"The LLM Killer?" The paper times itself against diffusion models, not autoregressive ones.

A YouTube video from the channel Code Bear, titled "NCM: The LLM Killer?", summarizes two research papers that arXiv received four days apart in October 2026: Next-Concept Modeling (NCM), submitted October 9, and Hybrid Reasoning Agents (HRA), submitted October 13. We checked the video's main figures against both papers, and the table below lists them. We read the video's auto-generated captions, not the picture, so anything shown on screen is outside this note, and we did not rerun either model.

Those figures hold up. The title does not follow from them. The video frames both papers as an attempt to break AI out of "producing one token at a time," the autoregressive habit. NCM's wall-clock speedups, though, are all against diffusion language models, another family of parallel text generators, and not against autoregressive ones. The video says as much in its own body: "All of NCM's speed tests are against diffusion models of similar size," and the largest model is "far smaller than the chatbots you use." The paper itself also counts function evaluations against small autoregressive baselines, which we take up below. A title is a claim too.

#What the video gets right

Each row is a figure the video states, where we found it, and whether it matches.

#Where the title outruns the first paper

The paper's own claim is narrower than the title. Its abstract frames the work around diffusion language models, which "offer parallel sampling for faster generation than autoregressive models," and states each speed result against the strongest of those: 21 times faster than DUO on LM1B, and 31 times faster than AURORA-LM on XSum, a summarization set. Figure 1, the wall-clock plot, shows only diffusion and flow baselines, as do the wall-clock panels of Figures 5 and 6. We found no autoregressive wall-clock time anywhere in the paper.

Autoregressive models do appear, as quality baselines and in function-evaluation counts. The tables put a 99M parameter autoregressive baseline at 64 function evaluations against NCM's 11 on translation and 8 on summarization (Table 2), and the Sudoku autoregressive rows at 128 against NCM's 12 (Table 3). A function evaluation is not a unit of time, and the paper itself cautions that "NFE alone does not capture the full inference efficiency picture, since the cost per step differs across different diffusion language models." We take that caution to apply to any comparison of call counts across model families. That is our reading.

On quality the picture is mixed. NCM is far ahead on hard Sudoku and ahead on XSum ROUGE-L (28.0 against 24.4 for the autoregressive baseline), and behind on translation without guidance (22.6 against 25.2; with minimum Bayes risk decoding NCM ties at 25.2, using 132 function evaluations against 64). In Table 8, the autoregressive sampling rows score lower, which is better, on generative perplexity as scored by GPT-2 large: 8.89 against NCM's 15.41 on TinyStories and 35.45 against 83.12 on OpenWebText, at 128 and 1,024 function evaluations against NCM's 12 and 32. The abstract's speed and perplexity claims are worded around the diffusion rows. On summarization, "matches" means a ROUGE-L of 28.0 for NCM against 26.6 for AURORA-LM without guidance, and 28.2 against 28.6 with guidance (Table 2). On OpenWebText the abstract says "competitive," not better.

Scale and scope narrow it further. The authors write that "We train a separate tokenizer and generator for each dataset" and that "Unless otherwise stated, our generators have approximately 100M parameters, comparable to or smaller than the baselines." The 725M model is a scaling study on one dataset, LM1B, which the conclusion calls "preliminary scaling results." Target lengths are fixed per dataset, from 48 tokens for XSum to 1,024 for OpenWebText (Table 4), and the limitations appendix says "NCM fixes both the hierarchy depth K and the per-level lengths" and lists "Scaling to larger datasets and model sizes" as "another exciting direction." We searched the paper's full text and found no instruction-following, chat or agent task. No chat or frontier model is compared in any table: named language models only score the generated text (GPT-2 large for perplexity in the main results; Mistral-7B, Llama-3.1-8B, Gemma-2-9B and Qwen-3-14B in an appendix check), supply a tokenizer, or, in one ablation, serve as a frozen prefix encoder (T5-small, not used in the reported results).

The 31x itself is the paper's stated figure for generating 1,024 sequences at batch size 16 on a single H100 GPU against AURORA-LM. We did not rerun it.

#The second paper gets the same check

The video's GAIA numbers are in the paper's Table 1. The framing around them is where we slowed down.

