Illustrative image: Artificial neural network diagram — Wikimedia Commons / Cburnett, CC BY-SA 3.0
📌 KEY HIGHLIGHTS
- What: LoCo is a new training method that stops AI models from "overthinking" — rambling on after they already have enough information to answer.
- Results reported: reasoning output 41.7–45.4% shorter on five maths benchmarks, 23.4–36.3% shorter on three code benchmarks, with accuracy maintained or improved.
- How: it finds the exact point in a model's thinking where extra reasoning stops helping, then trains the model (with DPO) to wrap up there instead.
- Status: a preprint posted to arXiv on 3 November 2026 — promising research, not yet peer-reviewed and not yet inside any product you can use.
- Why it matters: shorter reasoning means lower inference cost, so reasoning AI could get cheaper for Indian developers and users.
What happened?
On 3 November 2026, researchers from Xi'an Jiaotong University and Tencent WeChat AI posted a paper to arXiv titled "Think Less, Locally: Local Preference Learning for Reasoning Compression" (arXiv:2611.04814). It introduces LoCo, a method for compressing the long chain-of-thought that large reasoning models produce before answering.
The problem LoCo attacks is easy to recognise if you have used a reasoning model: it keeps deliberating long after it has enough information, re-checking and second-guessing. That extra thinking costs real money — every extra token a model generates is compute you pay for, whether through an API bill or a slower answer on your screen. Existing compression methods mostly try to shorten the whole reasoning chain, but they pay too little attention to the local decision — at this exact point, is more reasoning actually needed?
How does LoCo work?
LoCo works in three steps, all built on the model's own behaviour rather than on human-written examples:
1. Find the overthinking moments. LoCo scans the model's own reasoning traces and detects "candidate overthinking states" — points where the model has enough information to conclude, but keeps going anyway.
2. Freeze and rewrite. At each such point, LoCo freezes the reasoning written so far and generates a shorter, verified continuation from exactly that point — the same reasoning that stops sooner but still reaches the right answer.
3. Teach the choice with DPO. The long and short versions become preference pairs, and the model is trained with Direct Preference Optimisation (DPO) to prefer the shorter one in that state. The key idea: the model learns when to conclude, not just to be brief in general.
Notably, all the training data came entirely from self-generated mathematical traces — the model learned from its own maths reasoning. Yet the skill transferred: the same training improved code generation too, even though no code data was used. That suggests overthinking is a general habit of reasoning traces, not a quirk of one subject. It is the second efficiency breakthrough we have covered this week — see also our explainer on whether AI agents really need long context windows, where an NVIDIA study found shorter contexts can be 7x cheaper.
LoCo results: shorter reasoning, same accuracy
Here is what the paper reports across its eight benchmarks:
| Benchmark group | Shorter by | Accuracy |
|---|---|---|
| Five mathematics benchmarks | 41.7–45.4% | Maintained or improved |
| Three code generation benchmarks | 23.4–36.3% | Maintained or improved; best accuracy-efficiency trade-off (AUC) on all three |
Two honest caveats, which the coverage of the paper itself flags: this is a preprint, so the results have not yet been peer-reviewed; and the paper's abstract does not name the base models used or publish per-benchmark accuracy numbers. Treat the figures as the authors' reported claims, not as independently verified facts. The authors say code will be released at github.com/MuyaoYuan/LoCo, which will let other researchers check the work.
Illustrative image: Artificial brain illustration — Wikimedia Commons / Gengiskanhg, CC BY-SA 3.0
Why it matters for India readers
For Indian readers, the important number is not 45% — it is the cost behind it. India is one of the world's biggest markets for AI APIs and AI-powered apps, from coding assistants used by Indian developers to customer-support and tutoring products built by Indian startups. Those products pay per token, and reasoning models are the most expensive tokens in the stack because they "think" before they answer.
If a method like LoCo reaches production models, the same answer could cost roughly half as much reasoning compute — cheaper API bills for developers, faster answers for users, and advanced reasoning features reaching lower-priced tiers sooner. Shorter reasoning also helps on-device AI: less battery and memory spent on thinking, which is exactly what budget phones in India need. For a sense of what reasoning models cost today, see our explainer on Gemini Omni 1.2 Flash pricing in India.
None of that is guaranteed yet. LoCo is a research result, not a product feature — it is not in ChatGPT, Gemini or any other chatbot today, and the cost pressure behind it is the same one driving launches like OpenAI's GPT-6 Cyber. But the direction is clear: after a year in which models got smarter by thinking longer, the field is now learning to think just enough — good news for everyone who pays for AI by the token, and Indian builders are among the biggest beneficiaries.
Should you care?
✅ Yes, if you build with or pay for AI. Shorter reasoning is one of the most direct ways AI gets cheaper, and LoCo's numbers — if they hold up under independent testing — are large. Watch for the code release and for model providers adopting the idea; that is when it becomes real savings.
⏳ Wait, if you just use chatbots. Nothing changes in your apps today. This is a paper, not a launch — file it under "the AI you use in 2027 may cost less and answer faster", and keep an eye on follow-up work.
FAQs
What is LoCo in AI?
LoCo is a training method for reasoning compression, introduced in the November 2026 preprint "Think Less, Locally" by researchers at Xi'an Jiaotong University and Tencent WeChat AI. It teaches large reasoning models to recognise when they have thought enough and stop, cutting reasoning length by 41.7–45.4% on maths benchmarks and 23.4–36.3% on code.
What is overthinking in AI reasoning models?
Overthinking is when a reasoning model keeps deliberating — re-checking, hedging and restating — after it already has enough information to give the right answer. It inflates inference cost and waiting time without improving the result, which is the waste LoCo is designed to remove.
How much does LoCo reduce reasoning length?
According to the paper, LoCo reduces completion length by 41.7–45.4% on five mathematics benchmarks and 23.4–36.3% on three code generation benchmarks. The code gains are smaller, but they came from training on maths traces only, with no code-specific training data.
Does LoCo reduce accuracy?
The paper reports accuracy is maintained or improved on all eight benchmarks, with the best accuracy-efficiency trade-off on every code benchmark tested. But the abstract publishes no per-benchmark accuracy numbers, and the work is a preprint that has not been peer-reviewed — so independent verification is still pending.
Takeaway: AI models learned to think longer to get smarter — now they are learning when to stop. LoCo's reported 41–45% cut in reasoning, with no accuracy loss, is a preprint claim worth watching, because cheaper thinking is what brings advanced AI within reach of more Indian developers and users.
Follow World Tech on TechByCountry for tomorrow's update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Think Less, Locally: Local Preference Learning for Reasoning Compression — arXiv:2611.04814 (primary source, submitted 3 Nov 2026)
- AI Weekly — LoCo cuts math reasoning traces 41.7–45.4% with no accuracy loss (5 Nov 2026)
- AI Weekly — Tencent's LoCo DPO compresses reasoning lengths 42–45% on math (4 Nov 2026)
Comments
Post a Comment