Stopping sets short of failure gave slightly more 1RM-type strength than taking them to failure (SMD 0.24, 95% CI 0.06 to 0.42, I² 0%) and no clear difference in isometric strength, muscle size, muscular endurance or power, in a meta-analysis of 20 randomised trials (556 healthy adults, mean ages 19.0-67.6; GRADE moderate for every outcome); between studies (subgroups not tested against each other), the strength edge was significant in resistance-trained people (0.38) and not significant in untrained ones (0.09), and some included trials changed set number or volume together with failure. Read in full
Main caveat: The authors' own: mostly small trials; varied protocols (frequency, duration, load).
Wu and colleagues pooled randomised trials in healthy adults that compared resistance training stopped before failure with training taken to failure, and asked about dynamic strength, isometric strength, muscle size, muscular endurance and power. They searched four databases to December 2025 and included 20 trials with 556 participants (mean ages 19.0-67.6 years).
Non-failure training raised dynamic strength (mostly 1RM) a little more than failure training (SMD 0.24, 95% CI 0.06 to 0.42), with no heterogeneity and no sign of publication bias; for isometric strength (−0.12), hypertrophy (−0.15), muscular endurance (−0.02) and power (0.02) the differences were trivial and not significant, and the confidence intervals include zero.
In subgroups the strength advantage was significant in resistance-trained participants (0.38) but not in untrained ones (0.09), significant in programmes longer than 6 weeks, and significant in the velocity-based training subgroup (0.47); in that subgroup hypertrophy favoured failure (−0.79, 39 participants). GRADE certainty was moderate for every outcome.
The authors conclude that failure is not needed to maximise adaptations and may cost a little dynamic strength, and they call the subgroup results hypothesis-generating. Our reading: the direction is consistent with earlier work in the base, but the size of the strength edge rests partly on trials in which the non-failure group also did more sets.
Participants
20 RCTs, 556 healthy adults; mean ages 19.0-67.6 years (one trial in older adults, about 64-67 years).
Training status by the review's McKay classification: 38 sedentary, 442 recreationally active, 21 highly trained (national level), 55 international-level competitors; by lifting experience 203 trained, 353 untrained.
Sex: missing in one study; the overall split is not reported. Table 1 includes all-female samples (for example young women in Martorelli 2017, women in Sanborn 2000).
Group sizes in Table 1 range from about 6 to 32 per arm.
Protocol
Search: four databases (MEDLINE, PubMed, SPORTDiscus, Web of Science), searched up to 14 December 2025, plus 4 records found by hand; 3157 records, 40 full texts, 20 included. English only.
Inclusion: healthy adults (18 or older); RCT; non-failure training against failure training; at least one of strength, hypertrophy, muscular endurance or power.
Non-failure definitions: the same load with fewer reps (18 studies by the text), a reduced load with the same protocol (1), or voluntary interruption before failure (2). Three trials ended sets by velocity loss (10% or 20% in Table 1).
Programmes: non-failure groups 3-40 sets of 1-12 reps, failure groups 1-4 sets of 2-12 reps; 4-14 weeks (median 8.7), 2-4 sessions a week; single-joint in 10 trials, multi-joint in 11; fixed, incremental or wave-like loading.
Analysis: post-test means and SDs; random-effects inverse-variance SMD (Hedges' g), one effect size per study per outcome by a stated hierarchy, with a sensitivity analysis using all effect sizes; I²; funnel plot and Egger's test; subgroups for training status, volume equated or not, traditional vs velocity-based, ≤6 vs >6 weeks; no meta-regression (too few studies, low heterogeneity).
Risk of bias: Cochrane Collaboration's tool (six domains), two reviewers; GRADE for each outcome.
