Skip to content
Papers.

The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation research paper by Salesforce, 2026

Salesforce · Apr 18, 2026 · Training and scaling · 14 upvotes 5 months ago

Read on arXiv

What it shows

On-policy distillation suffers from miscalibration due to information mismatch between training and deployment contexts, which is addressed through a calibration-aware framework that improves both performance and confidence reliability.

Hugging Face's summary; not yet checked by hand.

More from Salesforce

All 33
PaperCitations
Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM AgentsKeeps an agent's raw past runs and writes a task-specific memory only when a new task arrives, rather than deciding up front what to keep.Agents and evaluation · Sep 20263 days ago-
Flattening Every Memory Peak in Long-Context Mixture-of-Experts TrainingBounds the four memory peaks that crash long-context mixture-of-experts training, from expert dispatch to optimizer state, with fixed-size GPU schedules.Architectures · Sep 202613 days ago-
RISE: Recursive Improvement via Self-Extrapolating Policy DistillationRISE improves language model post-training by recursively generating dense token-level supervision from the model's own reinforcement learning trajectory via self-extrapolation, avoiding external teachers.Training and scaling · Sep 20263 weeks ago-
EVOHARNESSBENCH: Can Your Agents Keep Pace with an Evolving Harness?The study introduces a benchmark to evaluate LLM agents under evolving tool, skill, and agent harnesses, revealing persistent gaps in retention, adaptation, and harness-induced forgetting.Agents and evaluation · Sep 20263 weeks ago-
Random Attention: Rethinking KV Cache Eviction for Efficient ReasoningRandom eviction of reasoning tokens matches selective KV cache compression because reasoning traces are self-protecting through redundancy, making scoring unnecessary once prompts are preserved.Reasoning · Sep 20263 weeks ago-
DarwinX: Evolving Agent Harnesses Through Natural SelectionDarwinX evolves agent harnesses via population selection with frozen models, improving verified performance across benchmarks without benchmark-specific patches.Agents and evaluation · Jul 20268 weeks ago-
StateAct: Program State, before Pixels, for Long-Horizon Computer-Use AgentsStateAct improves computer-use agents by grounding actions, verification, and memory in direct program state rather than screenshots, boosting success rates while reducing cost.Agents and evaluation · Jul 20262 months ago-
Evidence-Backed Video Question AnsweringEvidence-Backed Video Question Answering requires models to provide answers with precise spatio-temporal segmentation evidence, revealing gaps between reasoning and visual grounding that are improved by large-scale instruction tuning.Multimodal and robotics · Jul 20262 months ago-

Other training and scaling papers

Topic
PaperCitations
LoRA: Low-Rank Adaptation of Large Language ModelsLoRA fine-tunes a large model by training small low-rank matrices, cutting trainable parameters by 10,000 times.Microsoft · Jun 20215 years ago23.3k
Scaling Laws for Neural Language ModelsLanguage model loss falls as a smooth power law as model size, data and compute grow.OpenAI · Jan 20206 years ago9,065
Training Compute-Optimal Large Language ModelsChinchilla: for a fixed compute budget, train a smaller model on more data; parameters and tokens should grow together.Google DeepMind · Mar 20224 years ago3,756
Megatron-LM: Training Multi-Billion Parameter Language Models Using Model ParallelismMegatron-LM splits each Transformer layer across GPUs to train models with billions of parameters.NVIDIA · Sep 20197 years ago3,179
DeepSeek LLM: Scaling Open-Source Language Models with LongtermismDeepSeek LLM, an open-source language model project, develops a large dataset and employs SFT and DPO to achieve performance surpassing LLaMA-2 70B and GPT-3.5 in various benchmarks and open-ended evaluations.DeepSeek · Jan 2024 · Unverified2 years ago855
LoRA Learns Less and Forgets LessLoRA learns less than full fine-tuning on code and math, but forgets less of what the model already knew.Databricks · May 20242 years ago407
About this paper
Authors
Jiaxin Zhang, Xiangyu Peng, Qinglin Chen and 3 more
arXiv
2604.16830 · PDF
Citations
Not counted yet · Semantic Scholar
Upvotes
14 · Hugging Face
Code
github.com/SalesforceAIResearch/CaOPD
Lab
Salesforce · on Companies · on Acquisitions · on Quarterly · on Paydays · on TechConf

Changes

What changed
New paperAdded to the listSep 26, 2026today

New papers by email

Monday afternoons, only in weeks with new papers from the labs.

Double opt-in. Unsubscribe any time.