Policy optimisation, reward modelling and RLHF — how models are trained by feedback rather than by labels.
arXiv:2606.05122v2 Announce Type: replace Abstract: Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own ou...
arXiv:2604.06079v2 Announce Type: replace Abstract: Graphics Program Synthesis is pivotal for interpreting and editing visual data, effectively facilitating the reverse-engineering of static visuals...
The paper introduces a data‑free learning objective called relative trajectory balance for training diffusion models to sample from a posterior defined by a diffusion prior and an arbitrary black‑box constraint or likelihood. It proves asymptotic correctness of this objective and demonstrates its use across vision, language, and multimodal tasks, including classifier guidance, language infilling, and text‑to‑image generation. Additionally, the method is applied to continuous control with a score‑based behavior prior, achieving state‑of‑the‑art results in offline reinforcement learning.
The paper introduces an Item Response Theory (IRT)–based indicator that identifies likely mislabeled items in large language model (LLM) benchmarks with 95% precision among the top 200 examples across seven preference and multiple-choice datasets, using responses from 114 models. It outperforms a supervised classifier and attributes the mislabels to mechanical labeling heuristics, inherited annotation errors, and inherently ambiguous items. The IRT analysis also reveals that reward models tend to specialize in stylistic preference rather than factual knowledge, and pinpoints a frontier reward model that aligns with detected mislabels at 78% accuracy compared to 38% for other models, suggesting benchmark contamination or over‑optimization.
OmniFusion is an end‑to‑end multilingual multimodal translation system that fuses a pretrained multimodal foundation model (Omni 2.5‑7B) with a translation large language model (SeedX PPO‑7B). By connecting hidden states from multiple layers of the multimodal model to the translation LLM, OmniFusion can translate speech, speech‑and‑image, and text‑and‑image inputs while reducing simultaneous speech‑translation latency by about one second compared to cascaded pipelines. The approach improves overall translation quality by leveraging both audio and visual context.
CommerceVibe is a system that generates e‑commerce creatives by synthesizing executable HTML/CSS code conditioned on product images, design requirements, and product information. It uses dual‑feedback reinforcement learning, combining rule‑based checks for text readability, product visibility, and layout validity with visual feedback from a vision‑language model that evaluates perceptual and commercial aspects. After fine‑tuning a large language model on 28,000 examples and applying dual‑feedback reinforcement learning, CommerceVibe achieves a weighted score of 94.0/100 on a 1,300‑case benchmark, outperforming both its SFT‑only counterpart and external models, and is validated by expert blind evaluations.
GRAIN is a single-agent reinforcement learning framework that improves large language models’ robustness to real‑world shifts in node identifiers and task formulations by treating reasoning as a semantic parsing and tool‑execution pipeline. It introduces a Structure Invariance Reward that validates intermediate graphs against ground‑truth topologies, encouraging the model to learn genuine text‑to‑structure mappings instead of overfitting to surface patterns. On the new GRIT benchmark, GRAIN surpasses multi‑agent baselines by 16.45% in accuracy, reduces latency by about 24%, and halves the out‑of‑distribution gap of fine‑tuned models while remaining robust on large‑scale graphs beyond the training distribution.
arXiv:2608. 27313v1 Announce Type: cross Abstract: We establish a global finite-sample guarantee for synchronous quantile temporal-difference learning (QTD) in tabular distributional reinforcement learning.
PACIFIER is a graph‑learning framework that generates intervention sequences to reduce opinion polarization in online social networks modeled by the Friedkin‑Johnsen framework. It uses history‑aware node representations and both greedy and reinforcement‑learning strategies to score actions, supporting multiple moderation types and continuous opinions. Trained on small synthetic graphs, PACIFIER transfers effectively to large real‑world Twitter networks, outperforming baselines by up to 35.3% and achieving near‑oracle performance while being significantly faster.
AdaThinking-E introduces a reinforcement learning framework that uses one-token entropy regulation to enable large language models to decide adaptively whether to engage in deep reasoning. By measuring entropy in the predicted probability distribution at key decision tokens, the model learns to explore different thinking strategies during training and converge to confident, efficient decision policies. Experiments show the method improves accuracy on complex tasks while reducing computational overhead on simpler ones across various document reasoning benchmarks.
The paper introduces a new approach to safety in contextual bandits with continuous actions, focusing on high‑probability constraints on the realized cost rather than expected cost. It presents the High‑Probability Constrained UCB algorithm, which balances reward exploration with conservative safety estimation, and provides theoretical regret guarantees for linear models and extensions to general function classes. Experiments demonstrate that this realized‑cost safety framework significantly reduces safety violations compared to expected‑cost constrained methods.
