Jan-nano Technical Report

Alan Dao (Gia Tuan Dao); Dinh Bach Vu

REVIEW 5 major objections 5 minor 19 references

Jan-nano Technical Report

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 4B-parameter model trained without next-token prediction reaches 83.2% on SimpleQA with tool access, beating far larger models.

desk verdict Plausible 4B tool-augmented QA result, but the evaluation protocol is undocumented enough that the headline 83.2% SimpleQA is not verifiable. read the letter →

arxiv 2506.22760 v2 pith:634KLNZQ submitted 2025-06-28 cs.CL

Alan Dao (Gia Tuan Dao) , Dinh Bach Vu This is my paper

classification cs.CL

keywords smalllanguagemodelstool-augmentedfactualitySimpleQAreinforcementlearningwithverifiablerewardsretrieval-augmentedgenerationmodelspecializationconsumerhardwareinference

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a 4-billion-parameter model specialized for tool use can beat models orders of magnitude larger on factual question answering. Jan-nano, built on a 4B base model and trained with reinforcement learning from verifiable rewards, scores 83.2% on the SimpleQA benchmark when connected to search and scraping tools, a 24-point gain over the base model under the same tool setup and higher than a 671B-parameter model scored with tools. The training pipeline deliberately skips supervised next-token prediction, using reward signals for correct answers and successful tool calls across three stages. The point of the claim is that intelligence for knowledge tasks can come from strategy—knowing how to find information—rather than from parameter count.

What carries the argument

The central mechanism is multi-stage reinforcement learning with verifiable rewards (RLVR), paired with a local retrieval server that simulates a search engine. The server exposes two tools—websearch(query) returning snippets and scrape(id) returning full documents—and training is conducted on MuSiQue multihop questions. Three stages progress from basic tool mechanics to answer correctness to context extension, with a force non-thinking regularization that suppresses chain-of-thought-style overplanning, and a JSON-in-XML tool format keeps rewards easy to parse.

What would settle it

Run the released model on a fresh, randomly drawn half of SimpleQA using the same minimal tool setup with no configuration changes; if the score falls back toward the 59–71% range instead of near 83.2%, the headline gain comes from tuning on the test slice.

Watch

Extended reading notes

Core claim

The report's central claim is that a 4B model specialized for tool use can outperform even 671B-parameter models on tool-augmented short-form factuality. Jan-nano reaches 83.2% on SimpleQA, compared with 78.2% for the 671B model and 59.2% for the originating 4B base model under the same tool setup. The authors attribute the result to a design that prioritizes finding information over storing it, supported by the overthinking analysis: in head-to-head traces, the 4B model's direct, unfiltered searches retrieve the correct document, while an 8B model's added year filters discard relevant results and lead to hallucinated answers. They further claim the entire skill was acquired through multi-stage RLVR without any supervised next-token prediction stage.

Load-bearing premise

The load-bearing premise is that the 10% subset of SimpleQA used to choose no-thinking mode represents the full benchmark well enough that the configuration decision does not inflate the reported 83.2%.

Editorial extensions

If this is right

Tool-augmented factuality can be decoupled from model scale: a 4B model can beat a 671B model on SimpleQA when both use the same search-tool setup.
A small model can learn complex tool use without any supervised next-token training, using only verifiable rewards.
Suppressing chain-of-thought overthinking improves both speed and answer accuracy for search-based question answering.
Longer context (128K versus 8K) contributes a small gain (80.7% to 83.2%), so context length appears secondary to tool strategy.
The model's skill transfers partially across tool-calling formats: 76.2% on a Python-code agent framework versus 80.7% on a JSON-based framework, indicating format alignment matters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If these numbers replicate, the bottleneck for factuality shifts from model size to tool quality: a better search API and scraper could push the same 4B model toward or beyond 85%, as the authors themselves note.
The overthinking result suggests a general design rule for agent training: reward designs that penalize unnecessary actions may help small models in other tool-use domains, not just fact lookup.
A direct test of the no-SFT claim would be to train the same base model with standard supervised fine-tuning on the same MuSiQue tool traces and compare SimpleQA scores; the paper does not report this ablation.
Because the headline evaluation couples model and search tool, the claim is best read as 'small model plus good search beats giant model plus good search,' not as small models being intrinsically more factual.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's the thing about Jan-nano: the model is public, the training approach is a clean extension of Search-R1/ReSearch/DAPO-style RLVR to a 4B model, and the force-non-thinking observation (larger models over-filter and hallucinate) is worth taking seriously. That's the good part.

