Papers

Field

Top tools here

Open problems in Learning verifiable rewards

Poor credit assignment in diffusion model RL1 paper raises it · 1 addresses it
Limits efficiency and stability when fine-tuning diffusion models with black-box rewards.
- DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Modelsraises · addresses
Vanishing group-relative reward signals1 paper raises it · 1 addresses it
All rollouts receiving the same reward lose useful trajectory-level information for learning.
- Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresightraises · addresses
Memory inefficiency in LLM RL post-training1 paper raises it · 1 addresses it
High memory demands prevent scaling RL to larger models or longer training runs.
- LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketchesraises · addresses
Sparse terminal rewards degrade value tracking1 paper raises it · 1 addresses it
Standard critics fail to model value evolution along trajectories, harming credit assignment.
- TPPO: Tracking-Value PPO For Language Model Reasoningraises · addresses
Reward hacking in rubric-based RL1 paper raises it · 1 addresses it
Additive reward aggregation allows policies to exploit criteria compensation, reducing task performance.
- Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubricsraises · addresses

Named by an AI from these papers' abstracts.

Open problems

Recently funded

Topics in this field

learning verifiable rewards · reinforcement learning rl · reinforcement learning verifiable336 papers45 this week
optimal regret · near optimal · bandits221 papers32 this week
offline reinforcement learning · offline reinforcement · reinforcement learning rl190 papers32 this week
reinforcement learning rl · reinforcement learning · markov decision processes108 papers16 this week
policy gradient · reinforcement learning rl · reinforcement learning verifiable94 papers14 this week

