Output Vector Editing for Memorization Mitigation in Large Language Models

Large language models memorize and reproduce sequences from their training data, creating privacy, copyright, and security risks. Existing neuron-level mitigation methods equate editing with zeroing out neuron activations, but the activation only controls whether a neuron engages; the output vector is what writes to the residual stream and, through superposition, encodes multiple features. We propose output vector editing, a constrained-optimization weight edit that locates a small set of MLP neurons responsible for a memorized continuation and minimally modifies their output vectors to introduce a distractor in vocabulary space, redirecting their residual-stream contributions while leaving activations unchanged. Evaluating on four models from 360M to 7B parameters (SmolLM-360M, OLMo-1B, OLMo-7B, Llama2-7B), we center on OLMo-7B (whose open weights and pretraining corpus enable systematic mining) and mine 6831 memorized sequences, achieving up to 87.9% suppression. The 2.7$\times$ gap over zero ablation on the same located neurons shows the suppression comes from the output-vector edit, not localization alone. Four edit modes span a spectrum from aggressive suppression to minimal redirection; in ensemble they cover 96.5% of memorized sequences, while our recommended single-mode configuration reaches 81.5% with no catastrophic locality failures. We further identify a mechanistic boundary at ${\sim}14%$ of sequences unreachable by MLP-only editing; while these failures are not attention-driven overall, ablating the top contributing attention heads recovers 60--64% of them, with stronger recovery on continuations that copy tokens from the prefix, positioning attention as a complementary fallback rather than a primary mechanism. Edit mode ordering and the success-locality trade-off transfer across all four models, with success rates scaling with model size rather than family.

Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models

Text-guided image editing with visual autoregressive (VAR) generators requires controlling both what the model samples and where the sampled change is written back into the image code. Existing VAR editors mainly operate on token streams, features, or flat next-token logits, leaving two native structures of bitwise-residual VAR models underused: the per-bit Bernoulli prediction head and the additive multi-scale residual code field from which the image is assembled. We propose BitResEdit, a training-free editor for bitwise-residual VAR generators such as Infinity. BitEdit performs source-negative guidance by tilting the post-CFG per-bit log-odds along a source--target contrast computed on a shared edited prefix, then projects each update into a closed-form Bernoulli-KL trust region around the clean CFG sampler. ResEdit converts the sampled bits into per-scale continuous-code residuals, gates them with a localization mask, and re-injects them through the generator's native sum-of-scales. Together they couple decision-time bit guidance with combination-time code composition, so masked-out latent features are preserved exactly by code arithmetic while localized, scale-aware edits are applied inside the target region. On PIE-Bench with Infinity-2B, BitResEdit attains the strongest text alignment among same-backbone VAR editors, improving CLIP on the edited region by +1.07 over the strongest prior editor while keeping background preservation competitive with it. Ablations show BitEdit and ResEdit play complementary roles in target alignment and background preservation.

Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?

Large language models (LLMs) offer a promising approach to machine translation (MT) for extremely low-resource languages by incorporating linguistic resources through in-context learning. However, LLMs often struggle to apply grammatical information effectively during translation. Inspired by recent progress in chain-of-thought reasoning, we investigate whether low-resource MT can benefit from structured intermediate steps of linguistic analysis and grammatical reasoning. We propose a pipeline for automatically generating step-by-step linguistic reasoning traces from Universal Dependencies treebanks, dictionaries, and grammar-rule banks. We evaluate these traces in three settings: in-context learning (ICL), supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT), on Xibe and Chintang as test cases. Our results show that linguistic reasoning traces are most effective as inference-time guidance: in ICL, reliable sentence-specific traces substantially improve translation performance across most models, languages, and metrics. In contrast, using the linguistic reasoning traces as training data yields smaller and less consistent gains, as models learn the trace format but often generate erroneous content. These findings suggest that LLMs can leverage grammatical information for low-resource MT when given reliable linguistic analyses, while learning to generate such analyses remains a major bottleneck.

Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models

Causal tracing of factual recall has been studied predominantly in dense transformer language models, where interventions localize information flow to layers or feed-forward modules. Sparse mixture-of-experts (MoE) language models introduce a sharper question: when a factual prediction is mediated by a routed MoE block, which routed expert contributions matter? We formulate expert-aware causal tracing for sparse MoE language models. Using CounterFact facts, we first corrupt the model's factual preference by adding noise to subject-token embeddings, and then test whether clean MoE-block outputs or clean expert-level updates restore the true-vs-foil logit contrast. For Qwen3-30B-A3B-Base, a layer sweep selects and validates layer 44, and expert-level tracing identifies L44E069 as an expert repeatedly selected in the clean run whose held-out patch outperforms other active same-layer expert patches. For Mixtral-8x7B-v0.1, layer-level tracing validates a mid-layer signal, but the signal is not localized to the selected singleton expert; a coalition check instead recovers it with routed multi-expert updates. These results suggest that MoE factual tracing can be made expert-aware, while also showing that expert-level localization is model- and protocol-dependent rather than universal.

ReverseMath: Answer Inversion for Scalable and Verifiable Mathematical Problem Generation

Mathematical reasoning benchmarks are vital for evaluating large language models (LLMs), but many are static and repeatedly exposed through public evaluation and training pipelines, making it difficult to separate genuine reasoning from memorization. Meanwhile, manually constructing new math problems with reliable answers remains costly. We introduce ReverseMath, a scalable method for generating new math problems through answer inversion. Given a problem and its answer, ReverseMath masks a numerical value in the original problem, treats the original answer as a known condition, and rewrites the problem so that the masked value becomes the new answer. The generated problem reverses the original input-output relation, making its answer known by construction. We study ReverseMath for both evaluation and training. For evaluation, paired original/reversed problems reveal substantial behavioral shifts: models sometimes fail on reversed problems and even incorrectly output the original answer, suggesting memorization-like behavior. For training, ReverseMath provides automatically labeled reversed problems as data augmentation for reinforcement learning (RL). Experiments show that including ReverseMath-generated data improves mathematical reasoning performance across multiple benchmarks, demonstrating its value as both an analysis tool and a scalable source of verifiable training data.

Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs

Large reasoning models (LRMs) achieve strong mathematical reasoning performance in English, but remain much less reliable in many low- and medium-resource languages. This gap is often explained as a failure to understand non-English problem statements. We show that this view is incomplete: even when the problem is given in English, controlling the model's reasoning language can substantially reduce accuracy, suggesting that language also affects reasoning execution itself. To study this effect, we introduce DATG, a Directed Acyclic Trace Graph framework that maps reasoning traces to language-independent mathematical anchors and dependencies. This allows us to align target-language traces with reference DAGs and measure whether they cover required mathematical nodes, respect dependency edges, and avoid harmful mathematical actions. Experiments on the Qwen3 series across 12 languages show that non-English reasoning often suffers from reduced anchor coverage and weaker dependency fidelity, especially in low-resource languages. Motivated by this diagnosis, we propose Loop-Retry and Formula-Retry, two simple test-time controls targeting DATG-exposed failure modes, and show that they consistently improve target-language reasoning performance in low-resource languages.

ChunkFT: Byte-Streamed Optimization for Memory-Efficient Full Fine-Tuning

This work presents \textsc{ChunkFT}, a memory-efficient fine-tuning framework that reformulates full-parameter fine-tuning around a dynamically activated working set. \textsc{ChunkFT} enables gradient computation for arbitrary sub-tensors without modifying the network architecture, providing an algorithmic foundation for optimizing arbitrary sub-networks while avoiding standard dense gradient computation. We provide a theoretical convergence analysis of \textsc{ChunkFT} in the deterministic setting. Empirically, we apply \textsc{ChunkFT} to fine-tune Llama 3-8B and Llama 3-70B using a single RTX 4090-24GB GPU and 2$\times$ H800-80GB GPUs, respectively. Full-parameter fine-tuning of a 7B model with a 1K input length requires only 13.72GB of GPU memory. The results demonstrate the effectiveness of \textsc{ChunkFT} in memory usage, running time, and optimization quality. Moreover, downstream evaluations on language understanding, mathematical reasoning, and MT-Bench show that \textsc{ChunkFT} consistently outperforms existing memory-efficient baselines. Notably, \textsc{ChunkFT} achieves performance comparable to, and in some cases exceeding, full-parameter fine-tuning. Our repository is on https://github.com/misonsky/chunk.