The comparison model. The video calls GPT-5.6 Luna "a frontier model," and the paper lists it under "Frontier LLM Agents." OpenAI's own page says Luna "is designed for cost-sensitive, high-volume workloads" and "roughly corresponds to the nano model tier used in earlier GPT-5 families." OpenAI's page for GPT-6 Astra calls that model "our most capable model for the most demanding work," and no Astra row appears in Table 1. The paper does not mention Astra, so we cannot say whether it was available when the experiments ran. Its data pipeline does name a "GPT-6.1-Sol" among the external models that assign each reasoning step to text or latent mode, and that model is not a Table 1 row. The 14.5 point gap (52.4 minus 37.9) is against OpenAI's cost-sensitive tier.

The matched agent. The authors trained a text-based agent with the ARPO method on "the same RL training data and backbone models as HybridAgent," and it scores 43.7 on GAIA. That is 5.8 points above Luna before any silent reasoning, and the hybrid agent's margin over it is 8.7 points. The strongest frontier GAIA row, Kimi-K3, scores 41.7.

The spread. The paper prints no GAIA question counts. Seventeen of the table's 20 GAIA rows fit scores out of 39, 52 and 12 questions (103 in all) to within 0.1 point, including the four rows used here; two other splits we tried, 40, 50 and 10 and 30, 30 and 30, fit none. Taking that as an assumption, which is ours, and using unpaired Newcombe score intervals, the hybrid's margin over the text-based ARPO agent has a rough 95 percent interval from 4.8 points below zero to 21.9 above. Its margin over Luna runs from 1.0 to 27.4 above zero. Both agents were scored on the same benchmark, so a paired analysis would likely be tighter. The paper's own "±0.017" is a 95 percent interval over 5 runs of the same agent, which on our reading measures run-to-run variation, not the luck of which questions were asked. The paper does not print per-question results, so we could not pair them and these intervals are rough.

The caption. On Humanity's Last Exam the 14B agent's average is 16.0, against 28.4 to 31.4 for three frontier agents. The Table 1 caption nonetheless says "At 8B and 14B scales, HybridAgent consistently outperforms advanced LLM agents, agentic frameworks, and matched text and latent baselines." The HLE columns do not support that sentence, and neither do the DeepSearchQA columns, where Luna scores 23.3 and 49.4 against the 14B agent's 21.7 and 46.8. The video was right to say the agent loses on HLE.

The headline number. The abstract says "Compared with frontier LLM Agents and agentic frameworks spanning both text-based and latent methods, HybridAgent consistently improves performance by 10.4%," and puts the efficiency figures (35.5 to 54.7 percent fewer tokens, 1.8 to 2.6 times faster, 2.7 to 3.7 times cheaper) in the same sentence. Section 5.1 gives the 10.4% again, this time against "corresponding text/latent baselines" at the RL stage, and adds 17.4% for the 14B agent over "the strongest agentic baseline." The body measures the token and step reductions against "matched text-based LLM agents" (Figure 6 caption) and the cost against "text agents under the same backbones" (Table 9 caption). The video's "text-only agents they trained the same way" is the body's comparison, so on this point the video is more careful than the abstract.

On "learned on its own." The video says the model was never rewarded for going quiet and "learned on its own" when to do it. That follows the authors' framing: the paper says "We impose no explicit reward for mode usage," the abstract says HybridRL learns "when to think in latent space and when to act through text," and the lead author's thread says "The agent learns itself." The video does mention that 54,000 training examples were built automatically with no human labeling, but not what the automatic pipeline decided. The paper says "an automatic annotation pipeline" assigns each reasoning step "to either text or latent mode," and the exit from silent mode gets "dedicated supervision." On our reading, the split between text and silence was partly taught before RL began.

#When reasoning goes quiet, other fields keep a record

The second paper's central choice, moving some reasoning into hidden-state steps that never appear as words, raises an older question than language models: what record of the reasoning survives? Two regulated fields have answered it in writing.

The Public Company Accounting Oversight Board's standard on audit documentation, AS 1215, says the auditor "must document significant findings or issues, actions taken to address them (including additional evidence obtained), and the basis for the conclusions reached in connection with each engagement." The record must let "an experienced auditor, having no previous connection with the engagement" follow the work, and where judgment is heavier, as with accounting estimates, it expects "commensurately more extensive documentation." The standard asks for the basis of a conclusion, not only the action taken.

