Searcharxiv⌕ Search

arXiv subjects

Aditi Raghunathan

Publications and source records attributed to Aditi Raghunathan.

At least 19 recordsLinked to original sources

Regularization Amplifies Subliminal Learning

Recent work has shown that hidden biases can be transferred through data with no apparent semantic relationship to those biases. Can regularization limit this covert influence? By constraining adaptation, regularization might be expected to suppress behavioral changes on seemingly unrelated inputs. We systematically study this hypothesis in subliminal learning, as well as in related settings of covert influence, including phantom transfer and logit-linear selection. We find the opposite: regularization amplifies the transfer of hidden biases. In particular, subliminal learning is typically absent under full fine-tuning, but low-rank adaptation or explicit penalties can induce strong transfer. Phantom transfer and logit-linear selection show similar trends. Adding more data, an implicit form of regularization, also strengthens covert influence. Together, these findings suggest that the simplicity biases induced by regularization may facilitate covert influence. We further hypothesize that the level of regularization that maximizes covert transfer depends on task complexity.

cs.LG↗

Is Persona Generalization Predictable?

Persona generalization occurs when fine-tuning a language model in one domain induces behavioral changes in others -- for example, the broadly unsafe behavior observed in emergent misalignment. We discover a surprising "on-the-line" phenomenon in persona generalization: in-domain and out-of-domain persona rates across various fine-tuning hyperparameters follow an approximately probit-linear relationship, and the specific probit-linear relationship is surprisingly shared even across model families and model sizes, including Qwen and Llama. Training choices that leave the set of trainable parameters unchanged, such as learning rate, weight decay, and training duration, largely move models along the same line, reducing transfer mainly by weakening in-domain learning. Changing which layers are fine-tuned, however, can change the slope. For unsafe behavior, fine-tuning only later layers yields shallower slopes, retaining strong in-domain learning while reducing out-of-domain misalignment. This "on-the-line" phenomenon unifies unexpected, isolated observations of persona generalization within a quantitative framework for predicting and controlling behavioral transfer.

cs.CL↗

Pretraining Data Scale Reverses the Muon-AdamW Fine-Tuning Gap

Muon has recently emerged as a promising optimizer for pretraining large language models, offering a speedup over AdamW in minimizing pretraining loss. However, it remains unclear whether Muon's advantage persists after fine-tuning. In this paper, we study the downstream behavior of Muon and AdamW and uncover a surprising trend: beyond a sufficiently large pretraining data budget, AdamW outperforms Muon after fine-tuning. Moreover, as model size increases, AdamW overtakes Muon at progressively smaller pretraining token-parameter ratios. To explain this reversal, we view fine-tuning as a perturbation of the outputs and introduce the margin-to-spread ratio, which measures how far the correct logit lies above the wrong ones relative to how widely the wrong logits are spread. In language models, Muon has the larger margin-to-spread ratio at small data budgets and AdamW at large ones, mirroring the fine-tuning trend. In a linear associative memory, we prove that after one update the margin-to-spread ratio of gradient descent gains on that of Muon as the data budget grows, and simulations show that this trend persists after extended training.

cs.LG↗

Annotations Mitigate Post-Training Mode Collapse

Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.

cs.CL↗

Exact Unlearning of Finetuning Data via Model Merging at Scale

Approximate unlearning has gained popularity as an approach to efficiently update an LLM so that it behaves (roughly) as if it was not trained on a subset of data to begin with. However, existing methods are brittle in practice and can easily be attacked to reveal supposedly unlearned information. To alleviate issues with approximate unlearning, we instead propose SIFT-Masks (SIgn-Fixed Tuning-Masks), an exact unlearning method based on model merging. SIFT-Masks addresses two key limitations of standard model merging: (1) merging a large number of tasks can severely harm utility; and (2) methods that boost utility by sharing extra information across tasks make exact unlearning prohibitively expensive. SIFT-Masks solves these issues by (1) applying local masks to recover task-specific performance; and (2) constraining finetuning to align with a global sign vector as a lightweight approach to determine masks independently before merging. Across four settings where we merge up to 500 models, SIFT-Masks improves accuracy by 5-80% over naive merging and uses up to 250x less compute for exact unlearning compared to other merging baselines.

