Releases·Tool Release
onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction
Researchers introduce onPanda, an interactive tool for annotating LLM alignment data and agent trajectories via token-level correction. Annotators fix the first inappropriate token in a model response and let generation continue from the corrected prefix, producing data that stays on-policy while cutting annotation time. They also release Panda-CVL, an annotated dataset, plus a token-level correction benchmark.
0 engagement1 sourceMon, Sep 21, 2026, 05:56 PM
AI Overview
- onPanda lets annotators correct the first bad token via candidates or free-form editing, then regenerate, cutting median annotation time 52% vs manual post-editing.
- Token-level corrections yield fine-grained supervision with precise positions and paired positive-negative samples, suiting on-policy SFT and preference data.
- The tool connects to external tools/harnesses for interactive agent trajectory annotation; researchers release Panda-CVL dataset and a token-level correction benchmark.
Related
Releases·Tool Release
Steer LLMs and Agents at the Token Level: An interactive tool for token visualization & control, model inspection and data annotation.
A developer released onPanda (on-Policy Alignment Data Annotator), an interactive tool for token-level visualization, editing, and control of LLM and agent outputs. Built over two years since September 2024, it targets model inspection, prompt engineering, and alignment data annotation.
A developer released onPanda (on-Policy Alignment Data Annotator), an interactive tool for token-level visualization, editing, and control of LLM and agent outputs. Built over two years since September 2024, it targets model inspection, prompt engineering, and alignment data annotation.
141 engagement1 sourceredditSat, Sep 19, 2026, 04:21 AM
Products·launched
OnPanda – token-level steering and inspection UI for LLMs and agents
OnPanda is an interactive web tool for visualizing tokens, editing model output at the token level, and annotating data. Users hover over a token, pick an alternative or edit it, and continue generation, including edits to reasoning and tool calls. It targets hackers, engineers, and power users who need deep model inspection and efficient data annotation.
OnPanda is an interactive web tool for visualizing tokens, editing model output at the token level, and annotating data. Users hover over a token, pick an alternative or edit it, and continue generation, including edits to reasoning and tool calls. It targets hackers, engineers, and power users who need deep model inspection and efficient data annotation.
1 engagement1 sourcehackernewsFri, Sep 18, 2026, 07:26 PM
Developments·Paper
Step-Level On-Policy Distillation: Interpolating Between On-Policy Distillation and Supervised Fine-Tuning
A new paper proposes Step-Level On-Policy Distillation (SOPD), which combines the long-horizon correction of supervised fine-tuning with the efficiency of on-policy distillation. The method addresses a key limitation of token-level OPD, which only provides fragmented corrections along erroneous student trajectories. SOPD aims to improve student model alignment with teacher logits on student-generated trajectories.
A new paper proposes Step-Level On-Policy Distillation (SOPD), which combines the long-horizon correction of supervised fine-tuning with the efficiency of on-policy distillation. The method addresses a key limitation of token-level OPD, which only provides fragmented corrections along erroneous student trajectories. SOPD aims to improve student model alignment with teacher logits on student-generated trajectories.
0 engagement1 sourcearxivMon, Aug 17, 2026, 09:37 AM
Developments·Paper
Experts Rise Where LLMs Disagree: Using Cross-Model Disagreement to Target Expert Effort in LLM Codebook Revision for Large-Scale Annotation
A research paper proposes using cross-LLM disagreement to surface cases needing expert attention when revising annotation codebooks. In experiments on tutoring-session transcripts, expert feedback elicited via rationale labeling reached 64.9% LLM-labeling accuracy versus 57.8% for an expert-revised codebook. The approach cuts months of codebook development to days, reducing expert workload without sacrificing labeling performance.
A research paper proposes using cross-LLM disagreement to surface cases needing expert attention when revising annotation codebooks. In experiments on tutoring-session transcripts, expert feedback elicited via rationale labeling reached 64.9% LLM-labeling accuracy versus 57.8% for an expert-revised codebook. The approach cuts months of codebook development to days, reducing expert workload without sacrificing labeling performance.
0 engagement1 sourcearxivTue, Sep 22, 2026, 06:19 PM
Developments·Paper
Deep Interaction: A New Method for Correcting Reasoning Errors in LLMs
Researchers propose Deep Interaction, a human-AI interaction method that allows users to directly correct specific reasoning steps in Chain-of-Thought outputs, reducing error recurrence compared to traditional regeneration or follow-up turns.
Researchers propose Deep Interaction, a human-AI interaction method that allows users to directly correct specific reasoning steps in Chain-of-Thought outputs, reducing error recurrence compared to traditional regeneration or follow-up turns.
0 engagement1 sourcearxivWed, Jul 15, 2026, 05:16 PM