--- name: nowait-reasoning-optimizer-fd71e79f3601 description: Implements the NOWAIT technique for efficient reasoning in R1-style LLMs. Use when optimizing inference of reasoning models (QwQ, DeepSeek-R1, Phi4-Reasoning, Qwen3, Kimi-VL, QvQ), reducing chain-of-thought token usage by 27-51% while preserving accuracy. Triggers on "optimize reasoning", "reduce thinking tokens", "efficient inference", "suppress reflection tokens", or when working with verbose CoT outputs. --- # NOWAIT Reasoning Optimizer Implements the NOWAIT technique from the paper "Wait, We Don't Need to 'Wait'! Removing Thinking Tokens Improves Reasoning Efficiency" (Wang et al., 2025). ## Overview NOWAIT is a training-free inference-time intervention that suppresses self-reflection tokens (e.g., "Wait", "Hmm", "Alternatively") during generation, reducing chain-of-thought (CoT) trajectory length by **27-51%** without compromising model utility. ## When to Use - Deploying R1-style reasoning models with limited compute - Reducing inference latency for production systems - Optimizing token costs for reasoning tasks - Working with verbose CoT outputs that need streamlining ## Supported Models | Model Series | Type | Token Reduction | |--------------|------|-----------------| | QwQ-32B | RL-based | 16-31% | | Phi4-Reasoning-Plus | RL-based | 23-28% | | Qwen3-32B | RL-based | 13-16% | | Kimi-VL-A3B | Multimodal | 40-60% | | QvQ-72B-Preview | Multimodal | 20-30% | **Important**: NOWAIT works best with RL-based models. Distilled models (Qwen3-4B/8B/14B) show degraded performance when reflection tokens are suppressed. ## Quick Start ### 1. Basic Implementation ```python from scripts.nowait_processor import NOWAITLogitProcessor # Initialize processor for your model's tokenizer processor = NOWAITLogitProcessor(tokenizer) # Use during generation outputs = model.generate( inputs, logits_processor=[processor], max_new_tokens=32768 ) ``` ### 2. Keywords Suppressed See `references/keywords.md` for the complete list. Core keywords: ``` wait, alternatively, hmm, but, however, check, double-check, maybe, verify, again, oh, ah ``` ## How It Works 1. **Initialize Keywords**: Identify reflection keywords from empirical analysis 2. **Expand to Token Variants**: Map keywords to all token variants in vocabulary (e.g., "wait" → " wait", "Wait", " Wait", ".wait", "WAIT") 3. **Suppress During Inference**: Set logits of reflection tokens to large negative values during decoding ``` Logits (Before) Logits (After) Wait 0.8 → Wait -inf First 0.6 → First 0.6 Hmm 0.5 → Hmm -inf Let 0.4 → Let 0.4 ``` ## Key Findings ### Why It Works - NOWAIT doesn't eliminate self-reflection entirely—it guides models to skip **unnecessary** "waiting" reasoning - Models still perform essential verification at key decision points - Results in more linear, straightforward reasoning paths ### RL vs Distilled Models | Model Type | NOWAIT Effect | Recommendation | |------------|---------------|----------------| | RL-based (QwQ, Phi4, Qwen3-32B) | Stable accuracy, significant token reduction | ✅ Recommended | | Distilled (Qwen3-4B/8B/14B) | Accuracy degradation on hard tasks | ⚠️ Use with caution | Distilled models rely heavily on CoT structure from training data—removing reflection tokens disrupts their reasoning patterns. ## Integration Examples ### HuggingFace Transformers ```python from transformers import AutoModelForCausalLM, AutoTokenizer from scripts.nowait_processor import NOWAITLogitProcessor model = AutoModelForCausalLM.from_pretrained("Qwen/QwQ-32B") tokenizer = AutoTokenizer.from_pretrained("Qwen/QwQ-32B") processor = NOWAITLogitProcessor(tokenizer) response = model.generate( tokenizer(prompt, return_tensors="pt").input_ids, logits_processor=[processor], max_new_tokens=32768, do_sample=True, temperature=0.7 ) ``` ### vLLM ```python from vllm import LLM, SamplingParams from scripts.nowait_processor import get_nowait_bad_words_ids llm = LLM(model="Qwen/QwQ-32B") bad_words_ids = get_nowait_bad_words_ids(llm.get_tokenizer()) sampling_params = SamplingParams( max_tokens=32768, bad_words_ids=bad_words_ids ) ``` ## Expected Results | Task Type | Original Tokens | NOWAIT Tokens | Reduction | |-----------|-----------------|---------------|-----------| | Math (AIME) | 15,000 | 10,500 | 30% | | Visual QA (MMMU) | 2,900 | 1,450 | 50% | | Video QA (MMVU) | 1,700 | 1,250 | 27% | ## Limitations - Less effective on very simple problems