cs.LG↗

Large Flow Language Models

Continuous flow language models offer a path to parallel generation with tools for control and acceleration unavailable to discrete diffusion, yet effective techniques for building them at the billion-parameter scale remain less established. We present Lafa (Large flow language model), a fast, high-performing flow language model scaled to 8B parameters in a data-efficient manner. Lafa is a blockwise flow model with causal attention on the prefix and bidirectional, asynchronous flow within each block. This design enables efficient training by adapting pretrained autoregressive (AR) models and fast inference via prefix caching and blockwise parallel decoding. We train Lafa at 0.5B, 7B, and 8B scales with fewer than 30B tokens, more efficient than previous diffusion models, and evaluate it on instruction following, reasoning, and coding, where the learned flow matches or surpasses the quality of AR and can be distilled into flow maps for one- and few-step generation. Notably, flow training preserves or even improves pretrained AR capabilities via a multi-token prediction effect, which we leverage in a self-speculative decoding scheme that verifies flow-based drafts autoregressively within a single model. This yields 2.54x the throughput of an AR model at comparable quality, outperforming multi-token prediction and discrete diffusion methods. Our findings show that continuous flows can serve as a scalable foundation for language modeling.

cs.CL↗

Scalable Weak-to-Strong Verifier Hardening for Agent Environments

Agent environments are essential for large language model (LLM) development. They use programmatic verifiers to turn complex outcomes into rewards. Verifiers can be attacked into awarding full credit to cheating agents without completing the intended task, causing inaccurate evaluation and misalignment in training. As agent environments scale in numbers and difficulty, so must the repairing of exploitable verifiers. To this end, we ask whether today's model with scaled compute can automatically harden agent environments to prevent potentially stronger future models from reward hacking. We introduce a scalable, fully automatic approach to weak-to-strong verifier hardening. It first sets up an isolated verifier sandbox that agents cannot directly modify, then uses a hacker-fixer loop to find and repair remaining verifier vulnerabilities. We apply this loop to Terminal Bench, KernelBench, and SETA, substantially reducing attack success across these agent environments. The iterative design scales with the hardening budget, producing progressively stronger defenses as the number of loop iterations increases and outperforming the static, non-iterative baseline. An earlier version of our hacker-fixer loop has been adopted by the official Terminal Bench repository as a way to audit exploits in its environments. Project page: https://verifier-hardening.github.io/.

cs.CR↗

Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning

A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wisdom is that on-policy training is a prerequisite for continual learning. In practice, however, data containing new knowledge or capabilities are often off-policy. While methods such as on-policy self-distillation (OPSD) try to bridge this gap by converting off-policy data into on-policy signal, they have been shown to cause reasoning collapse. In this paper, we show that off-policy merging beats OPSD for continual learning. We first show that SFT learns a useful signal from new data, but naively applying its update interferes with existing capabilities. We reduce this interference with a simple recipe we term grafting, which changes where the update is learned and how it is applied: (1) learning the update on an earlier donor checkpoint, ideally even before the end of pretraining, and applying the weight update to the post-trained model; (2) scaling the weight update, equivalent to a form of model merging; and (3) optionally, masking the most sensitive update directions when the new data distribution is far from the post-trained model. Across continual learning settings including (1) distilling from expert traces, (2) self-improvement with STaR and Pedagogical RL, and (3) injecting knowledge after pretraining cutoff, grafting Pareto-dominates both SFT and OPSD in new-task and old-task performance, while avoiding expensive on-policy sampling. Therefore, our work challenges on-policy training as a necessity for continual learning on RL-trained models.