The bad part is that the central number, 83.2% on SimpleQA with MCP, is not tied to a documented protocol. The report says "the more conservative 83.2% result" after mentioning a forked MCP server that could beat 85%, which means the number is a selection among runs. Table 4 only labels the 80.7% run as "SimpleQA (Full)"; Table 1 gives 83.2% without saying full or subset. The 10% SimpleQA subset was used to decide Think vs No-Think, and there is no statement about whether the final eval used the full set or that same slice, or how the subset was sampled. So the headline may be the result of tuning on the test set.

The baseline comparison is also shaky. Qwen3-4B at 59.2% is presented without evidence that it ran with the same tool stack, prompts, grading, and temperature. The report itself warns that different MCP implementations affect direct comparisons, but doesn't apply that caveat to its own baseline. The no-SFT claim is asserted, not demonstrated—no ablation against an SFT baseline.

I don't think the paper is a waste of time. The model is on Hugging Face, and the task design (RLVR with local RAG server, three-stage rewards, force non-thinking) is a reasonable recipe that someone could reproduce or extend. But the report as written fails to substantiate the one number that matters. A serious referee would need to see the evaluation harness, the answer-matching rule, the full test set details, and at least one ablation for the no-SFT claim.

My recommendation: don't desk-reject it outright—the claims are important enough and the model ships—but it needs heavy revision before it can be taken as evidence. If it's a technical report rather than a journal paper, then the bar is lower, and the public weights carry some weight. Still, the headline should be re-evaluated or clearly labeled as preliminary.

Who this is for: practitioners doing tool-augmented small models, and people who care about honest benchmark reporting. I'd bring it to reading group to discuss what counts as a verifiable evaluation. I would not cite it in my own work yet.

Recommendation: send it to peer review with a request for a revised version that documents the evaluation protocol and ablations.

Referee Report

5 major / 5 minor

Summary. The paper introduces Jan-nano, a 4B-parameter model fine-tuned from Qwen3-4B using a three-stage Reinforcement Learning with Verifiable Rewards (RLVR) pipeline that the authors claim completely avoids SFT. The central empirical claim is that Jan-nano-128k with MCP achieves 83.2% on SimpleQA, a 24-point improvement over the Qwen3-4B MCP baseline (59.2%) and higher than DeepSeek-671B with MCP (78.2%). The report also claims that disabling the thinking mode ("force non-thinking") prevents overthinking and improves tool-use performance, and it releases several public model variants on Hugging Face.

Significance. If the headline result were fully supported, it would be a notable demonstration that a small, tool-specialized model can outperform models orders of magnitude larger on tool-augmented factuality, and the public release of four model variants is a practical strength. The authors' explicit choice of a conservative MCP setup, rather than a forked server they believe could exceed 85%, is also a positive sign. However, the evaluation protocol is not documented well enough to verify the central claim: the exact test set, grading rule, answer-matching procedure, run count, and baseline setup are missing or ambiguous. As written, the paper's significance cannot be assessed reliably.

major comments (5)

