Steering MoE LLMs via Expert (De)Activation

Mixture-of-Experts (MoE) in Large Language Models (LLMs) routes each token through a subset of specialized Feed-Forward Networks (FFN), known as experts. We present SteerMoE, a framework for steering MoE models by detecting and controlling behavior-linked experts. Our detection method identifies experts with distinct activation patterns across paired inputs exhibiting contrasting behaviors. By selectively (de)activating such experts during inference, we control behaviors like faithfulness and safety without retraining or modifying weights. Across 11 benchmarks and 6 LLMs, our steering raises safety by up to +20% and faithfulness by +27%. In adversarial attack mode, it drops safety by -41% alone, and -100% when combined with existing jailbreak methods, bypassing all safety guardrails and exposing a new dimension of alignment faking hidden within experts.

Enhancing Robustness of Autoregressive Language Models against Orthographic Attacks via Pixel-based Approach

Autoregressive language models are vulnerable to orthographic attacks, where input text is perturbed with characters from multilingual alphabets, leading to substantial performance degradation. This vulnerability primarily stems from the out-of-vocabulary issue inherent in subword tokenizers and their embeddings. To address this limitation, we propose a pixel-based generative language model that replaces the text-based embeddings with pixel-based representations by rendering words as individual images. This design provides stronger robustness to noisy inputs, while an extension of compatibility to multilingual text across diverse writing systems. We evaluate the proposed method on the multilingual LAMBADA dataset, WMT24 dataset and the SST-2 benchmark, demonstrating both its resilience to orthographic noise and its effectiveness in multilingual settings.

Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning

Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of NLP tasks, but they remain fundamentally stateless, constrained by limited context windows that hinder long-horizon reasoning. Recent efforts to address this limitation often augment LLMs with an external memory bank, yet most existing pipelines are static and heuristic-driven, lacking any learned mechanism for deciding what to store, update, or retrieve. We present Memory-R1, a reinforcement learning (RL) framework that equips LLMs with the ability to actively manage and utilize external memory through two specialized agents: a Memory Manager that learns to perform structured memory operations {ADD, UPDATE, DELETE, NOOP}, and an Answer Agent that selects the most relevant entries and reasons over them to produce an answer. Both agents are fine-tuned with outcome-driven RL (PPO and GRPO), enabling adaptive memory management and use with minimal supervision. With as few as 152 question-answer pairs and a corresponding temporal memory bank for training, Memory-R1 outperforms the most competitive existing baseline and demonstrates strong generalization across diverse question types and LLM backbones. Beyond presenting an effective approach, this work provides insights into how RL can unlock more agentic, memory-aware behaviors in LLMs, pointing toward richer, more persistent reasoning systems.

ImpliRet: Benchmarking the Implicit Fact Retrieval Challenge

Retrieval systems are central to many NLP pipelines, but often rely on surface-level cues such as keyword overlap and lexical semantic similarity. To evaluate retrieval beyond these shallow signals, recent benchmarks introduce reasoning-heavy queries; however, they primarily shift the burden to query-side processing techniques -- like prompting or multi-hop retrieval -- that can help resolve complexity. In contrast, we present ImpliRet, a benchmark that shifts the reasoning challenge to document-side processing: The queries are simple, but relevance depends on facts stated implicitly in documents through temporal (e.g., resolving "two days ago"), arithmetic, and world knowledge relationships. We evaluate a range of sparse and dense retrievers, all of which struggle in this setting: the best nDCG@10 is only 15.07%. We also test whether long-context models can overcome this limitation. But even with a short context of only ten documents, including the positive document, GPT-4.1 scores only 35.06%, showing that document-side reasoning remains a challenge. Our codes are available at github.com/ZeinabTaghavi/IMPLIRET.Contribution.

Hateful Person or Hateful Model? Investigating the Role of Personas in Hate Speech Detection by Large Language Models

Hate speech detection is a socially sensitive and inherently subjective task, with judgments often varying based on personal traits. While prior work has examined how socio-demographic factors influence annotation, the impact of personality traits on Large Language Models (LLMs) remains largely unexplored. In this paper, we present the first comprehensive study on the role of persona prompts in hate speech classification, focusing on MBTI-based traits. A human annotation survey confirms that MBTI dimensions significantly affect labeling behavior. Extending this to LLMs, we prompt four open-source models with MBTI personas and evaluate their outputs across three hate speech datasets. Our analysis uncovers substantial persona-driven variation, including inconsistencies with ground truth, inter-persona disagreement, and logit-level biases. These findings highlight the need to carefully define persona prompts in LLM-based annotation workflows, with implications for fairness and alignment with human values.