cs.LG↗

Visual General Intelligence: A White Paper

This paper reconsiders intelligence from a vision-centered perspective and examines whether intelligence emerging from visual experience and learning may provide a pathway toward AGI. In the language domain, beginning with the introduction of the Transformer architecture, the GPT series has demonstrated transfer to unseen tasks through autoregressive language modeling on web-scale text combined with aggressive scaling. This raises a natural question, namely, what capabilities and forms of intelligence can emerge from visual modalities such as images, videos, and geometry? In this paper, we discuss whether visual intelligence can serve as a pathway toward AGI, referred to in this paper as visual general intelligence (VGI), by bringing together contributors from diverse standpoints and affiliations. Our aim is not to offer a single definition of visual intelligence, but to clarify the principles that computer vision should pursue in the AGI era, the visual input modalities, the benchmarks, the learning paradigms, and the relationship between vision, when taken as the core, and other modalities such as language.

cs.CV↗

Jailbreaking in the Haystack

Recent advances in long-context language models (LMs) have enabled million-token inputs, expanding their capabilities across complex tasks like computer-use agents. Yet, the safety implications of these extended contexts remain unclear. To bridge this gap, we introduce NINJA (short for Needle-in-haystack jailbreak attack), a method that jailbreaks aligned LMs by appending benign, model-generated content to harmful user goals. Critical to our method is the observation that the position of harmful goals play an important role in safety. Experiments on standard safety benchmark, HarmBench, show that NINJA significantly increases attack success rates across state-of-the-art open and proprietary models, including LLaMA, Qwen, Mistral, and Gemini. Unlike prior jailbreaking methods, our approach is low-resource, transferable, and less detectable. Moreover, we show that NINJA is compute-optimal -- under a fixed compute budget, increasing context length can outperform increasing the number of trials in best-of-N jailbreak. These findings reveal that even benign long contexts -- when crafted with careful goal positioning -- introduce fundamental vulnerabilities in modern LMs.

cs.CR↗

Understanding and Mitigating Premature Confidence for Better LLM Reasoning

Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps, limiting the gains from additional test-time compute. Improving reasoning quality directly would require process reward models, but the step-level annotations needed to train them are expensive and scarce. We find such a signal in how the model's confidence evolves during reasoning: premature confidence, the tendency to commit to an answer early and use the remaining tokens to rationalize it, strongly predicts flawed reasoning across tasks and model scales. We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early -- rewarding gradual confidence growth and penalizing early commitment, with no external labels or reward models. The method improves accuracy and reasoning quality from 1.5B to 8B parameters across arithmetic (Countdown), math (DAPO, AIME), and science (ScienceQA): on Countdown, accuracy improves 3.2x (+42.0pp) and flawed reasoning drops 48pp; on AIME, Pass@64 improves 6.6pp. Consistent with this mechanism, the method also improves faithfulness: on a safety benchmark, our models more transparently surface misleading content in their reasoning traces rather than concealing it. Controlled experiments reveal that the problem and its remedy scale together: premature confidence grows with model size and task difficulty, and so do the gains from addressing it.

cs.AI↗

Early Data Exposure Improves Robustness to Subsequent Fine-Tuning

How can we train models whose post-trained capabilities survive subsequent fine-tuning? Rather than focusing on downstream interventions to mitigate forgetting of upstream capabilities, we study how upstream training choices - that is, the manner in which a capability is acquired - shape how robustly that capability is retained. We investigate this question in a controlled three-stage language-model pipeline: pretraining, post-training to acquire a target capability, and downstream fine-tuning on a new objective. Across 135M and 1B models, two post-training domains, and two downstream fine-tuning tasks, we find that immediate post-training performance does not reliably predict retention after subsequent fine-tuning: training recipes that look equivalent immediately after post-training can retain the target capability very differently after subsequent fine-tuning. In particular, early exposure - mixing post-training data into pretraining - consistently improves the frontier between retained upstream performance and downstream performance. In compute-matched experiments, where the target data must be allocated between pretraining and post-training, we find that the optimum lies at neither extreme. Together with our other empirical and theoretical findings, this supports the view that post-training drives immediate specialization while early exposure improves robustness to later forgetting. Replay and dropout, typically used to mitigate forgetting as it occurs during fine-tuning, provide complementary gains to early exposure when applied during post-training. Our findings suggest that robustness to subsequent fine-tuning should be treated as a first-class objective of upstream training, addressed preventatively through choices like early exposure rather than reactively during fine-tuning itself.

