Keertana Chidambaram | Searcharxiv

arXiv2026

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

Chain-of-thought (CoT) monitoring is a safety strategy where the reasoning of a large language model "actor" is inspected by a "monitor" (often another language model) for signs of unsafe planning, deception, or misalignment. We find that planting harmful but benign-sounding reasoning in the actor's context can steer it to perform adversarial actions while evading monitors, an attack we term "plan injection". We initially discover this attack in the multiple-choice question-answering monitorability setting proposed by Lanham et al. (2023), using the investigator-agent elicitation framework of Li et al. (2025). We generalize the attack and show that the discovered behavior scales to harder tasks (achieving 25-33% monitor evasion rates across different monitorability benchmarks) and larger models such as DeepSeek-R1. Across the settings we study, actor models not only follow injected plans but also paraphrase them as their own reasoning, without explicit attribution to the injections. Finally, we find cases where extra monitor resources cause harm - giving the monitor access to the injected plan drops detection by as much as 50% in the Bio-Math task and in a case study on monitor reasoning budget, we find transcripts where additional thinking tokens are spent rationalizing the injected plan rather than flagging it.

cs.AI↗

arXiv2026

Pigeonholing: how bad prompts hurt models, causing collapse and mistakes

While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.

cs.CL↗

arXiv2026

Exponential Reward Weighting for Fine-Tuning Generative Recommenders under Sparse and Noisy Feedback

In recommendation systems, users interact with only a small fraction of a vast item catalog, producing feedback that is both sparse and noisy. This challenges post-training generative recommenders: reward models trained from logged interactions often fail to generalize, while directly optimizing imperfect rewards can lead to reward over-optimization. We propose Exponential reward-weighted fine-tuning (Exp-RSFT), where each logged interaction is weighted by $\exp(r/λ)$, avoids this failure by optimizing directly on the logged rewards, with the temperature $λ$ regularizing against their noise. We theoretically show that Exp-RSFT's suboptimality decomposes into two costs: a coverage cost arising from limitations of the logging policy and a noise cost from imperfect feedback. The temperature $λ$ balances these competing effects, yielding an optimal tradeoff between exploiting high-reward behavior and robustness to noise. Across three public benchmarks and a large-scale industrial dataset, we verify this theoretical prediction: performance follows an inverted-U trend as a function of $λ$, while PPO and DPO often over-optimize unreliable reward models and degrade recommendation quality. Exp-RSFT consistently improves ranking performance without requiring online exploration or preference data.

cs.IR↗

arXiv2026

Robust Post-Training for Generative Recommenders: Why Exponential Reward-Weighted SFT Outperforms RLHF

Aligning generative recommender systems to user preferences via post-training is critical for closing the gap between next-item prediction and actual recommendation quality. Existing post-training methods are ill-suited for production-scale systems: RLHF methods reward hack due to noisy user feedback and unreliable reward models, offline RL alternatives require propensity scores that are unavailable, and online interaction is infeasible. We identify exponential reward-weighted SFT with weights $w = \exp(r/λ)$ as uniquely suited to this setting, and provide the theoretical and empirical foundations that explain why. By optimizing directly on observed rewards without querying a learned reward model, the method is immune to reward hacking, requires no propensity scores, and is fully offline. We prove the first policy improvement guarantees for this setting under noisy rewards, showing that the gap scales only logarithmically with catalog size and remains informative even for large item catalogs. Crucially, we show that temperature $λ$ explicitly and quantifiably controls the robustness-improvement tradeoff, providing practitioners with a single interpretable regularization hyperparameter with theoretical grounding. Experiments on three open-source and one proprietary dataset against four baselines confirm that exponential reward weighting is simple, scalable, and consistently outperforms RLHF-based alternatives.

cs.LG↗

arXiv2025

Direct Preference Optimization With Unobserved Preference Heterogeneity: The Necessity of Ternary Preferences