where CoT overhead is already minimal - Distilled models may suffer accuracy loss on challenging tasks - Some domains may require model-specific keyword tuning ## References - Paper: arXiv:2506.08343v2 - Complete keyword list: `references/keywords.md` - Implementation: `scripts/nowait_processor.py` --- Source attribution: Dicklesworthstone Original project: https://github.com/Dicklesworthstone/pi_agent_rust/blob/f28fc46ee0db8e710d95b8b55592cd38d60f2b26/tests/ext_conformance/artifacts/templates-davila7/cli-tool/components/skills/productivity/nowait/SKILL.md License: MIT License information: https://opensource.org/license/mit Original upstream instructions; downloaded from Mega AI Hub. ## Upstream license notice MIT License (with OpenAI/Anthropic Rider) Copyright (c) 2026 Jeffrey Emanuel Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: ADDITIONAL RIDER / RESTRICTION (OpenAI / Anthropic): This rider is part of the "conditions" of this License. In the event of any conflict between this rider and any other portion of this License, this rider controls. "Restricted Parties" means OpenAI, L.L.C.; Anthropic, PBC; any of their respective Affiliates; and any person or entity acting directly or indirectly on behalf of, for the benefit of, or under the direction of any of the foregoing (including any officer, director, employee, contractor, agent, consultant, service provider, or representative). Notwithstanding any other provision of this License, no rights are granted to any Restricted Party. Any purported license, sublicense, assignment, transfer, or other permission to any Restricted Party is null and void absent the express prior written permission of Jeffrey Emanuel. You may not provide, disclose, distribute, sublicense, sell, lease, lend, host, make available, or otherwise permit access to the Software or any derivative work of the Software (as defined in applicable copyright law) (collectively, "Derivative Works") to or for any Restricted Party. For purposes of this rider, "use" includes, without limitation: copying, modifying, merging, publishing, distributing, sublicensing, selling, transferring, making available, hosting, deploying, executing, benchmarking, testing, analyzing, indexing, or incorporating the Software or any Derivative Works into any dataset, training corpus, evaluation harness, or pipeline for machine learning or other automated systems. This rider applies to the Software and all Derivative Works. As a condition of use, you agree that this rider is a precondition to exercising any rights under this License, and you agree that any distribution of the Software or any Derivative Works must include this rider provision unmodified. Any breach of this rider automatically and immediately terminates the permissions granted by this License. Upon termination, you must immediately cease all use and distribution of the Software and any Derivative Works and destroy all copies under your control. You agree that a breach of this rider would cause irreparable harm and that Jeffrey Emanuel may seek injunctive or other equitable relief to enforce this rider, in addition to any other remedies available at law. To the maximum extent permitted by applicable law, the prevailing party in any action to enforce this rider shall be entitled to recover reasonable attorneys' fees and costs. For purposes of this rider, "Affiliate" means any entity that directly or indirectly controls, is controlled by, or is under common control with the specified party. "Control" means ownership of more than 50% of the voting securities or other ownership interest, or the power to direct management or policies by contract or otherwise. The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. # Mega AI Hub resource metadata Creator / publisher: Dicklesworthstone Original project: https://github.com/Dicklesworthstone/pi_agent_rust/blob/f28fc46ee0db8e710d95b8b55592cd38d60f2b26/tests/ext_conformance/artifacts/templates-davila7/cli-tool/components/skills/productivity/nowait/SKILL.md Attribution basis: Source repository record Fields of work: Data & Analytics, Development & Engineering Roles: Backend Developer, Data Scientist, ML / AI Engineer Specialties: API, Backend Tools: Huggingface Transformers Country applicability: Country not specified Classification basis: Source catalog labels Instruction SHA-256: 99b9f625034afa654e8197332f032449d2faaee228e37383751cf1403874bcda