cs.LG↗

Posterior Refinement: Fast Language Generation via Any-Order Flow Maps

Non-autoregressive generation offers a powerful paradigm for iterative refinement, allowing models to recursively critique, erase and regenerate arbitrary subsets of tokens. However, existing non-autoregressive models fail to realize this potential. Masked Diffusion Models (MDMs) suffer from factorization error, causing sample quality to collapse when generating multiple tokens simultaneously. Flow Map Language Models (FMLMs) circumvent this bottleneck via joint sequence transport for excellent few-step generation, but sacrifice the inference-time flexibility of MDMs. We introduce FMLM+, a framework that bridges this gap by equipping FMLM with masking-style noise schedules. While generating the full sequence in a single step, FMLM+ simultaneously scores the global consistency of each token a posteriori. We leverage this to introduce Posterior Refinement, a novel inference-time refinement strategy that enables the model to adaptively self-correct its outputs, matching the performance of discrete baselines with 32x fewer NFEs. Across diverse benchmarks, we demonstrate that FMLM+ with Posterior Refinement improves the speed--quality tradeoff over both MDM and FMLM families, providing a scalable foundation for high-fidelity language modeling.

cs.CL↗

Natively Unlearnable Large Language Models

Unlearning aims to remove the influence of specific training data sources, but this has proved challenging because the contributions of different sources are entangled within the model. Isolating source contributions to disjoint parameters makes removal easier, though it obstructs joint learning across sources. We propose NULLs (Natively Unlearnable LLMs), a model class that satisfies the two opposing goals of isolating source-specific contributions and learning jointly across sources, by training a set of shared backbone neurons alongside a pool of sparsely activated sinks. During training, information specific to a source naturally concentrates in its sinks while information shared across sources accumulates in the backbone. A source is then unlearned at deployment by disabling its corresponding sinks, with no gradient updates and no access to the retained data. We show that NULLs scales to Wikipedia's ~6M articles, isolating each as an independent source. Unlearning a single article removes knowledge specific to it while preserving facts shared with semantically related articles, closely matching retraining from scratch. We note that unlearning with NULLs is also robust: in a case study of unlearning the Harry Potter books, NULLs resists both adversarial extraction and relearning that reverses post-hoc unlearning. Finally, NULLs preserves general language capabilities, matching a standard transformer on downstream benchmarks. Together, these results suggest that source-level unlearning need not be an afterthought. It can be built natively into LLM training while retaining the benefits of shared representation learning.

cs.LG↗

Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops

Agent benchmarks score submissions with outcome verifiers that are typically hand-written and brittle, leaving them open to reward hacking. We audit 1,968 tasks across five terminal-agent benchmarks and find 323 (16%) hackable by frontier models given only the task description. This corrupts both leaderboard rankings and RL training signal, yet the standard response is manual and reactive. We introduce the hacker-fixer loop, a method for building exploit-resistant verifiers without per-task manual patching. The loop alternates three LLM agents: a hacker tries to pass the verifier without solving the task, a fixer patches the verifier to reject each discovered exploit, and a solver confirms the patched verifier still admits legitimate solutions. The loop iterates: each patch reshapes what the verifier rewards, surfacing the next exploit. We further add verifier access, and let patches transfer across tasks, to broaden the exploits the loop discovers. On KernelBench, the loop drives the attack success rate from 62% to 0% on a held-out corpus of publicly reported exploits. We also find that weaker agents in the loop can defend against much stronger hackers: Gemini 3 Flash's loop drives the stronger Gemini 3.1 Pro and Claude Opus 4.7's attack success rate from 76% and 61% to 0% on KernelBench, and Gemini 3.1 Pro's from 39% to 17% on Terminal Bench across 77 tasks. We release Terminal Wrench (323 hackable environments, 3,632 hack trajectories) as a snapshot of the current attack surface, our patched verifiers, the exploits the loop discovered, and our implementation as a basis for future work.

