# E5-small LoRA vs TMR, which AI detector is more accurate?

Source: https://toolbenchmarks.com/compare/e5-lora-vs-tmr/  
Tested: 30 October 2026. Publisher: Tool Benchmarks. Data licence: CC BY 4.0.

> E5-small LoRA ranks higher, with a balanced accuracy of 55.6% against TMR's 50.0%, but neither is usable. TMR flagged 83.3% of human texts and E5 flagged 50.0%. With line breaks removed TMR flagged 23.8% of human texts, fewer than E5, but its catch rate fell from 83.3% to 38.7%, well below E5's unchanged 61.1%.

| Measure | E5-small LoRA detector | TMR (Target Mining RoBERTa) |
|---|---|---|
| Balanced accuracy | 55.6% | 50.0% |
| Catches AI text | 61.1% | 83.3% |
| Flags human text | 50.0% | 83.3% |
| Catches humanized AI text | 47.6% | 81.0% |
| AUROC | 0.582 | 0.605 |
| Flags human text: Academic abstracts | 36.0% | 16.0% |
| Flags human text: Q&A answers | 75.0% | 100.0% |
| Flags human text: Encyclopedia articles | 52.0% | 100.0% |
| Flags human text: Fiction | 12.5% | 100.0% |
| Flags human text: News reports | 70.8% | 100.0% |

## Two RAID detectors that both over-flag

Both models were trained on RAID, a public benchmark of human and AI writing. TMR's model card says its training method was designed to keep false positives low. On our test texts, human ones written before ChatGPT and AI ones from current Claude models, both flagged most of the human writing in most genres.

The genre pattern differs. TMR flagged every human forum answer, encyclopedia article, fiction passage and news report. E5-small LoRA flagged 75.0% of forum answers and 70.8% of news but only 3 of 24 fiction passages. Academic abstracts reverse the order: TMR flagged 16.0% of them and E5 36.0%, while both caught few AI abstracts, 16.0% and 21.0%.

## Line breaks explain TMR, not E5

TMR's problem has a known trigger. As written, it flagged all 101 human texts with more than one paragraph. With the line breaks replaced by spaces, it flagged 26 of them, and its rate on all human texts fell to 23.8%.

E5 flagged 54 of the same 101 texts, with or without line breaks, because its tokenizer ignores them. Whatever drives its false positives is in the words themselves, which makes it harder to work around than TMR's.

Catch rates follow the same pattern. Without breaks, TMR caught 38.7% of AI texts. E5 stayed at 61.1%. So TMR stripped of its quirk flags fewer humans than E5 and also misses far more AI text.

## Catch rates and famous texts

As published, TMR's numbers on AI text look strong: 88.1% of Claude Fable 5.1 texts against E5's 49.2%, and 81.0% of humanized Claude Opus 5.5 texts against 47.6%. Those come from the same habit of flagging nearly everything. TMR flagged 28 of 30 famous public-domain texts; E5 flagged 6.

Both are fast, at 0.03 seconds per 1,000 words for E5 and 0.09 for TMR.

## A better option than both

If you are stuck with TMR, removing line breaks before scoring cuts its false positive rate to 23.8%, but its catch rate falls to 38.7%. E5 has no such workaround. In the same test Desklib flagged 6.3% of human texts while catching 58.3% of AI text, and gave the same verdicts with or without line breaks. The [paragraph break guide](/guides/paragraph-breaks-change-ai-detector-results/) compares how all seven handled formatting; the [TMR](/ai-detectors/tmr/) and [E5-small LoRA](/ai-detectors/e5-lora/) pages cover the rest.

## Questions people ask

### Which flags more human writing, E5-small LoRA or TMR?

TMR, at 105 of 126 human texts (83.3%), against 63 of 126 (50.0%) for E5-small LoRA. Academic abstracts were the exception, where TMR flagged 16.0% and E5 36.0%.

### Why does TMR flag so much more than E5-small LoRA?

Mostly because of paragraph breaks. TMR flagged all 101 multi-paragraph human texts as written, and 26 of them once the breaks were replaced with spaces. E5's verdicts did not depend on line breaks.