Topics in this field

Active contributors · 90 days

Active contributors

Activity · 12 weeks

Activity

Recent papers

All papers in Research

arXiv cs.LGSep 29
Uncertainty-Normalized Margins for Direct Preference Optimization
Sadegh Khorasani, Petrus Mikkola, Matthias Grossglauser
Direct preference optimization (DPO) models binary preferences through a Bradley-Terry model with a common noise scale, without explicitly accounting for preference strength or prompt-dependent uncertainty from human feedback. We introduce uncertainty-normalized margin DPO (UNM-DPO), which combines strength-dependent margins with a learned prompt scale. Motivated by a heteroskedastic Bradley-Terry model, we develop two training objectives. Both compare the implicit rewards of preferred and rejected responses, derived from response log-probability ratios to a reference policy. Advantage-only (AO) divides this reward difference by the prompt scale before subtracting the margin; whole-residual (WR) subtracts the margin before dividing by the scale. For the WR comparison model, we establish a necessary and sufficient condition under which known margins make the prompt scale identifiable. We introduce a practical procedure for learning the scale. Building on WR, we introduce ULNM-DPO-WR, which normalizes each response's implicit reward by its length. We evaluate our methods against DPO and related baselines on HelpSteer2 and HelpSteer3, using the Skywork reward model as a judge. With…
PDF arXiv
arXiv cs.LGSep 29
Revisiting scaling laws for reward optimization
Ali Aouad, Aymane El Gadarri, Vivek F. Farias
Scaling laws for optimization against reward models in AI alignment have pinned down how performance depends on optimization effort---measured by a KL-divergence budget relative to a reference policy. Beyond a certain budget, over-optimization (or reward hacking) can arise: because we optimize against a proxy reward model (distinct from true rewards), performance can plateau or degrade. Naturally, the proxy reward's accuracy depends on how much preference data (often in the form of pairwise comparisons) was used to train it. However, existing research does not cleanly identify how performance jointly scales with the amount of training data and the divergence budget. Our main contribution is to provide an empirically accurate and theoretically grounded scaling law in such context. Performance roughly scales as $Θ(\sqrt{\min\{\log(M),K\}})$, where $M$ is the number of comparisons in training data and $K$ is the policy's divergence budget. We develop an information-theoretic model to establish this upper bound and prove it is tightly achievable through a constructive procedure. Informed by this, we conduct extensive empirical evaluations using a real-world annotation setup, whereby…
PDF arXiv
arXiv cs.LGSep 29
What Pretraining and Midtraining Make Learnable from Rewards?
Chiwun Yang, Xiaoyu Li
A reward can identify a correct answer while leaving the computation needed for new inputs undetermined. We study how pretraining and midtraining supply the information and computation that make reward adaptation effective. In sequential state computation and contextual memory, we characterize mechanisms that agree on every training reward yet demand different held-out answers. Task-independent source observations resolve this ambiguity. We construct finite sampled Adam paths from specified random initializations through source prediction and reward adaptation in the same parameters, proving how prediction acquires execution or retrieval and rewards learn their task-specific use. Experiments with pretrained Qwen2.5 checkpoints test this division of labor. Across eight worlds, Sequential models trained with correct source and first-operation supervision reach 82.61% success, versus 44.15% for a private-random source control. Memory replay preserves retrieval during reward adaptation, and an independent eight-world confirmation achieves 75.32% task success versus 49.86% after matched alternative-retrieval training. GSM8K and HotpotQA separate accuracy at reward entry, subsequent…
PDF arXiv
arXiv hep-phSep 29
Searching for BSM Experimental Signatures with Large Lagrangian Models
Ibrahim Elsharkawy, Victoria Knapp-Perez, Wahid Bhimji, Aishik Ghosh
The search for physics Beyond the Standard Model (BSM) is generally limited not by the supply of theory descriptions but by the lack of discriminating experimental observations. A case in point is dark matter, where the overwhelming gravitational evidence only goes so far in distinguishing between models within a vast theory space. Exploring the space of testable model signatures may help identify overlooked experimental observables and indicate the utility of future experiments. A challenge is designing a search through model signatures outside what is found in the literature. Our primary contribution is hAIthem, a framework that combines the self-guided exploration of reinforcement learning (RL) with the broad literature-derived knowledge of LLMs. We build an RL agent that learns to find which portions of a theory's high-dimensional parameter space are not excluded under some subset of constraints by playing a Battleship-style "game" against a suite of phenomenology tools. The agent is built as a Large Lagrangian Model (LLaM), an autoregressive transformer that reads a tokenized Lagrangian, is pretrained at scale (here on ~1 billion tokens from ~10,000 Lagrangians), and is…
PDF arXiv
arXiv cs.LGSep 29
Achieving an $O(1/N)$ Optimality Gap in Average-Reward Weakly-Coupled MDPs
Yige Hong, Xiangcheng Zhang, Qiaomin Xie, Yudong Chen, Weina Wang
We study average-reward weakly-coupled Markov decision processes (WCMDPs), where a WCMDP consists of $N$ smaller MDPs, called arms, that share multiple per-step budget constraints. We consider the setting where the arms have identical model parameters, multiple actions, and state- and action-dependent costs. For restless bandits (RBs), a well-studied special case of WCMDPs, prior work has developed policies that achieve an $O(1/\sqrt{N})$ optimality gap under general conditions, and has further identified conditions under which policies can achieve a better-than-$1/\sqrt{N}$ optimality gap. However, for general WCMDPs, no prior result achieves an optimality gap better than $1/\sqrt{N}$. In this paper, we identify conditions analogous to those for RBs under which a better-than-$1/\sqrt{N}$ optimality gap is achievable, and design a policy that attains an $O(1/N)$ optimality gap. Notably, unlike prior approaches based on generalizing priority orderings, our policy is not priority-based but rather is designed to induce locally linear mean-field dynamics.
PDF arXiv
arXiv cs.LGCodeSep 29▲ 20
Tail-Influence Sampling for CVaR Policy Evaluation
Pauline Bourigault, Xiaotong Ji, Matthieu Zimmer, Rasul Tutunov, Haitham Bou-Ammar
Policies with similar mean returns can differ sharply in rare failures, yet estimating lower-tail conditional value-at-risk (CVaR) accurately can require many costly rollouts. When different conditional components of a stochastic workflow can be queried separately, we ask how to allocate a fixed evaluation budget to estimate a fixed policy's CVaR most accurately. We derive a tail influence for each queryable conditional law that aggregates how its uncertainty affects CVaR across every Bellman reuse. Its variance yields the fixed-design efficiency bound and the oracle Neyman allocation. Tail-Influence Sampling (TIS) estimates these influence scales from a pilot model and reallocates fresh queries toward kernels that matter most for the tail; a visitation-anchored variant protects against pilot underallocation. Under fixed dimension and a positive quantile margin, TIS attains oracle asymptotic variance and first-order MSE including pilot cost, while the anchored variant is within a factor two of the oracle. We also characterize an exact-grid regime in which tail- and mean-optimal allocations coincide. On CliffWalking, TIS reduces MSE by 41% versus learned occupancy and 76% versus…
PDF arXiv Code