The paper investigates whether the high costs of training chain-of-thought reasoning models can be reduced through algorithmic design. It introduces an autocurriculum approach that lets the model select which problems to focus on during training, showing that this method provably improves both supervised fine‑tuning and reinforcement learning. For supervised fine‑tuning, autocurriculum requires exponentially fewer reasoning demonstrations by targeting prompts where the model struggles, while for reinforcement learning it decouples computational cost from the quality of the reference model, making the burn‑in cost nearly independent of target accuracy.
SpeechGym is an audio‑native environment that lets two omni‑modal models converse entirely in native audio, eliminating external ASR/TTS and API boundaries while preserving the tasks, tools, and success checks of a standard text‑based agent benchmark. By training end‑to‑end, the framework addresses perceptual failures—such as misheard arguments that cascade into failed calls—and behavioural failures, both of which are automatically labeled for free. Using per‑turn process rewards to overcome reward sparsity, agents trained in SpeechGym transfer to an independent voice benchmark, doubling task success and improving efficiency in turns and tokens.
The paper introduces Echo-GRPO, a method that rewrites privileged reasoning traces into a model’s own idiolect to align off‑policy supervision with the student policy’s vocabulary. By preserving semantics through Dual‑Reference Decoding, Echo‑GRPO mitigates gradient clipping on critical reasoning tokens and improves reasoning distillation. The approach is instantiated as VideoEcho‑R1 for video reasoning, yielding consistent gains across multiple multimodal LLM backbones and benchmarks, and it can be applied as a plug‑in to both RL and supervised fine‑tuning frameworks.
CARE: Causally-Aligned Reasoning Exploration for Medical Large Language Models proposes a new framework to improve medical reasoning in LLMs. It introduces two key conditions—Causal Sufficiency and Proximal Learnability—to curate high-quality training trajectories, using agreement-based self-verification and dynamic entropy bounds. Experiments on medical multimodal and text-only benchmarks show that CARE outperforms competitors, reducing incorrect reasoning and enhancing training stability.
The paper introduces Safe Contrastive Reinforcement Learning (Safe-CRL), a method that corrects bias in contrastive RL caused by failure-terminated Markov decision processes. By applying mass-weighted InfoNCE and a log-survival-mass score, Safe-CRL uses only a one-bit failure signal to improve survival and goal-reaching performance across twelve robot navigation and locomotion tasks. The approach demonstrates complex failure-avoidance behaviors and completes the theoretical foundation of contrastive RL under failure termination.
The paper introduces MCCE, a hybrid framework that combines a frozen closed‑source large language model (LLM) with a lightweight, trainable model for multi‑objective discrete optimization. By maintaining a trajectory memory and refining the small model through reinforcement learning, the two models jointly enhance global exploration and learning. Experiments on drug‑design benchmarks demonstrate that MCCE achieves state‑of‑the‑art Pareto front quality, outperforming existing baselines.
MemToC is a controlled benchmark that tests how large language models resolve conflicts between their internal memory and tool outputs. It contains 6,504 episodes built from 542 factual questions, each paired with a model‑generated closed‑book answer and a tool return whose correctness is known, creating four distinct source‑correctness scenarios. Across five 7‑9B open‑weight models, tool responses overwhelmingly dominate closed‑book answers, and only a minority of instruction‑tuned models correctly retain a verified answer when the tool is wrong, while most follow a correct tool or repeat a wrong tool.
The paper introduces Reward‑Informed Sparse Autoencoders (RI‑SAEs), which use reinforcement‑learning rewards to curate data for training sparse autoencoders on language‑model activations. On Llama‑3.1‑8B, a sparse subset of features separates high‑reward from low‑reward reasoning continuations, but control experiments show this separation largely reflects solution completeness rather than true reasoning quality. The authors conclude that reward filtering can cheaply reuse RL signals for interpretability, though most of the discovered features capture completion form rather than deep reasoning.
SPEAR (Symbolic Process Evaluation and Alignment Reward) is a training‑free, plug‑and‑play reward method for on‑policy distillation in reinforcement learning. It converts natural‑language reasoning traces into domain‑adaptive symbolic milestones and uses the longest common subsequence to align student exploration with teacher milestones, producing a dense, order‑aware reward that enforces logical consistency without an external neural verifier. Experiments on math, science, and commonsense reasoning tasks show that SPEAR effectively bridges the reasoning gap between student and teacher models through sequence‑level distillation with efficient dense process rewards.