[§2, Table 1 and Appendix A.2, Table 4] The 83.2% headline in Table 1 is not tied to a documented evaluation protocol. Appendix A.2/Table 4 labels only Jan-nano (LangGraph MCP) at 80.7% as "SimpleQA (Full)", while Table 1 lists Jan-nano-128k (MCP) at 83.2% without stating whether it was measured on the full SimpleQA benchmark, the 10% subset introduced in Table 2, or some other slice. The text in §2 says the authors "opted for the more conservative 83.2% result", which implies a selection among runs or implementations that is never specified. Without a precise statement of the test set, grading rule, answer-matching criterion, temperature, and number of runs, the central result is not verifiable.
[§5.1, Table 2 and §2] The force-non-thinking design choice was selected using a 10% SimpleQA subset, but the report does not describe how this subset was sampled, whether it is disjoint from the final evaluation, or whether the 83.2% figure was obtained on the full benchmark or on that same subset. Because the final evaluation uses the same benchmark family, the headline number is not an independent held-out evaluation unless overlap is explicitly ruled out. The authors should state the subset sampling procedure, report both full-set and subset results, and clarify all overlaps.
[§2, Table 1] The Qwen3-4B baseline (59.2%) is central to the claimed 24-point improvement, but no evidence is given that this baseline was evaluated with the identical mcp-server-serper, LangGraph ReAct template, grading procedure, temperature, and test subset as Jan-nano. The report itself cautions in §2 that different MCP implementations and evaluation setups affect direct comparisons, yet the baseline is not subjected to that caveat. Without a same-setup comparison, the improvement claim does not follow from the data presented.
[§4 and Abstract] The claim that the RLVR system "completely eliminates reliance on next token prediction training (SFT)" is not substantiated. Section 4 gives no reward weights for the three stages, no DAPO hyperparameters, no number of RL steps, no convergence criteria, and no ablation comparing against an SFT-initialized control. The phrasing is also imprecise because the model is initialized from Qwen3-4B, which was itself trained with next-token prediction; the claim should be restated as "no additional SFT" and supported by the full training configuration.
[§5.1, Table 2 and Appendix A.1] The overthinking conclusion is not directly supported by the reported numbers. Table 2 contains only two Think-mode rows (4B Think 71.0%, 8B Think 62.2%) on a 10% subset, with no No-Think accuracy or latency values, even though the text claims no-think variants deliver significantly faster response times while maintaining competitive accuracy. The qualitative conversation logs in Appendix A.1 illustrate a failure pattern but do not quantify the overthinking effect. Since the adoption of force non-thinking is a load-bearing design decision, this evidence gap should be filled.

minor comments (5)

[Abstract, §4, §6] The term "RL VR" appears with a space in the abstract and Section 4, while "RLVR" is used elsewhere; please standardize the notation.
[Figure 1] Figure 1 is never discussed in the main text; the caption says it visualizes "focus during different training stages", but the figure's axes, panels, and takeaway should be explained in a paragraph.
[Appendix A.1, Table 3] The conversation logs in Table 3 are extremely dense and difficult to read; consider separating the 4B and 8B traces into distinct columns or sub-tables, and fix the recurring "ASSIST ANT" typo.
[§2] The SimpleQA grading protocol is not described: the paper does not state how answers are judged correct, whether "not attempted" answers are included, or whether an exact-match or LLM-based grader is used. This detail should be added in the evaluation section.
[Tables 1 and 4] The benchmark label is inconsistent: Table 1 says "SimpleQA" while Table 4 says "SimpleQA (Full)". These labels should be reconciled so that readers know which configuration produced which number.

Circularity Check

1 steps flagged · score 4.0 of 10

The 83.2% SimpleQA headline follows a design choice made on a 10% SimpleQA subset, so the benchmark result is not fully held out.

fitted input called prediction [Section 5.1 'Overthinking Problem in Language Models' (Table 2) and Section 2 'Performance Benchmarks' (Table 1)]
"Analysis across model configurations reveals that disabling the thinking mechanism provides substantial speed benefits while maintaining competitive accuracy. Through comparative evaluation of same model sizes with and without thinking capabilities, performance differences remain relatively modest, yet no-think variants deliver significantly faster response times. This speed advantage, combined with improved user experience from reduced latency, motivated our adoption of force non-thinking training in subsequent versions."