Papers

Topics in this field

learning verifiable rewards · reinforcement learning rl · reinforcement learning verifiable336 papers45 this week
optimal regret · near optimal · bandits221 papers32 this week
offline reinforcement learning · offline reinforcement · reinforcement learning rl190 papers32 this week
reinforcement learning rl · reinforcement learning · markov decision processes108 papers16 this week
policy gradient · reinforcement learning rl · reinforcement learning verifiable94 papers14 this week

Active contributors · 90 days

Recent papers

All papers in Research

arXiv cs.LGSep 29
Uncertainty-Normalized Margins for Direct Preference Optimization
Sadegh Khorasani, Petrus Mikkola, Matthias Grossglauser
Direct preference optimization (DPO) models binary preferences through a Bradley-Terry model with a common noise scale, without explicitly accounting for preference strength or prompt-dependent uncertainty from human feedback. We introduce uncertainty-normalized margin DPO (UNM-DPO), which combines strength-dependent margins with a learned prompt scale. Motivated by a heteroskedastic Bradley-Terry model, we develop two training objectives. Both compare the implicit rewards of preferred and rejected responses, derived from response log-probability ratios to a reference policy. Advantage-only (AO) divides this reward difference by the prompt scale before subtracting the margin; whole-residual (WR) subtracts the margin before dividing by the scale. For the WR comparison model, we establish a necessary and sufficient condition under which known margins make the prompt scale identifiable. We introduce a practical procedure for learning the scale. Building on WR, we introduce ULNM-DPO-WR, which normalizes each response's implicit reward by its length. We evaluate our methods against DPO and related baselines on HelpSteer2 and HelpSteer3, using the Skywork reward model as a judge. With…
PDF arXiv
arXiv cs.LGSep 29
Revisiting scaling laws for reward optimization
Ali Aouad, Aymane El Gadarri, Vivek F. Farias
Scaling laws for optimization against reward models in AI alignment have pinned down how performance depends on optimization effort---measured by a KL-divergence budget relative to a reference policy. Beyond a certain budget, over-optimization (or reward hacking) can arise: because we optimize against a proxy reward model (distinct from true rewards), performance can plateau or degrade. Naturally, the proxy reward's accuracy depends on how much preference data (often in the form of pairwise comparisons) was used to train it. However, existing research does not cleanly identify how performance jointly scales with the amount of training data and the divergence budget. Our main contribution is to provide an empirically accurate and theoretically grounded scaling law in such context. Performance roughly scales as $Θ(\sqrt{\min\{\log(M),K\}})$, where $M$ is the number of comparisons in training data and $K$ is the policy's divergence budget. We develop an information-theoretic model to establish this upper bound and prove it is tightly achievable through a constructive procedure. Informed by this, we conduct extensive empirical evaluations using a real-world annotation setup, whereby…
PDF arXiv
arXiv cs.LGSep 29
What Pretraining and Midtraining Make Learnable from Rewards?
Chiwun Yang, Xiaoyu Li
A reward can identify a correct answer while leaving the computation needed for new inputs undetermined. We study how pretraining and midtraining supply the information and computation that make reward adaptation effective. In sequential state computation and contextual memory, we characterize mechanisms that agree on every training reward yet demand different held-out answers. Task-independent source observations resolve this ambiguity. We construct finite sampled Adam paths from specified random initializations through source prediction and reward adaptation in the same parameters, proving how prediction acquires execution or retrieval and rewards learn their task-specific use. Experiments with pretrained Qwen2.5 checkpoints test this division of labor. Across eight worlds, Sequential models trained with correct source and first-operation supervision reach 82.61% success, versus 44.15% for a private-random source control. Memory replay preserves retrieval during reward adaptation, and an independent eight-world confirmation achieves 75.32% task success versus 49.86% after matched alternative-retrieval training. GSM8K and HotpotQA separate accuracy at reward entry, subsequent…
PDF arXiv
arXiv hep-phSep 29
Searching for BSM Experimental Signatures with Large Lagrangian Models
Ibrahim Elsharkawy, Victoria Knapp-Perez, Wahid Bhimji, Aishik Ghosh
The search for physics Beyond the Standard Model (BSM) is generally limited not by the supply of theory descriptions but by the lack of discriminating experimental observations. A case in point is dark matter, where the overwhelming gravitational evidence only goes so far in distinguishing between models within a vast theory space. Exploring the space of testable model signatures may help identify overlooked experimental observables and indicate the utility of future experiments. A challenge is designing a search through model signatures outside what is found in the literature. Our primary contribution is hAIthem, a framework that combines the self-guided exploration of reinforcement learning (RL) with the broad literature-derived knowledge of LLMs. We build an RL agent that learns to find which portions of a theory's high-dimensional parameter space are not excluded under some subset of constraints by playing a Battleship-style "game" against a suite of phenomenology tools. The agent is built as a Large Lagrangian Model (LLaM), an autoregressive transformer that reads a tokenized Lagrangian, is pretrained at scale (here on ~1 billion tokens from ~10,000 Lagrangians), and is…
PDF arXiv
arXiv cs.LGSep 29
Achieving an $O(1/N)$ Optimality Gap in Average-Reward Weakly-Coupled MDPs
Yige Hong, Xiangcheng Zhang, Qiaomin Xie, Yudong Chen, Weina Wang
We study average-reward weakly-coupled Markov decision processes (WCMDPs), where a WCMDP consists of $N$ smaller MDPs, called arms, that share multiple per-step budget constraints. We consider the setting where the arms have identical model parameters, multiple actions, and state- and action-dependent costs. For restless bandits (RBs), a well-studied special case of WCMDPs, prior work has developed policies that achieve an $O(1/\sqrt{N})$ optimality gap under general conditions, and has further identified conditions under which policies can achieve a better-than-$1/\sqrt{N}$ optimality gap. However, for general WCMDPs, no prior result achieves an optimality gap better than $1/\sqrt{N}$. In this paper, we identify conditions analogous to those for RBs under which a better-than-$1/\sqrt{N}$ optimality gap is achievable, and design a policy that attains an $O(1/N)$ optimality gap. Notably, unlike prior approaches based on generalizing priority orderings, our policy is not priority-based but rather is designed to induce locally linear mean-field dynamics.
PDF arXiv
arXiv cs.LGCodeSep 29▲ 20
Tail-Influence Sampling for CVaR Policy Evaluation
Pauline Bourigault, Xiaotong Ji, Matthieu Zimmer, Rasul Tutunov, Haitham Bou-Ammar
Policies with similar mean returns can differ sharply in rare failures, yet estimating lower-tail conditional value-at-risk (CVaR) accurately can require many costly rollouts. When different conditional components of a stochastic workflow can be queried separately, we ask how to allocate a fixed evaluation budget to estimate a fixed policy's CVaR most accurately. We derive a tail influence for each queryable conditional law that aggregates how its uncertainty affects CVaR across every Bellman reuse. Its variance yields the fixed-design efficiency bound and the oracle Neyman allocation. Tail-Influence Sampling (TIS) estimates these influence scales from a pilot model and reallocates fresh queries toward kernels that matter most for the tail; a visitation-anchored variant protects against pilot underallocation. Under fixed dimension and a positive quantile margin, TIS attains oracle asymptotic variance and first-order MSE including pilot cost, while the anchored variant is within a factor two of the oracle. We also characterize an exact-grid regime in which tail- and mean-optimal allocations coincide. On CliffWalking, TIS reduces MSE by 41% versus learned occupancy and 76% versus…
PDF arXiv Code

