Recursive self-improvement (RSI) in production agents involves transforming accumulated experience into reliable model updates while adapting to evolving business policies. We present CARE (Customer-service Agent Retrospective and Evolve), a grounded RSI framework that operationalizes this process in customer service. Its goal is to replace a frontier proprietary incumbent with an in-house mid-sized model that continues learning after deployment. After a one-time cold start on governed production traces, the current model serves production traffic alongside the incumbent, and its own failures drive the next update. A 24/7 inspection process establishes failures through evidence-blinded, independently reviewed judgments. Attribution separates model errors from defects in Skills, tools, or the harness, routing component failures to repair. Model-attributable trajectories are truncated at the first erroneous decision and paired with verified repairs. On-policy self-distillation (OPSD) then updates the model using the same model as teacher, with the diagnosis and repair supplied as privileged information unavailable to the student. To support continued evolution as business requirements change, training mixes failures from different Skill/SOP versions, each judged against the policy in force when it occurred. Candidates run an executable simulation test compiled from the latest SOP and then frozen, followed by offline inspection of the resulting traces. Promotion requires a combined simulation-and-inspection score at least at parity with the deployed version, confirmation through an online comparison with the incumbent, and human approval.
Xueqi Ma, Yanbei Jiang, Xingjun Ma, James Bailey, Sarah M. Erfani
Aligning large language models (LLMs) with multiple, potentially conflicting human values remains challenging, as existing multi-objective alignment methods often require costly fine-tuning and provide limited insight into the internal mechanisms underlying value-related behaviors. In this work, we investigate multi-value alignment from a mechanistic interpretability perspective by identifying value-sensitive attention heads (value heads) in LLMs. Using a probing-based value localization framework, we discover sparse sets of attention heads associated with diverse dimensions, spanning both human values (e.g., harmlessness, honesty) and stylistic preferences (e.g., coherence, complexity). Analyses across multiple LLM families reveal functional specialization and structured interactions among value heads, including conflicts between certain values. Targeted head ablations further induce substantial, value-specific behavioral changes, providing causal evidence of their functional importance. Building on these findings, we propose an inference-time activation steering strategy that enables controllable alignment with individual values or combinations of values without updating the underlying model parameters, supported by theoretical analysis. Experiments across multiple value alignment benchmarks demonstrate effective single- and multi-value control while largely preserving non-target model performance. Our findings connect mechanistic interpretability with flexible multi-value alignment, offering a foundation for more interpretable and pluralistic LLM behavior. The code is available at https://github.com/sihuo-design/ValueHeads.
Yanbei Jiang, Amr Keleg Mohamed bin Zayed University of Artificial Intelligence, Ryandito Diandaru, Jey Han Lau The University of Melbourne, Lea Frermann University of Melbourne, Biaoyan Fang Oracle, Fajri Koto Mohamed bin Zayed University of Artificial Intelligence
While the real world is inherently stochastic, Large Language Models (LLMs) are predominantly evaluated on single-round inference against fixed ground truths. In this work, we shift the lens to distribution alignment: assessing whether LLMs, when prompted repeatedly, can generate outputs that adhere to a desired target distribution, e.g. reflecting real-world statistics or a uniform distribution. We formulate distribution alignment using the attributes of gender, race, and sentiment within occupational contexts. Our empirical analysis reveals that off-the-shelf LLMs and standard alignment techniques, including prompt engineering and Direct Preference Optimization, fail to reliably control output distributions. To bridge this gap, we propose a novel fine-tuning framework that couples Steering Token Calibration with Semantic Alignment. We introduce a hybrid objective function combining Kullback-Leibler divergence to anchor the probability mass of latent steering tokens and Kahneman-Tversky Optimization to bind these tokens to semantically consistent responses. Experiments across six diverse datasets demonstrate that our approach significantly outperforms baselines, achieving precise distributional control in attribute generation tasks.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表