cs.CR↗

Self-Trained Verification for Training- and Test-Time Self-Improvement

Self-improvement at scale has been a longstanding goal for reasoning models, and there are two natural places to do it: at test time, through verification-refinement (V-R) loops; and at training time, through self-training methods. Both are gated by the same bottleneck: the verifier. V-R loops stall when verifier scores inflate while accuracy stagnates, and when feedback is too generic to act on; self-training fails similarly when bad self-generated data are added to training. Better verification would unlock both, but the capability we want to train, i.e., catching self-generated errors, lacks training signal. To address this challenge, we propose self-trained verification (STV). Our key observation is that, while a model cannot catch these errors alone, it can when shown the reference solution. We turn this asymmetry into a supervision target and train the verifier to imitate a more informed version of itself. At test time, STV substantially improves V-R loops on hard problems, while alternatives (e.g., SFT, RL on verifier scores, and even meta-verifiers) do not. STV roughly doubles accuracy on hard math and lifts it 14x on scientific reasoning tasks (1.5% to 21%). At training time, we additionally train the generator using RL with STV verifier's feedback inside the V-R loop - a procedure we call verifier-in-the-loop training (ViL). Starting from an RL-converged generator, ViL yields a further 33% gain in pass@1. More notably, the generator's standalone pass@1, with no verifier at test time, climbs 30% relative past where standard RL had converged. Hence, the next frontier in reasoning on hard problems may lie in how we train for and with verification. Website: https://ar-forum.github.io/stv-webpage

cs.LG↗

Flow Map Language Models: One-step Language Modeling via Continuous Denoising

Language models based on discrete diffusion have attracted widespread interest for their potential to provide faster generation than autoregressive models. Despite their promise, these models typically produce samples whose quality sharply degrades in the few-step regime, preventing a dramatic speedup in practice. Here, we show that language models based on continuous flows over one-hot token embeddings can outperform discrete diffusion in both quality and speed. Importantly, our continuous formulation defines a unique flow map that can be learned directly for efficient few-step inference, a structure we show is unavailable to discrete methods. In this setting, we show that both the flow and its associated flow map can be learned with simple cross-entropy objectives that respect the simplex geometry of the data, and we identify three distinct choices for flow map distillation whose performance we compare in practice. Using these insights, we build a flow language model (FLM), a continuous flow that matches state-of-the-art discrete diffusion baselines on the One Billion Words (LM1B) and OpenWebText (OWT) datasets. We then distill FLM into a flow map language model (FMLM), whose one-step generation exceeds the 8-step quality of recent few-step discrete diffusion language models. Our work challenges the widely-held hypothesis that discrete noising processes are necessary for generative modeling over discrete modalities and paves the way toward accelerated language modeling at scale. Code is available at https://github.com/david3684/flm.

cs.CL↗

Base Models Look Human To AI Detectors

As AI-generated text enters the real-world at scale, institutions increasingly use commercial AI-text detectors, especially in education and academic-integrity workflows. We report a surprising empirical finding about such systems: when evaluated by GPTZero and Pangram, generated text from base models is often judged overwhelmingly human, whereas text generated by their instruction-tuned counterparts is not. Building on this observation, we propose Humanization by Iterative Paraphrasing (HIP), a detector-agnostic pipeline that minimally fine-tunes a base model into a paraphraser and applies it iteratively. Compared with the baselines we test, HIP yields a stronger trade-off between semantic preservation and detector evasion on commercial detectors. Across Llama-3 and Qwen-3 families, spanning model sizes from 0.6B to 70B, HIP consistently improves detector human-likeness. Our findings suggest that current detectors are tracking artifacts of instruction tuning and local context more than any invariant notion of machine-generated text. This, in turn, calls for detector designs that model these factors more explicitly.

cs.CL↗