Activity · 12 weeks

Open problems in Learning verifiable rewards

Poor credit assignment in diffusion model RL1 paper raises it · 1 addresses it
Limits efficiency and stability when fine-tuning diffusion models with black-box rewards.
- DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Modelsraises · addresses
Vanishing group-relative reward signals1 paper raises it · 1 addresses it
All rollouts receiving the same reward lose useful trajectory-level information for learning.
- Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresightraises · addresses
Memory inefficiency in LLM RL post-training1 paper raises it · 1 addresses it
High memory demands prevent scaling RL to larger models or longer training runs.
- LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketchesraises · addresses
Sparse terminal rewards degrade value tracking1 paper raises it · 1 addresses it
Standard critics fail to model value evolution along trajectories, harming credit assignment.
- TPPO: Tracking-Value PPO For Language Model Reasoningraises · addresses
Reward hacking in rubric-based RL1 paper raises it · 1 addresses it
Additive reward aggregation allows policies to exploit criteria compensation, reducing task performance.
- Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubricsraises · addresses

Named by an AI from these papers' abstracts.

Fields · AI · updated 43m ago

Learning verifiable rewards

Found by the topic model: learning verifiable rewards, optimal regret, offline reinforcement learning.

139papers this week−32.2% vs last week

All papers · 949 With code · 29

Newest Most relevant

874 papers this month