The 10% SimpleQA subset in Table 2 is used to compare Think versus No-Think modes, and Section 5.1 states that this comparison 'motivated our adoption of force non-thinking training in subsequent versions.' The headline result, 83.2% on SimpleQA (Abstract and Table 1), is then reported on the same benchmark family whose slice was used for that design choice. The report never states whether 83.2% was computed on the full SimpleQA set or on the same 10% subset; only Appendix Table 4 labels a different run (80.7%, LangGraph MCP) as 'SimpleQA (Full).' The configuration is therefore selected using a slice of the target test set, and the headline is presented as an independent benchmark result.

full rationale

The central training claim is not circular in the strongest sense: the model is trained on MuSiQue, not SimpleQA, so the headline accuracy is not derived by construction from the training data, and there are no load-bearing self-citations or imported uniqueness theorems. The RLVR pipeline and the 80.7% and 83.2% numbers are empirical measurements against an external benchmark. The one significant circularity-adjacent step is the use of a 10% SimpleQA subset to choose force non-thinking training, followed by reporting SimpleQA as the headline. Because the paper does not state that 83.2% was measured on the full benchmark, and because the only explicitly 'Full' number in Appendix A.2 is 80.7%, the 83.2% result cannot be verified as independent of the mode-selection slice. This is best characterized as moderate evaluation circularity rather than a derivation that reduces by construction: no equation in the report forces 83.2% from the subset scores, and the model weights are public, so the number could in principle be reproduced independently. However, as written, the central comparison relies on a benchmark that was partially used to set the configuration, which warrants a score of 4. The separate assertion that RLVR 'completely eliminates' SFT is an overstatement given that the model is fine-tuned from Qwen3-4B, which is itself an SFT-trained model; that is a correctness concern rather than a circularity concern.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; RLVR is a training framework rather than a postulated entity. The main unevidenced commitments are domain assumptions about transfer from MuSiQue to SimpleQA and about the equivalence of the simulated and real search environments.

free parameters (5)

Training subset composition = 7,000 2-hop, 2,150 3-hop, 1,175 4-hop questions (10,325 total)
Hand-selected imbalance with no sensitivity analysis showing that this composition does not drive the results.
Retrieval pipeline parameters = top-15 candidates reranked to top-10, 150-character previews
Chosen for training efficiency; the effect of these choices on downstream SimpleQA performance is not reported.
Stage 1 reward weights = not specified
The reward balances correct answer, tool execution, format adherence, and XML compliance, but no weights, curves, or optimization details are given.
Force non-thinking mode = enabled
Selected after comparing Think versus No-Think modes on a 10% SimpleQA subset, so the deployment choice is tied to the evaluation benchmark.
Headline evaluation configuration = 128K context with minimal MCP server
The 83.2% headline comes from the 128K variant while the standard variant is 80.7%; the authors also report a forked MCP version exceeding 85% but chose not to use it.

assumptions (5)

domain assumption Qwen3-4B is a suitable base model for RLVR tool-use training.
Stated in Section 1 as the base model; no comparison to other 4B bases is provided.
domain assumption Training on MuSiQue transfers to SimpleQA performance.
Training uses MuSiQue-Ans, evaluation uses SimpleQA, and the only evidence of transfer is the final reported scores.
domain assumption DAPO with verifiable rewards can train tool-use behavior without SFT.
Section 4 relies on DAPO, but no ablation compares the no-SFT recipe against SFT or against DAPO initialized differently.
domain assumption The local RAG server approximates real search engine behavior.
Section 4.1 assumes the simulated snippets and scrape results behave like the Serper-based MCP tools used at evaluation time; the mismatch is not quantified.
domain assumption SimpleQA scores across models in Table 1 are comparable despite different MCP implementations.
The paper itself notes that different MCP implementations and evaluation setups may affect direct comparisons.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Jan-nano Technical Report." pith.science (2026). https://pith.science/paper/634KLNZQ

