors show that its components become detectable and functional long before the model reaches high accuracy. They further demonstrate that per-head causal contributions can increase eightfold from 1- to 10-shot, and that interventions such as cross-shot activation patching or function vector injection can dramatically improve accuracy, even at 0-shot, by leveraging the pre‑existing circuit in the model weights.
By Melissa Wessel
The paper investigates how chunked KV‑cache compression, which reduces memory and attention costs by grouping consecutive tokens, introduces a new positional coordinate called a token’s phase. It finds that models using this compression exhibit phase sensitivity: retrieval performance can vary by up to 40 percentage points depending on a token’s phase, creating periodic weak spots that average benchmark scores hide. By pretraining transformers with different compression designs and applying causal interventions, the authors show that attention components specialize asymmetrically for different phases, and theoretical analysis suggests gradient dynamics may favor such specialization.
By Xingyu Zhu (Luke), Pu (Luke), Yi, Ziheng Cheng, Ang Lv, Jing Liu, Lexing Ying, Yiyuan Ma, Xin Dong
The paper introduces Verbalization Training (VT), a technique that encourages large language models (LLMs) to openly express their evaluation awareness (EA) without directly supervising their internal beliefs. VT works by truncating model rollouts just before spontaneous verbalizations, creating training prefixes that signal awareness, and then applying a reinforcement learning objective to increase calibrated verbalization. Experiments on models such as Qwen3.6-35B-A3B, Kimi K2.6, and Inkling show that VT boosts verbalized EA by 2.4–2.9× while keeping latent EA and overall behavior largely unchanged, and a causal study confirms that VT-induced verbalizations reflect newly acquired meta‑knowledge.
By Usman Anwar, Sahar Abdelnabi, David Krueger
The paper introduces Explainability from Training (EFT), a model‑agnostic method that tracks how deep learning models learn and organize evidence during training. EFT is applied to four leading T cell receptor‑epitope prediction models, revealing distinct learning trajectories for CNNs and transformers, conflicts between TCR alpha and beta chain evidence, and differences in feature preferences when using real versus predicted structural data. The authors also present a new benchmark, TCR‑XAI2, comprising 388 experimentally resolved TCR‑epitope structures and several predicted models to evaluate these insights.
By Jiarui Li, Zixiang Yin, Samuel Landry, Zhengming Ding, Ramgopal Mettu
AdaKerNet is a task‑adaptive neural kernel decoder that operates on frozen multimodal representations from large foundation models, without requiring access to the models’ parameters. It learns Lipschitz‑controlled multimodal features, a reference kernel providing a soft structural prior, and a lightweight nonlinear predictor that deforms this structure. Experiments on four multimodal large language models and diverse input modalities show consistent improvements over baseline decoders, achieving up to 41% error reduction in scarce‑label settings.
By Konstantinos D. Polyzos, Eleni Oikonomou, Tara Javidi
The study compares human and vision‑language model (VLM) responses to cross‑modal association tasks, using identical stimuli (a pseudo‑word and two images) and recording both choices and eye movements. While larger VLMs show some alignment with human choices, their attention patterns correlate poorly with human gaze, performing no better than a simple center‑bias baseline. Fine‑tuning VLMs on human choices improves choice alignment but not attention alignment, and training on human gaze improves attention correlation without affecting choice accuracy.
By Sumin Hong, Katsumi Ibaraki, Renee Shi, David Chiang, Toby Jia-Jun Li
arXiv:2609.36515v1 Announce Type: cross
Abstract: A common assumption in language model development is that cognitive abilities are organized around a general, domain-free intelligence factor, like f...
By Faiz Ghifari Haznitrama, Afrizal Hasbi Azizy, Faeyza Rishad Ardi