Cockpit voice recorders are the aviation version. They record "flight crew verbalizations of intentions and coordination," and the recording is a rolling one that "overwrites the oldest data." The FAA's final rule, published February 2, 2026, raises the required recording time "from the currently mandated 2 hours to 25 hours for all affected future manufactured aircraft," and sets its three compliance timeframes "in response to the Act and comments received," the Act being the 2024 FAA Reauthorization Act. The rule's background cites a 2018 recommendation from the National Transportation Safety Board (NTSB) for 25 hours. That recommendation stems from a July 2017 incident at San Francisco International Airport where recorder data was overwritten, leaving an investigation that "lacked direct evidence of the flight crew's decision making, coordination, and perception of its environment." The rule says the NTSB's recommendation report lists 14 events from 2003 to 2018 in which installed recorders were overwritten.

Our reading, not the papers' and not the regulators': both fields keep or require a record of reasoning where one exists, audit scales it with the judgment involved, and aviation lengthened it after recordings proved too short.

The HRA paper's main evidence about its own visible record is Table 7. On 560 math problems (the weak monitor was scored on MATH500 only), the hybrid agent's visible text is 43 to 49 percent shorter, carries 26 to 28 percent more key reasoning steps per 100 tokens, and lets an outside monitor predict whether an answer was correct about as well as a text agent's trace does: higher in three of four comparisons and 0.001 lower in the fourth. That is a result about correctness on math problems. We found no monitoring test in the paper on agent tasks, and none of whether a monitor catches deception or other misbehavior: a search of its full text finds no "deception," no "reward hacking" and no "sabotage," and "misbehavior" appears once, in the title of a cited paper.

By the audit standard's test, whether someone with no previous connection to the work could follow what a hybrid agent did, and why, from its visible record alone is still open. That is an inference, and a test on agent tasks showing the visible text supports the same audit as a text-only agent's would settle it.

That open question is the one behind the video's safety beat. Peter Barnett, whose profile says "AI governance at @MIRIBerkeley" and "Views my own," replied to project lead Jiaru Zou's announcement thread on October 14: "This undermines one of the only ways we have to (imperfectly) monitor AIs. This is very bad for safety." His reply goes on to say the authors should not continue the line of work.

The paper takes the opposite view in its Section 8: latent thoughts "are readable solely through the provider-owned HybridLink," so "no text exists to leak," and "providers can assess the safety of reasoning and actions over the full readout, including the portion users never see." The video reads out a compressed version of that passage, in quotation marks, as if it were the paper's own sentence.

Four of the paper's eight authors list Together AI, whose home page says of its platform: "Accelerate inference, model shaping and pre-training on a research-optimized platform." An argument built on the provider holding the readout is also an argument that suits providers. That is our inference about incentive, and nothing we read shows it shaped the paper. What the paper does test is whether the hidden steps matter. The swap results above show that they do, and that replacing each one with its nearest word's embedding barely changes the score on this one benchmark.

#Before you repeat a multiplier

1. Read the legend, not the title. For wall-clock time the comparison class is in the figure: here, diffusion and flow models, not autoregressive ones.
2. Find the tier. A row grouped under "Frontier LLM Agents" is a label. The vendor's own page says which tier it is. Here it was the cost-sensitive one, and a matched text agent already sat 5.8 points above it.
3. Ask for the spread. When the question count is small and per-question results are not printed, a margin can be hard to tell from zero. By our unpaired arithmetic, the hybrid agent's margin over its matched text agent could be zero, though a paired analysis would likely be tighter.
4. When a claim moves work out of view, ask what record remains. Audit and aviation answer that in writing.

Related field notes: A video explained a new reasoning paper accurately, then put a number in the title that the paper never reported and An 8B model beat 32B baselines by changing the harness.

Whether NCM holds up at the scale of the models people actually use is a fair question. The authors list scaling as future work, and a title should not answer it for them.

#Sources

All read October 19, 2026.

• Code Bear (YouTube), NCM: The LLM Killer?
• arXiv, Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (2610.14162) and its PDF
• arXiv, Hybrid Reasoning Agents (2610.18861) and its PDF
• OpenAI, GPT-5.6 Luna model page
• OpenAI, GPT-6 Astra model page
• X, Jiaru Zou's announcement of Hybrid Reasoning Agents, October 14, 2026
• X, Peter Barnett's reply, October 14, 2026 and his profile
• Together AI, home page
• PCAOB, AS 1215 Audit Documentation
• Federal Register, FAA 25-Hour Cockpit Voice Recorder (CVR) Requirement, New Aircraft Production, February 2, 2026