Interpretability

TL;DR for operators The central finding of Jonathan F. Jao and Andrew Kyle Lampinen’s Value Probes: Learned Value Functions for Beam-Search Red Teaming1 is that automated safety auditing benefits from two separable improvements: structured exploration of conversations and learned estimates of which conversations deserve further exploration. On Gemma 3 4B, beam search with randomly selected surviving branches reached a mean peak behavior-judge score of 0.74, compared with 0.51 for token-matched Petri auditing. Adding a Max Value Probe raised the score to 0.85. Similar patterns appeared at 12B and 27B. ...

TL;DR for operators If two parameterizations compute exactly the same model function, an interpretability conclusion presented as a property of that model needs a reason to remain valid when the internal coordinates change. The Linear Representation Hypothesis Needs a Group Action by Louie Hong Yao, Yuhao Li, and Shengchao Liu1 makes that requirement explicit. ...

TL;DR for operators A model can expose enough information for a linear classifier to recover sentiment without giving you a stable sentiment vector that will reappear when the data sample changes. Said et al.1 test this directly across English, Hausa, and Yoruba. Every tested probe remains above chance, yet final-token direction agreement ranges from 0.737–0.870 for English, 0.589–0.762 for Hausa, and only 0.101–0.399 for Yoruba. Across repeated topic allocations, the English > Hausa > Yoruba ordering holds in all 77 complete final-token comparisons. ...

TL;DR for operators A latent-space monitor is valuable only if the internal signal it reads is sufficiently difficult for the monitored model to manipulate. Measuring Activation Control in Large Language Models tests that assumption directly.1 Across 25 open-weight instruction-tuned models, the authors find meaningful but coarse control over concept-related internal representations. Models can raise a concept signal, suppress it toward its ordinary baseline, order several requested intensity levels, and move signal toward broad regions of a sentence. They are much less successful at targeting a particular layer or restricting modulation to particular token groups. ...

TL;DR for operators When a model is midway through a difficult task, the expensive choice is often whether to leave it alone, verify it, or spend more inference on correction. Doing that for every request wastes compute and can damage answers that were already on track. Lihao Sun and colleagues show that the model’s internal state changes systematically as reasoning progresses, and that late-stage changes contain a useful warning signal for eventual failure.1 Their trajectory-based correctness features reach a seed-averaged best-layer ROC-AUC of about 0.85, versus 0.765 for the strongest reported LogitLens baseline and 0.649 for reasoning step count alone. ...

TL;DR for operators A system generating several candidate answers eventually needs a ranking decision: which response should be shown, which should be discarded, and which should receive additional review. A reward model commonly reduces that decision to one scalar score derived from the prompt and response text. Oprea and Bâra test whether that score improves when the model is also given four explicit signals—response length, toxicity, refusal behavior, and prompt-response semantic similarity—and allowed to interpret those signals jointly with the text representation.1 On Anthropic HH-RLHF, the answer is consistently yes across ten evaluated model configurations. The strongest DeBERTa-v3 reward model moves from 0.74 to 0.84 ROC-AUC and from 0.72 to 0.83 pairwise accuracy. ...

TL;DR for operators Teams are easy to label. Understanding who actually listens to whom is harder. Hong’s paper on learned coordination conventions proposes a diagnostic for inspecting how cooperative reinforcement-learning agents route information between predefined roles.1 The central move is architectural: place role labels in both the querying agent’s representation and each ally’s representation, then use cross-attention to expose a role-to-role routing matrix. ...

TL;DR for operators A policy, prompt, adapter, steering vector, or internal patch can make a model look more orderly. That does not mean it controls the model. The paper’s central distinction is brutal and useful: order is visible structure; control is validated movement through the right receiver under the right conditions, with side effects bounded.1 ...

The current AI industry has a strangely theatrical relationship with intelligence. We judge models by the visible performance: the answer they print, the image they reconstruct, the attention map they expose, the number of reasoning steps they perform, the architectural flourish in the diagram. If the output looks sophisticated, we call the system capable. If the output looks wrong, we assume the capability is missing. This is convenient, measurable, and often completely misleading. Naturally, it is popular. ...

Most companies do not actually want an AI system that “thinks longer.” They want one that knows when extra thinking is worth the bill. That distinction is becoming more important. Reasoning models are moving from demo-stage math puzzles into document review, financial research, compliance analysis, customer support escalation, and agentic workflows. In these settings, reasoning has three costs: latency, compute, and misplaced confidence. A model that spends 30 seconds producing an elegant wrong answer has not reasoned. It has performed expensive theatre. Very fluent theatre, admittedly. ...