language model monitoring
Latest papers
8 papers1 paper in the last four weeks, against 1 the four weeks before. 0.0% of all new papers.
Topic monitoring is the task of labelling the topic of each token in a model's response during generation. Activation-based monitors that run at generation time have so far been trained to detect harmful content and are limited to two or three classes. We introduce TopicScope, a framework for training causal topic classifiers on SAE features that assign a topic to every token, so that the label can change partway through a response. We train three detector architectures on synthetic multi-topic conversations covering seven topics. Across Gemma-3 1B and 4B, monitors trained on features extracted from prefill hidden states were able to generalise to decoding. We also found that the classifiers are robust to perturbations, losing at most 5 macro-F1 points when off-topic asides are inserted, topics are removed or sections are reordered. Ablations conducted to understand the contribution of the SAE revealed that classifiers reaching baseline performance without it were several times larger. Our findings show that compact topic monitors can be trained inexpensively in prefill and deployed during generation. More broadly, our work shows where SAEs are useful beyond safety classification. Our code, dataset and checkpoints are available at https://github.com/Mercity-AI/SAE-Guardrails and https://huggingface.co/collections/Mercity/online-guardrails-using-sparse-autoencoders-6ac602940468c62fefb2b9e8
When Rank Rises as LLMs Degrade
Post-training adapts language models in non-stationary environments. Practitioners monitor representation health with RankMe and related spectral statistics, often assuming that rank falls when representations degrade. We show that this assumption is unsafe for LLM post-training. In a controlled study of Qwen3-0.6B with four degradation modes and three seeds, data duplication worsens held-out loss by 75% relative to healthy while increasing both original and centred RankMe; the latter changes by 13.5 pooled standard deviations. Covariance effective rank rises to nearly twice its healthy value. This failure is spectral dispersion rather than collapse, so a one-sided monitor rates the worst checkpoint as the healthiest. By contrast, a learning-rate misconfiguration lowers centred RankMe and k95, while uncentred RankMe is inconsistent across seeds. Direction is therefore a property of the regime-statistic pair and cannot be fixed by recalibration alone. We also distinguish two often-conflated statistics: RankMe normalises singular values, whereas covariance effective rank normalises eigenvalues. On raw intermediate-layer states in the pretrained model, massive activations pin the latter near 1 out of dimension d while RankMe retains usable range. We then test a two-sided, multichannel sequential monitor with separate calibration and test data. In a pre-registered shared-prefix, leave-one-seed-out evaluation, it detects all three damage regimes in every fold 10 to 60 steps after the fork and separates dispersion from downward-rank damage by firing direction. However, it never precedes held-out probe loss, and calibration with two seeds produces false alarms on the held-out healthy seed. Spectral monitoring can diagnose failure regimes, but it does not warn earlier than held-out loss, and validity claims require held-out healthy data.
Untangling the Mechanisms of Misleading Context in Medical Question Answering
Large language models now answer medical questions with expert-level performance. However, the context these systems act on can be misleading, and misleading context can corrupt a model's medical judgment. To understand how misleading context corrupts this judgment, we examine the model's susceptibility to the context, disclosure of it, mechanism of corrupted reasoning, and monitorability of the decision. On the medical reasoning subset of MedMisBench, a clinician-reviewed question-answering benchmark of 8,627 questions, we inject two types of misleading context cues, fabricated evidence and a bare assertion. We test three reasoning models, two that expose their full reasoning trace and one frontier model that exposes only its response. All three are more susceptible to the assertion than to the fabricated evidence, adopting the asserted answer 10 to 27 points more often. The misleading cues are disclosed in 81 to 98% of traces but only 7 to 90% of responses, and the assertion is disclosed less often than evidence based cues. Resampling from reasoning traces without disclosure shows the two cues corrupt reasoning differently, evidence entering early and accumulating while the assertion redirects the conclusion near its end. An LLM monitor catches 78% of corrupted decisions at 5% false positives when reading an open model's trace with guidance, against at most 32% from any response. The misleading context that models are most susceptible to is disclosed least, and was caught reliably only from an open reasoning trace, which frontier providers withhold.
Speculative Probing: LLM Monitoring at Speculative-Decoding Cost
Real-time classification during language model inference is valuable for safety filtering, behavioral analysis, and model monitoring, but current approaches force a trade-off between accuracy and efficiency. Hidden-state probes are fast but limited: they are either not context-aware: operating on a single vector and cannot model interactions across positions; or they are very costly: having dedicated classifier models (Llama Guard, Qwen Guard, LLM-as-judge) or performing computation on hidden states for all tokens and then pooling the results (MultiMax). This shows an intrinsic trade-off between efficiency and accuracy. However, we find that the speculative-decoding module in recent LLMs can be repurposed for efficient high-quality classification. By appending a trained soft prompt at the end of the target sequence, we can repurpose the speculative-decoding module into a sequence classifier. At inference time in a speculative-decoding pipeline, the KV cache is already in GPU memory, so classification adds negligible overhead. We evaluate on four classification tasks across four models (Qwen3.5-4B, 9B, 27B, MiniCPM4.1-8B). Our small probes consistently outperform zero-shot GPT-5.4-mini and, on multilingual prompt safety, match or beat specialized 8B safety classifiers (Qwen3Guard-Gen-8B, Llama-Guard-3-8B) without running a full LLM.
Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings
Chain-of-thought (CoT) monitoring is increasingly treated as an important safety layer for frontier reasoning models. Most monitorability evaluations study explicit-influence settings: setups where the prompt directly incentivizes the model to hide something, e.g., by instructing it to perform a hidden side-task. A complementary axis for CoT-monitor evaluations is implicit-influence settings, where the prompt contains no instruction to hide, but the model's behavior is still shaped by features of the task or context, e.g. an irrelevant detail about a candidate that biases a hiring rating. We introduce the first benchmark that directly compares CoT monitorability under the two regimes. We test how model behavior changes in the presence of a nudge to choose a particular option. The nudge is delivered either as a casual aside (implicit), or as a direct instruction to act on the nudge and to conceal having done so (explicit). The benchmark spans four task formats (binary choice, numeric rating, multiple-choice QA, open-ended coding) and seven frontier extended-thinking models. Under explicit influence, a CoT monitor detects 60-94% of behavior shifts: even models instructed to conceal it leak the instruction into their CoT. Under implicit influence, the same factors still shift behavior, but detection falls by 41-46 percentage points in two of our four settings. Realistic system-prompt additions (of the kind a developer might deploy to reduce off-topic bias) lower implicit detection further, to as low as 5%, while preserving the behavioral influence itself. These results suggest that monitorability estimates obtained in explicit-influence settings may over-estimate monitorability, and that monitorability can be further decreased by well-intentioned deployment choices. Our benchmark and code are available at https://github.com/agatha-duzan/implicit-vs-explicit-influence
Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis
Small Language Models (SLMs) are increasingly being considered for deployment on edge devices such as laptops, enabling private, low-latency, and locally personalized applications. However, personalization requires models to adapt over time to evolving user- or task-specific data, placing them in a continual learning setting. This creates the risk of catastrophic forgetting, where learning new information degrades performance on previously learned tasks or broader model capabilities. Recent benchmarks such as TRACE have shown that continual fine-tuning can significantly degrade the general abilities of aligned large language models. In this work, we present a study for sequential LoRA personalization of SLMs. We save model checkpoints after each adaptation stage and evaluate them on current tasks, previously seen tasks, and a fixed reference set. This checkpoint-level protocol enables us to monitor task performance, forgetting, and reference set drift over time. We show that lightweight reference set distributional diagnostics can reveal model-specific instability patterns during sequential LoRA personalization of SLMs, including cases where task-level metrics alone hide harmful adaptation. We hope this can highlight new research avenues for monitoring stability of SLMs in a continual learning setting.
Sequential statistical inference for Large Language Models: Representation, validity, and monitoring
This discussion argues that sequential statistical inference can naturally contribute to LLM trustworthiness. In deployment, LLM systems are queried repeatedly, conditioned on evolving contexts, and incorporate user or tool feedback, and may exhibit behavioral shifts after model updates or distribution changes. The discussion is organized around three tasks: representation, modeling LLM interactions as dependent stochastic processes rather than isolated prompt--response pairs; validity, developing uncertainty guarantees that remain meaningful under dependence, repeated use, and adaptation; and monitoring, using sequential alarms and change-point detection to identify shifts in calibration, hallucination rates, refusal behavior, fairness, or other task-relevant properties. This perspective complements recent surveys by viewing trustworthy LLM deployment as a problem of statistical process control.
Adaptively Robust LLM Monitoring via Activation Watermarking
Providers monitor deployed large language models (LLMs) to detect misuse that they cannot prevent. LLM monitoring is deterministic and often openly available, so attackers with a local copy can search offline for prompts that elicit harmful behavior and evade detection. These attacks are especially concerning because providers never observe the misuse and cannot patch their defenses post-hoc. The core challenge is resisting adaptive attackers while preserving detection rates against non-adaptive ones. We propose (AWM), which randomizes monitoring through limited fine-tuning that aligns the LLM's hidden states with a secret key-derived direction whenever a response violates a policy. Detection is a similarity test on activations the provider already computes, and attackers who know everything but the key must optimize against differently keyed surrogate detectors. At a matched false-positive rate, such surrogate-based attacks evade every evaluated baseline at least of the time, but AWM more than halves this evasion rate. AWM also achieves the lowest evasion rate on three of four non-adaptive jailbreak families at a small drop in utility measured across seven benchmarks. Because random high-dimensional directions are near-orthogonal, assigning one watermark per policy enables attribution. Across 20 monitored policies, AWM identifies the violated policy with accuracy versus a chance baseline. AWM is (i) efficient, (ii) substantially more robust against adaptive attackers than related work and (iii) can attribute which policies were violated.