Research Papers Topics
Training language models and agents from reward — RLHF, RLVR, GRPO, and verifiable-reward methods that drive reasoning gains.
Reinforcement Learning · Alibaba Qwen Team
Bebop proves multi-token prediction acceptance is bounded by rising RL entropy. A total-variation loss plus rejection sampling holds acceptance near 95%, flattens the entropy slope to -0.06, for up to 1.8x faster RL.
Reinforcement Learning · University of Zurich
RL with a chrF reward teaches an LLM to translate from in-context linguistic packets rather than memorize languages. On five unseen languages it averages 0.3335 chrF vs 0.2300 for SFT, which drops below the base model.
SDPG adds a full-vocabulary self-distillation loss to verifier RL, learning from a hint-conditioned teacher. On Qwen3-4B it lifts AIME25 from 0.242 (GRPO) to 0.327 and AMC23 from 0.714 to 0.870 last-checkpoint.
Reinforcement Learning · Tencent
CPPO replaces PPO's one-size threshold with stricter early-token clipping and a running prefix-divergence budget, lifting Qwen3-30B-A3B-Base AIME from 49.23 (DPPO) to 54.79.
Reinforcement Learning · Xi'an Jiaotong University
Flow-DPPO swaps PPO ratio clipping for an exact per-step Gaussian KL term, lifting GenEval2 to 48.1 on SD3.5 (vs 39.9 for Flow-GRPO) while cutting policy drift roughly 4x.
AI Agents · The Chinese University of Hong Kong
Orchestra-o1 orchestrates text, image, audio, and video sub-agents and hits 72.8% on OmniGAIA with a GPT-5 brain (+10.3 over Gemini-3-Pro). Its trained 8B orchestrator reaches 30.0%, best among open omnimodal agents.
AI Agents · City University of Hong Kong
RHO tunes an LLM agent harness from past unlabeled trajectories using self-consistency and pairwise self-preference, lifting SWE-Bench Pro from 59% to 78% in one round with no external grading.
Reinforcement Learning · Tencent
DRPO swaps DPPO's hard divergence mask for a smooth advantage-weighted quadratic regularizer, keeping the Binary-TV trust region but with bounded gradient weights, and trains Qwen3 LLMs more stably under FP8.
A CASIA survey maps agentic environments for LLM agents along eight attribute axes and eight domains, unifying synthesis, evaluation, and co-evolution. Sharpest finding: environments barely fit multi-agent settings.
Reinforcement Learning · Alibaba Qwen Team
APPO branches RL rollouts at high-uncertainty, high-influence tokens instead of tool-call boundaries, lifting Qwen2.5-7B by 3.9 points over ARPO across 13 math, multi-hop, and deep-search benchmarks.
Reinforcement Learning · Alibaba Qwen Team
Z-Reward predicts a distribution over rubric scores instead of one scalar. A 9B student hits 88.6% human-preference accuracy with a single output token, and downstream T2I tuning gains 41.3% net GSB over SFT.
AI Agents · University of Science and Technology of China
Role-Agent makes a single LLM act as agent and environment at once, generating its own process reward and curriculum. It beats GiGPO by 4.2% on ALFWorld and 6.9% on WebShop with Qwen2.5-1.5B.
Reinforcement Learning · Zhejiang University
N-GRPO perturbs rollout embeddings with semantic neighbors, lifting DeepSeek-R1-Distill-Qwen-1.5B average Pass@32 to 79.17 and AIME25 Pass@32 to 50.28.
Text-to-Image · The Chinese University of Hong Kong
InterleaveThinker adds planner and critic agents around frozen image generators, reaching 66.3 to 67.2 average on UEval and lifting FLUX.2-klein WISE from 0.47 to 0.73.
MaxProof turns MiniMax-M3 into a generator, verifier, fixer, and ranker; with population-level test-time scaling it reports 35/42 on IMO 2025 and 36/42 on USAMO 2026.
AI Agents · University of Illinois Urbana-Champaign
Harness-1 is a 20B RL search agent that hands working memory to the environment, hitting 0.730 average curated recall and beating the next open subagent by +11.4 points.
Theorem Proving · Google Research
HOList turns machine learning for higher-order theorem proving into a concrete research object, with evidence anchors, method tradeoffs, and limits for practical use.
Reinforcement Learning · Tianjin University
When you RL-tune an LLM across math, code, QA, and writing in sequence, math drops from 66.49 to 57.66 even though gradients look orthogonal. A short math refresh pulls it back to 66.04 without wrecking the other three.
Long Context · Tsinghua University
Tsinghua's LongTraceRL mines distractors from real search-agent trajectories and adds entity-level rubric rewards, lifting a Qwen3-4B reasoner from 53.3 to 59.0 average across five long-context benchmarks (+5.7).
OpenSkill lets agents build skills and their own verifiers from the open web, hitting 43.6% on SkillsBench (+8.9 over the best baseline) with zero target-task answers.
Reinforcement Learning · Tsinghua University
CHERRL injects four known judge biases to reliably reproduce reward hacking in rubric RL; an agent reading only training logs pinned the onset with 11-step total interval error and missed none of six runs.
SAAS uses self-aware RL to cut a Qwen2.5-7B search agent's average queries from 2.19 to 0.97 per question, while keeping accuracy near the best baseline (48.7% vs 49.8%).
Reinforcement Learning · University of Edinburgh
SCOPE co-evolves a task-writing Challenger and a retrieval Solver, judged by a frozen copy of the base model, lifting eight open-ended benchmarks by up to +10.4 points with zero curated prompts.
TaskMem trains a multimodal agent to write its own memory with RL, lifting streaming-video QA accuracy to 67.9% on VideoMME and 45.4% on EgoLife, gains of 6.3 and 7.0 points over the Qwen3-VL-30B baseline.
DeepSeek-Prover-V1.5 combines Lean feedback, reinforcement learning, and RMaxTS search, reaching 63.5% on miniF2F and 25.3% on ProofNet.
PPO keeps policy-gradient RL stable with a clipped surrogate objective — almost as well-behaved as TRPO but far simpler — which made it the default RL engine behind RLHF for ChatGPT and InstructGPT.
AI Agents · Zhejiang University
SDAR adds a gated, token-level self-distillation signal from a skill-augmented teacher on top of GRPO, lifting multi-turn agents by up to +10.2 points on WebShop and +9.4 on ALFWorld for small Qwen models.
DeepSeek-R1 learns to reason from reinforcement learning on whether its answer is correct — with no human reasoning examples — matches OpenAI o1 on AIME and MATH-500, and ships open MIT-licensed weights.