@misc{pith2026250622760,
  author       = {Pith},
  title        = {Pith review of: Jan-nano Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/634KLNZQ}},
  note         = {Machine review of arXiv:2506.22760}
}

read the original abstract

Most language models face a fundamental tradeoff where powerful capabilities require substantial computational resources. We shatter this constraint with Jan-nano, a 4B parameter language model that redefines efficiency through radical specialization: instead of trying to know everything, it masters the art of finding anything instantly. Fine-tuned from Qwen3-4B using our novel multi-stage Reinforcement Learning with Verifiable Rewards (RLVR) system that completely eliminates reliance on next token prediction training (SFT), Jan-nano achieves 83.2% on SimpleQA benchmark with MCP integration while running on consumer hardware. With 128K context length, Jan-nano proves that intelligence isn't about scale, it's about strategy.

Figures

Figures reproduced from arXiv: 2506.22760 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 9 canonical work pages

[1]

Dutch businessman convicted drug trafficking 9/11

Use web search to find information... ASSIST ANT: Thought:I will start by searching for the Dutch businessman... Action: result = web search(query="Dutch businessman convicted drug trafficking 9/11", filter year=2001) ASSIST ANT: Observation:Search Results:

work page 2001
[3]

URL https: //arxiv.org/abs/2503.09516. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨ uttler, Mike Lewis, Wen tau Yih, Tim Rockt¨ aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks,

arXiv
[7]

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W

URL https://arxiv.org/abs/2505.09388. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering,

arXiv
[9]

6 Tuesday 23rd September, 2025 Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu

URL https: //arxiv.org/abs/2503.14476. 6 Tuesday 23rd September, 2025 Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deep- researcher: Scaling deep research via reinforcement learning in real-world environments,

arXiv 2025
[10]

URL https: //arxiv.org/abs/2504.03160. 7 Tuesday 23rd September, 2025 A Conversation Logs Demonstrating Overthinking Behaviors A.1 Comparative Analysis of Y ear Filter Constraints Leading to Hallucination We present two representative examples demonstrating how the 8B model’s use of additional parameters system- atically leads to search failures and hallu...

arXiv 2025
[11]

Rudi Dekkers (July 27, 1956 – April 11,

[Rudi Dekkers - Wikipedia]... Rudi Dekkers (July 27, 1956 – April 11,

work page 1956
[13]

Dutch nationals 9/11 hijackers drug trafficking

[The threat of Russian Organized Crime]... ASSIST ANT: Thought:The initial search did not yield specific results. I need to adjust the query... Action: result = web search(query="Dutch nationals 9/11 hijackers drug trafficking", filter year=2001) ASSIST ANT: Observation:Search Results: [A list of irrelevant search results from 2001 about terrorism, the Pa...

work page 2001
[15]

ThunderCats 1985 production manager

Action: production info = web search(query="ThunderCats 1985 production manager") ASSIST ANT: Observation:Search Results:

work page 1985

Show all 19 references

[16]

For the shows ThunderCats (1985 original series)

[Masaki Iizuka - ThunderCats wiki - Fandom] Masaki Iizuka is known producer and production manager. For the shows ThunderCats (1985 original series)

1985
[17]

ASSIST ANT:The production manager of ThunderCats (1985) is Masaki Iizuka, as confirmed by multiple sources including Fandom, IMDb, and archived articles

[Thundercats (TV Series 1985–1989) - Full cast & crew - IMDb]... ASSIST ANT:The production manager of ThunderCats (1985) is Masaki Iizuka, as confirmed by multiple sources including Fandom, IMDb, and archived articles. ASSIST ANT: Final Answer:Masaki Iizuka ASSIST ANT:Here are...