From Knowledge to Noise: CTIM-Rover and the Pitfalls of Episodic Memory in Software Engineering Agents

We introduce CTIM-Rover, an AI agent for Software Engineering (SE) built on top of AutoCodeRover (Zhang et al., 2024) that extends agentic reasoning frameworks with an episodic memory, more specifically, a general and repository-level Cross-Task-Instance Memory (CTIM). While existing open-source SE agents mostly rely on ReAct (Yao et al., 2023b), Reflexion (Shinn et al., 2023), or Code-Act (Wang et al., 2024), all of these reasoning and planning frameworks inefficiently discard their long-term memory after a single task instance. As repository-level understanding is pivotal for identifying all locations requiring a patch for fixing a bug, we hypothesize that SE is particularly well positioned to benefit from CTIM. For this, we build on the Experiential Learning (EL) approach ExpeL (Zhao et al., 2024), proposing a Mixture-Of-Experts (MoEs) inspired approach to create both a general-purpose and repository-level CTIM. We find that CTIM-Rover does not outperform AutoCodeRover in any configuration and thus conclude that neither ExpeL nor DoT-Bank (Lingam et al., 2024) scale to real-world SE problems. Our analysis indicates noise introduced by distracting CTIM items or exemplar trajectories as the likely source of the performance degradation.

Look Within or Look Beyond? A Theoretical Comparison Between Parameter-Efficient and Full Fine-Tuning

Parameter-Efficient Fine-Tuning (PEFT) methods achieve performance comparable to Full Fine-Tuning (FFT) while requiring significantly fewer computing resources, making it the go-to choice for researchers. We find that although PEFT can achieve competitive results on some benchmarks, its performance falls short of FFT in complex tasks, such as reasoning and instruction-based fine-tuning. In this paper, we compare the characteristics of PEFT and FFT in terms of representational capacity and robustness based on optimization theory. We theoretically demonstrate that PEFT is a strict subset of FFT. By providing theoretical upper bounds for PEFT, we show that the limited parameter space constrains the model's representational ability, making it more susceptible to perturbations. Experiments on 15 datasets encompassing classification, generation, reasoning, instruction fine-tuning tasks and 11 adversarial test sets validate our theories. We hope that these results spark further research beyond the realms of well established PEFT. The source code is in the anonymous Github repository\footnote{https://github.com/misonsky/PEFTEval}.

Why Do More Experts Fail? A Theoretical Analysis of Model Merging

Model merging dramatically reduces storage and computational resources by combining multiple expert models into a single multi-task model. Although recent model merging methods have shown promising results, they struggle to maintain performance gains as the number of merged models increases. In this paper, we investigate the key obstacles that limit the scalability of model merging when integrating a large number of expert models. First, we prove that there is an upper bound on model merging. Further theoretical analysis reveals that the limited effective parameter space imposes a strict constraint on the number of models that can be successfully merged. Gaussian Width shows that the marginal benefit of merging additional models diminishes according to a strictly concave function. This implies that the effective parameter space becomes rapidly saturated as the number of merged models increases. Furthermore, using Approximate Kinematics Theory, we prove the existence of a unique optimal threshold beyond which adding more models does not yield significant performance improvements. At the same time, we introduce a straightforward Reparameterized Heavy-Tailed method (RHT) to extend the coverage of the merged model, thereby enhancing its performance. Empirical results on 12 benchmarks, including both knowledge-intensive and general-purpose tasks, validate our theoretical analysis. We believe that these results spark further research beyond the current scope of model merging. The source code is in the anonymous Github repository https://github.com/wzj1718/ModelMergingAnalysis.

Do We Know What LLMs Don't Know? A Study of Consistency in Knowledge Probing

The reliability of large language models (LLMs) is greatly compromised by their tendency to hallucinate, underscoring the need for precise identification of knowledge gaps within LLMs. Various methods for probing such gaps exist, ranging from calibration-based to prompting-based methods. To evaluate these probing methods, in this paper, we propose a new process based on using input variations and quantitative metrics. Through this, we expose two dimensions of inconsistency in knowledge gap probing. (1) Intra-method inconsistency: Minimal non-semantic perturbations in prompts lead to considerable variance in detected knowledge gaps within the same probing method; e.g., the simple variation of shuffling answer options can decrease agreement to around 40%. (2) Cross-method inconsistency: Probing methods contradict each other on whether a model knows the answer. Methods are highly inconsistent -- with decision consistency across methods being as low as 7% -- even though the model, dataset, and prompt are all the same. These findings challenge existing probing methods and highlight the urgent need for perturbation-robust probing frameworks.

Understanding Gated Neurons in Transformers from Their Input-Output Functionality

Interpretability researchers have attempted to understand MLP neurons of language models based on both the contexts in which they activate and their output weight vectors. They have paid little attention to a complementary aspect: the interactions between input and output. For example, when neurons detect a direction in the input, they might add much the same direction to the residual stream ("enrichment neurons") or reduce its presence ("depletion neurons"). We address this aspect by examining the cosine similarity between input and output weights of a neuron. We apply our method to 12 models and find that enrichment neurons dominate in early-middle layers whereas later layers tend more towards depletion. To explain this finding, we argue that enrichment neurons are largely responsible for enriching concept representations, one of the first steps of factual recall. Our input-output perspective is a complement to activation-dependent analyses and to approaches that treat input and output separately.