Papers for

model developers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Large language models tracked by tracking how their knowledge changes

From Representations to Computational Trajectories: Intrinsic Fingerprinting for Large Language Model Provenance Verification

Abstract: Large language models (LLMs) are valuable digital assets, yet their weights can be easily modified after unauthorized acquisition, making model provenance difficult to verify. Model fingerprinting provides a promising solution for model ownership protection without modifying the model or embedding ownership information. However, parameter-based fingerprints can be sensitive to model transformations, limiting their robustness, while representation-based fingerprints may assign high similarity to unrelated models, limiting their uniqueness. To address these limitations, we propose TrFP, an intrinsic fingerprinting method based on computational trajectories. Instead of comparing model parameters or representations, TrFP captures how representations evolve across network depth and constructs compact fingerprints using permutation-invariant trajectory descriptors. Experiments across multiple LLM families show that TrFP reliably distinguishes related and unrelated models. It also remains robust under fine-tuning, pruning, quantization, model merging, parameter permutation, and adaptive attacks, while requiring less than 0.3 MiB of fingerprint storage. Overall, TrFP provides stronger provenance discrimination than several existing state-of-the-art methods, demonstrating the effectiveness of computational trajectories for reliable LLM provenance verification.

Mon 26 OctCryptography and Security
The gist
It can be hard to tell if a large language model (LLM) is the original or a copy that was changed without permission. The researchers offer a new way to create a unique “fingerprint” for each model by looking at how its understanding changes layer by layer, rather than just its stored information or output. This fingerprint stays reliable even if the model gets updated, trimmed, or mixed with others. Their method uses very little storage and works better than previous ways at telling models apart.
Open → 2610.34675v1

Harmful jailbreak tricks for one AI type work on another too

Harmful Responses Are Jailbreak Templates: Transferring Autoregressive Jailbreaks to Diffusion Language Models

Abstract: Diffusion large language models (dLLMs) enable flexible masked generation through bidirectional context modeling, but this capability also introduces a distinct jailbreak attack surface. We find that a harmful response to a given query naturally serves as an effective jailbreak template: simply masking a subset of its tokens can induce a dLLM to reconstruct harmful content. Based on this observation, we propose RMT-Jailbreak, a general framework that transfers existing autoregressive (AR) jailbreak attacks to dLLMs through their induced harmful responses. RMT-Jailbreak converts an AR-generated harmful response into a high-mask-ratio template using uniform token retention, and further applies prompt-based safety filtering to mask any remaining visible harmful content. The refined template is then provided to the target dLLM for bidirectional reconstruction. Extensive experiments across multiple AR jailbreak methods and target dLLMs demonstrate the effectiveness and generality of RMT-Jailbreak, revealing strong cross-paradigm transferability of jailbreak attacks.

Thu 22 OctMachine Learning
The gist
Diffusion large language models (dLLMs) can generate missing parts of text using clues from both before and after the gap. But this ability also lets people trick the models into giving harmful answers. The authors found that harmful responses themselves can be used as templates to make the model fill in the blanks with harmful content. They created a method called RMT-Jailbreak that takes harmful outputs from older autoregressive models and turns them into mask-based templates for attacking diffusion models. Experiments show this method works well, meaning jailbreak attacks can transfer between different AI generation approaches.
Open → 2610.31331v1

SkillEval reveals detailed skill profiles of large language models

SkillEval: Learning Interpretable Ability Profiles of LLMs via Cognitive Diagnosis Models

Abstract: Current evaluation of LLMs on benchmarks typically aggregates performance over individual examples into a single scalar score. This approach is overly coarse and ignores the diverse scope of skills covered by benchmark datasets. To address this limitation and better exploit benchmark data, we propose SkillEval, a new evaluation framework that performs cross-model, skill-level analysis. The framework integrates advances from two areas: (1) it leverages the meta-cognitive capabilities of LLMs to automatically discover interpretable skills from questions, and (2) it applies cognitive diagnosis models (CDMs) from psychometrics to estimate per-skill abilities of LLMs. We conduct experiments on 3,811 LLMs evaluated on 9,523 items across five benchmarks spanning diverse domains (MATH, BBH, GPQA, MuSR, IFEval). Compared to the traditional average-based evaluation paradigm, SkillEval produces reliable, more calibrated and informative yet concise model profiles that more accurately predict model behavior on unseen items. We also demonstrate the utility of SkillEval in two downstream tasks: (1) LLM routing, which dynamically assigns each query to a capable but efficient model, reducing inference cost by over 60% while achieving improved performance, and (2) Computerized Adaptive Testing (CAT), where an unseen model can be efficiently evaluated using a trained CDM. Furthermore, we show that these skill profiles provide useful insights, for example, how benchmarks share skills and how instruction-tuned models differ from their base models.

