#1 ·
SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering
Genming Cui, Xiangyu Li, Zixu Hao, Xin Ding
QuantizationLLM AgentsDistillation
Why it matters
Low-bit quantization is how teams run agents cheaply in production, but it silently degrades task success, and re-quantizing is the usual (slow) fix. This targets the actual failure mode instead: the specific turns where a quantized agent drifts from a full-precision reference. It is a practical lever for anyone balancing agent quality against inference cost.
SteerDistill is a quantization-aware distillation method for LLM agents. The authors trace quantization failures to behavioral divergence: the student overlooks task requirements or environment conditions at particular turns, and its rollouts then drift from successful full-precision trajectories. A full-precision teacher guides the student to revise those failed responses and iteratively refines its guidance from feedback on student execution, then the student is trained to reproduce the corrected behavior unaided. Across five agentic benchmarks and three models it beats the strongest quantized baseline by 1.67 to 2.18 points under W4A16 and stays within one point of BF16; on tau2-Bench it gains 1.84 percentage points of success rate over QeRL with 67 percent fewer training GPU hours.
arXiv | HuggingFace
#2 ·
Incentivizing Self-Learning in Large Language Models via Reinforcement Learning
Lingkai Kong, Yuzhe Gu, Haiteng Zhao, Zijian Wu
Reinforcement LearningSelf-LearningLong Context
Why it matters
RAG answers questions from pre-assembled context, but unfamiliar domains hand the model raw material and expect it to figure out what to read. Training the find-study-apply loop directly with fine-grained RL rewards is a step toward agents that can absorb new long-form material on the fly instead of depending on retrieval pipelines.
The paper formulates self-learning as an autonomous skill: navigate textbook exposition, retrieve and study the relevant principles, and construct proofs for exercises, with reference solutions withheld. A BookNavigator interface supports the study process, and RL with fine-grained rewards incentivizes it. The result lifts Intern-S2-Preview-35B overall proof accuracy from 51.5 to 79.5 percent across in-domain and out-of-domain textbooks and subjects, beating RL-trained textbook-free and fixed-retrieval baselines. Gains transfer to long-context benchmarks, with improvements of 1.7 and 2.3 percentage points over the base model on BrowseComp-Plus and CL-bench, suggesting a domain-agnostic self-learning strategy.
arXiv | HuggingFace
#3 ·
Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu
AI AgentsInference EfficiencyToken Cost
Why it matters
Web-facing agents read whole pages, documents, and tool responses when they usually need only sparse evidence, and the over-reading burns tokens, latency, and signal. This is a drop-in control layer that works across agent frameworks without modifying the model, so it is a deployable cost lever rather than a research curiosity.
SparseRead is a reusable reading control layer that governs what content an agent is admitted to see. A Read Intent Protocol tracks the agent's evolving needs and source coverage, adaptive reader backends choose how evidence is located and exposed, and a Budgeted Read Gate bounds cumulative raw exposure per episode with a faithful full-view fallback. Across six models including Claude Opus 5 and five scenarios it reduces total tokens by up to 92.9 percent and wall time by up to 89.0 percent while preserving or improving task quality in 97 percent of settings. On web deep-search tasks it cuts tokens by 31 to 43 percent and completion time by 23 to 37 percent, and integration with three agent frameworks shows it is portable.
arXiv | HuggingFace
Outlook
The theme is cost discipline: smarter quantization, self-acquired knowledge, and budgeted reading all attack the same problem, which is that agents are hungry. Expect quantization-aware distillation to spread to more model families and bit-widths, self-learning RL to move beyond math into new domains, and read-gate-style controls to become a standard layer in agent frameworks.