1985
[18]

production manager of ThunderCats 1985

ASSIST ANT: Thought:I will start by searching for the production manager of ThunderCats using a web search filtered to 1985 to ensure relevance. Action: search results = web search(query="production manager of ThunderCats 1985", filter year=1985) ASSIST ANT: Observation:Search...

1985
[19]

ThunderCats 1985 production team Hanna-Barbera

[Television/RadloAge] ... Starting September 1985, a brand new 90-minute concept! THE PA W-PA WS... presented by HANNA-BARBERA’S. [Other irrelevant results from 1985 newspapers and magazines.] ASSIST ANT: Thought:The initial search did not yield direct results... I will refine...

1985
[1985]

Use web search(...) ASSIST ANT: Thought:I will use the web search tool to find information about the production manager of ThunderCats from

Facts to look up: Production manager of ThunderCats (1985)... Use web search(...) ASSIST ANT: Thought:I will use the web search tool to find information about the production manager of ThunderCats from

1985
[2018]

org/abs/1809.09600

URL https://arxiv. org/abs/1809.09600. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu...

arXiv
[2020]

doi: 10.18653/v1/2020.coling-main.580

International Committee on Computational Linguistics. doi: 10.18653/v1/2020.coling-main.580. URL https://aclanthology.org/2020.coling-main.580/. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to ...

2020 doi
[2021]

Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

URL https://arxiv.org/abs/2005.11401. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Musique: Multihop questions via single-hop question composition,

2005 arXiv
[2022]

Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus

URL https://arxiv.org/abs/2108.00573. Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models,

arXiv
[2024]

URL https://arxiv.org/ abs/2411.04368. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianho...

arXiv
[2025]

Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa

URL https://arxiv.org/abs/2503.19470. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. In Donia Scott, Nuria Bel, and Chengqing Zong, editors, Proceedings of the 28th Internati...

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Dutch businessman convicted drug trafficking 9/11

Use web search to find information... ASSIST ANT: Thought:I will start by searching for the Dutch businessman... Action: result = web search(query="Dutch businessman convicted drug trafficking 9/11", filter year=2001) ASSIST ANT: Observation:Search Results:

work page 2001

[2] [3]

URL https: //arxiv.org/abs/2503.09516. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨ uttler, Mike Lewis, Wen tau Yih, Tim Rockt¨ aschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks,

arXiv

[3] [7]

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W

URL https://arxiv.org/abs/2505.09388. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering,

arXiv

[4] [9]

6 Tuesday 23rd September, 2025 Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu

URL https: //arxiv.org/abs/2503.14476. 6 Tuesday 23rd September, 2025 Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deep- researcher: Scaling deep research via reinforcement learning in real-world environments,

arXiv 2025

[5] [10]

URL https: //arxiv.org/abs/2504.03160. 7 Tuesday 23rd September, 2025 A Conversation Logs Demonstrating Overthinking Behaviors A.1 Comparative Analysis of Y ear Filter Constraints Leading to Hallucination We present two representative examples demonstrating how the 8B model’s use of additional parameters system- atically leads to search failures and hallu...

arXiv 2025

[6] [11]

Rudi Dekkers (July 27, 1956 – April 11,

[Rudi Dekkers - Wikipedia]... Rudi Dekkers (July 27, 1956 – April 11,

work page 1956

[7] [13]

Dutch nationals 9/11 hijackers drug trafficking

[The threat of Russian Organized Crime]... ASSIST ANT: Thought:The initial search did not yield specific results. I need to adjust the query... Action: result = web search(query="Dutch nationals 9/11 hijackers drug trafficking", filter year=2001) ASSIST ANT: Observation:Search Results: [A list of irrelevant search results from 2001 about terrorism, the Pa...

work page 2001

[8] [15]

ThunderCats 1985 production manager

Action: production info = web search(query="ThunderCats 1985 production manager") ASSIST ANT: Observation:Search Results:

work page 1985