The Value Axis: Language Models Encode Whether They're on the Right Track

Researchers have identified a linear encoding inside the Qwen3-8B language model that estimates whether its current strategy is likely to succeed, a signal they call the “value axis.” The finding, submitted on 15 Jun 2026, suggests models internally track expected goal success and modulate their confidence accordingly [1][2]. The work, posted to arXiv, investigates whether language models internally track the value of their current trajectory, defined as the likelihood that their ongoing strategy will achieve their goals [1]. Using synthetic, in-context reinforcement learning data, the authors constructed a “value” axis for Qwen3-8B [2]. Activations along this axis distinguish between high and low verbalized confidence, rollouts without and with backtracking, and correct versus corrupted code [1][2]. Steering the model toward high value causally suppresses self-correction and reduces explanatory verbosity, while steering toward low value induces backtracking and exploration [1][2]. The team also demonstrated that direct preference optimization can increase the internal value of rewarded behaviors — such as using a certain word — causing the model to act more confidently after exhibiting them [1][2]. In real-world settings, the researchers found that Qwen assigns low value to politically sensitive chat queries after post-training and that supervised fine-tuning increases internal confidence within the training domain [1][2]. The results indicate that language models linearly encode an estimate of expected goal success that modulates their confidence in pursuing a direction [1][2]. The study adds to a growing body of interpretability research that probes how large models represent internal states. While the primary paper focuses on a single model, the methodology opens a path for examining whether similar value-tracking mechanisms exist across architectures. The use of synthetic reinforcement learning data to isolate the axis suggests that such signals may be induced or amplified by training procedures, rather than emerging solely from pretraining on natural language.