Reinforcement Learning from Human Feedback (RLHF) has become central to aligning large language models with human values, typically by first learning a reward model from preference data which is then used to update the model with reinforcement learning. Recent alternatives such as Direct Preference Optimization (DPO) simplify this pipeline by directly optimizing on preferences. However, both approaches often assume uniform annotator preferences and rely on binary comparisons, overlooking two key limitations: the diversity of human evaluators and the limitations of pairwise feedback. In this work, we address both these issues. First, we connect preference learning in RLHF with the econometrics literature and show that binary comparisons are insufficient for identifying latent user preferences from finite user data and infinite users, while (even incomplete) rankings over three or more responses ensure identifiability. Second, we introduce methods to incorporate heterogeneous preferences into alignment algorithms. We develop an Expectation-Maximization adaptation of DPO that discovers latent annotator types and trains a mixture of LLMs accordingly. Then we propose an aggregation algorithm using a min-max regret fairness criterion to produce a single generative policy with equitable performance guarantees. Together, these contributions establish a theoretical and algorithmic framework for fairness and personalization for diverse users in generative model alignment.

cs.LG↗

arXiv2025

Direct Preference Optimization with Unobserved Preference Heterogeneity: The Necessity of Ternary Preferences

Reinforcement Learning from Human Feedback (RLHF) has become central to aligning large language models with human values, typically by first learning a reward model from preference data which is then used to update the model with reinforcement learning. Recent alternatives such as Direct Preference Optimization (DPO) simplify this pipeline by directly optimizing on preferences. However, both approaches often assume uniform annotator preferences and rely on binary comparisons, overlooking two key limitations: the diversity of human evaluators and the limitations of pairwise feedback. In this work, we address both these issues. First, we connect preference learning in RLHF with the econometrics literature and show that binary comparisons are insufficient for identifying latent user preferences from finite user data and infinite users, while (even incomplete) rankings over three or more responses ensure identifiability. Second, we introduce methods to incorporate heterogeneous preferences into alignment algorithms. We develop an Expectation-Maximization adaptation of DPO that discovers latent annotator types and trains a mixture of LLMs accordingly. Then we propose an aggregation algorithm using a min-max regret fairness criterion to produce a single generative policy with equitable performance guarantees. Together, these contributions establish a theoretical and algorithmic framework for fairness and personalization for diverse users in generative model alignment.

cs.AI↗

arXiv2025

Sequential Decision Making with Expert Demonstrations under Unobserved Heterogeneity

We study the problem of online sequential decision-making given auxiliary demonstrations from experts who made their decisions based on unobserved contextual information. These demonstrations can be viewed as solving related but slightly different problems than what the learner faces. This setting arises in many application domains, such as self-driving cars, healthcare, and finance, where expert demonstrations are made using contextual information, which is not recorded in the data available to the learning agent. We model the problem as zero-shot meta-reinforcement learning with an unknown distribution over the unobserved contextual variables and a Bayesian regret minimization objective, where the unobserved variables are encoded as parameters with an unknown prior. We propose the Experts-as-Priors algorithm (ExPerior), an empirical Bayes approach that utilizes expert data to establish an informative prior distribution over the learner's decision-making problem. This prior distribution enables the application of any Bayesian approach for online decision-making, such as posterior sampling. We demonstrate that our strategy surpasses existing behaviour cloning, online, and online-offline baselines for multi-armed bandits, Markov decision processes (MDPs), and partially observable MDPs, showcasing the broad reach and utility of ExPerior in using expert demonstrations across different decision-making setups.

cs.LG↗

arXiv2024

Personalized Adaptation via In-Context Preference Learning

Reinforcement Learning from Human Feedback (RLHF) is widely used to align Language Models (LMs) with human preferences. However, existing approaches often neglect individual user preferences, leading to suboptimal personalization. We present the Preference Pretrained Transformer (PPT), a novel approach for adaptive personalization using online user feedback. PPT leverages the in-context learning capabilities of transformers to dynamically adapt to individual preferences. Our approach consists of two phases: (1) an offline phase where we train a single policy model using a history-dependent loss function, and (2) an online phase where the model adapts to user preferences through in-context learning. We demonstrate PPT's effectiveness in a contextual bandit setting, showing that it achieves personalized adaptation superior to existing methods while significantly reducing the computational costs. Our results suggest the potential of in-context learning for scalable and efficient personalization in large language models.

cs.LG↗