Tue 20 OctComputation and LanguageArtificial IntelligenceMachine Learning
The gist
Current assessments of large language models (LLMs) mostly give a single overall score, which misses the variety of different skills these models have. The authors introduce SkillEval, a method that breaks down LLM performance by specific skills using ideas from psychology to better understand how models perform. They tested SkillEval on thousands of models across many tasks and found it can predict performance more accurately than simple averages. This method also helps assign tasks to the best model for efficiency and offers new ways to evaluate and compare models.
Open → 2610.28767v1

Netflix reveals new way to measure recommendation impact accurately

Beyond Raw Engagement: A Counterfactual Observability Framework for Recommender Systems at Netflix

Abstract: Understanding the performance of large-scale recommender systems remains an underexplored challenge, especially for content creators and model developers. The raw engagement signals available to them, such as views and clicks, conflate content quality, model behavior, presentation bias, and audience reach, making it hard to attribute outcomes to the right cause. In this work, we present a general evaluation framework that enhances observability across multiple recommender systems at Netflix and demonstrate its effectiveness through several production deployments. The framework treats recommender-system observability as a counterfactual measurement problem: estimating what the recommender would have done, and what engagement would have followed, in the absence of a specific content item or model decision. We articulate three stakeholder-centered observability principles for content creators and model developers, and propose measurement methodologies covering bias reduction, relativity, and incrementality, applicable to both single-stage and cascading recommender systems and serving both audiences from a single measurement foundation.

Sat 19 SeptInformation RetrievalArtificial Intelligence
The gist
Netflix has a new way to figure out how well its recommendation system really works. Often, just looking at clicks or views doesn’t tell the full story because many things can influence those numbers. The authors created a method that estimates what would have happened if a certain show or algorithm choice hadn't been made. This helps creators and engineers understand the real reasons behind user engagement more clearly. It works across different recommendation stages and helps reduce bias in measuring results.
Open → 2609.22747v1

Backdoors in large language models separate trigger detection from control

LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders

Abstract: Even though backdoors in LLMs have been a growing concern, their inner workings are still under heavy scrutiny. Trigger-based backdoors are easy to define behaviorally, a rare input that makes the model switch to a chosen response pattern, but the mechanism between triggers and their responses is less clear. We study this mechanism in a controlled, harmless language-switching setting, where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German. For this, we train sparse autoencoders (SAEs) across layers and transformer components, then compare triggered prompts with translation and pretraining controls to identify trigger-relevant feature directions. We show how SAE features separate triggered prompts from controls with near-perfect F1, but features that detect the trigger do not necessarily control the behavior. In intervention tests, attention and MLP features often fire reliably on triggered prompts, making them good detectors, but ablating them rarely suppresses the language switch and activating them rarely induces it. In contrast, residual-stream features can suppress triggered generation when ablated, and some selected features can induce target-language continuations without the trigger. In short, these token-trigger mechanisms decompose into distinct SAE feature directions, with separate features for trigger detection, residual-stream propagation, and later language tracking. This role-level decomposition is the part most likely to transfer to other trigger-based backdoors, even when the payload, layers, or circuit locations differ.

Mon 7 SeptComputation and Language
The gist
Some large language models can be tricked by special hidden phrases, called triggers, to respond in unusual ways. This paper looks inside the models to understand how these triggers cause a switch in the model’s language output, using a harmless test of switching English to French or German. The authors find that the model uses different internal features: some detect the trigger, while others control the switch in language. This separation helps explain how backdoors operate and suggests ways to find and control them in other cases.
Open → 2609.07746v1