SMoA: Spectrum Modulation Adapter for Parameter-Efficient Fine-Tuning

As the number of model parameters increases, parameter-efficient fine-tuning (PEFT) has become the go-to choice for tailoring pre-trained large language models. Low-rank Adaptation (LoRA) uses a low-rank update method to simulate full parameter fine-tuning, which is widely used to reduce resource requirements. However, decreasing the rank encounters challenges with limited representational capacity. Theory suggests that LoRA fine-tuning with rank r converges toward the top r singular values of the pre-trained weight matrix. As the rank increases, more principal singular directions are preserved, which generally improves the model's performance. However, a larger rank also introduces more trainable parameters, leading to higher computational cost. To overcome this dilemma, we propose SMoA, a \textbf{S}pectrum \textbf{Mo}dulation \textbf{A}dapter that enlarges the accessible family of spectrum-aware updates under a smaller parameter budget. SMoA partitions the layer into multiple aligned spectral blocks and applies one in-block Hadamard-modulated low-rank branch to each diagonal block, yielding broader coverage of pretrained spectral directions. We provide theoretical analysis and empirical results on multiple tasks. In our experiments, SMoA improves average performance in the current lower-budget setting over LoRA and competitive LoRA-style baselines.

DiM\textsuperscript{3}: Bridging Multilingual and Multimodal Models via Direction- and Magnitude-Aware Merging

Towards more general and human-like intelligence, large language models should seamlessly integrate both multilingual and multimodal capabilities; however, extending an existing multimodal model to many languages typically requires expensive multilingual multimodal data construction and repeated end-to-end retraining. We study a training-free alternative: injecting multilingual capability into an existing multimodal model by composing residual updates in the shared language model backbone. The key challenge is that multilingual and multimodal updates are heterogeneous, reflecting different functional roles in the shared model. To address this, we propose Direction- and Magnitude-aware Multilingual Multimodal merging (DiM3), which selectively composes the two updates at each parameter dimension while preserving the original vision encoder and multimodal projector. Experiments on multilingual benchmarks in both text-only and vision-language settings, covering 57 languages across LLaVA- and Qwen-based backbones, show that DiM3 consistently outperforms existing merging baselines, substantially improves multilingual performance over the original multimodal model, and remains competitive with dedicated multilingual multimodal fine-tuning while largely retaining general multimodal ability. We further show that DiM3 can be directly applied to already trained multilingual multimodal models and still yield additional gains. Further interpretability analysis shows that DiM3 primarily reshapes intermediate-layer semantic representations, strengthening cross-lingual alignment under both text-only and multimodal inputs while preserving higher-layer task-sensitive structure. Our repository is on https://github.com/wzj1718/DiM3.

GKnow: Measuring the Entanglement of Gender Bias and Factual Gender

Recent works have analyzed the impact of individual components of neural networks on gendered predictions, often with a focus on mitigating gender bias. However, mechanistic interpretations of gender tend to (i) focus on a very specific gender-related task, such as gendered pronoun prediction, or (ii) fail to distinguish between the production of factually gendered outputs (the correct assumption of gender given a word that carries gender as a semantic property) and gender biased outputs (based on a stereotype). To address these issues, we curate \gknow, a benchmark to assess gender knowledge and gender bias in language models across different types of gender-related predictions. \gknow allows us to identify and analyze circuits and individual neurons responsible for gendered predictions. We test the impact of neuron ablation on benchmarks for disentangling stereotypical and factual gender (DiFair and the test set of GKnow), as well as StereoSet. Results show that gender bias and factual gender are severely entangled on the level of both circuits and neurons, entailing that ablation is an unreliable debiasing method. Furthermore, we show that benchmarks for evaluating gender bias can hide the decrease in factual gender knowledge that accompanies neuron ablation. We curate GKnow as a contribution to the continuous development of robust gender bias benchmarks.