Key Results
| Outcome (non-failure vs failure; positive favours non-failure) | SMD (95% CI) | p | I² |
|---|---|---|---|
| Dynamic strength | 0.24 (0.06 to 0.42) | 0.010 | 0% |
| Isometric strength | −0.12 (−0.48 to 0.24) | 0.510 | 0% |
| Muscle hypertrophy | −0.15 (−0.39 to 0.09) | 0.220 | 0% |
| Muscular endurance | −0.02 (−0.45 to 0.41) | 0.930 | 30.96% |
| Muscular power | 0.02 (−0.28 to 0.31) | 0.910 | 0% |
Subgroups (Table 2; participants per subgroup in brackets):
| Outcome and subgroup | SMD (95% CI) | p |
|---|---|---|
| Dynamic strength, trained (178) | 0.38 (0.09 to 0.66) | 0.01 |
| Dynamic strength, untrained (283) | 0.09 (−0.07 to 0.25 in Table 2; −0.08 to 0.41 in the text) | 0.19 |
| Dynamic strength, traditional training (385) | 0.20 (0 to 0.39) | 0.05 |
| Dynamic strength, velocity-based training (76) | 0.47 (0.03 to 0.91) | 0.04 |
| Dynamic strength, ≤6 weeks (83) | 0.20 (−0.22 to 0.62) | 0.35 |
| Dynamic strength, >6 weeks (378) | 0.25 (0.06 to 0.45) | 0.01 |
| Dynamic strength, volume equated (256) | 0.21 (−0.03 to 0.45) | 0.09 |
| Dynamic strength, volume not equated (205) | 0.29 (0.02 to 0.56) | 0.03 |
| Hypertrophy, velocity-based training (39) | −0.79 (−1.41 to −0.16) | 0.01 |
| Hypertrophy, traditional training (217) | −0.04 (−0.30 to 0.22) | 0.75 |
| Power, trained (137) | 0.03 (−0.29 to 0.36) | 0.85 |
Publication bias: Egger's p 0.632 (dynamic strength), 0.821, 0.884, 0.489 and 0.803 for the other outcomes; none detected.
Sensitivity analysis with all effect sizes: broadly consistent with the main results (numbers in a supplementary table, not read).
GRADE: moderate for each outcome (details in a supplementary table, not read).
Risk of bias, in words: allocation concealment was the most often unclear domain, sequence generation was often thinly described, blinding of participants was generally not possible and assessor blinding was inconsistently reported; no count of low-, moderate- or high-risk trials is given in the text.
No subgroup was tested against another; the review's "moderators" are subgroups with and without a significant pooled effect.
Practical Application
Taking sets to failure is not needed for strength, size, endurance or power; stopping short gave slightly more 1RM strength and no significant difference on the other outcomes. Across 20 randomised trials the non-failure groups gained a little more dynamic strength (SMD 0.24, a small effect) and showed no significant difference on the other four outcomes, at moderate certainty.This supports the base's "short of failure" default in SYNTH-105, SYNTH-101, SYNTH-43 and SYNTH-44.
The edge for non-failure was significant in the subgroup of studies on people who already lift; the trained and untrained subgroups were not tested against each other. In resistance-trained participants it was 0.38; in untrained participants it was 0.09 and not significant, so in untrained people no advantage either way was shown. Inference, not tested here: for a beginner the choice can rest on comfort and recovery.
For muscle size alone, stopping short did not cost growth on average (−0.15, CI −0.39 to 0.09), but the point estimate leans to failure. In the small velocity-based training subgroup (39 people), failure grew more muscle (−0.79). Inference, not tested here: a gym-first athlete whose goal is size can take some last sets close to failure; this fits the base's other hypertrophy evidence (REF-678, REF-1483).
The review does not tell how far from failure to stop. It pools very different "non-failure" protocols (fewer reps at the same load, voluntary stops, 10-20% velocity loss) and gives no reps-in-reserve dose. The 2-3 reps in reserve used in SYNTH-43 and SYNTH-44 remain practice. Inference, not tested here: for an endurance athlete who lifts, stopping short also leaves more for the next ride or run; the review did not measure this.
Where it differs from the guide: none in direction. SYNTH-105 states its starting sets "short of failure" without a trial behind that choice; this review is that evidence for strength, with the caveat that part of the strength edge comes from trials where the non-failure group also did more sets.
Limitations
The authors' own: mostly small trials; varied protocols (frequency, duration, load); participants from recreationally active to competitive athletes with few within-trial comparisons; no meta-regression.
Several included comparisons mix failure with other variables (Table 1): in Kramer 1997 and Sanborn 2000 three non-failure sets were compared with one set to failure; Dorrell 2020 compared velocity-based with percentage-based loading; Pareja-Blanco 2017 compared 20% with 40% velocity loss; the non-failure groups often did more sets (Drinkwater 2005: 8 × 3 against 4 × 6). The dynamic-strength effect was not significant when volume was equated (0.21, CI −0.03 to 0.45).
Text and tables disagree in places: the untrained subgroup's interval differs between Table 2 and the text; the text says 16 publications showed non-failure increasing dynamic strength more than failure, while Table 1 marks many of these trials as equal; the protocol counts add to 21 for 20 trials; one trial's reference (Lasevicius 2019) names a different paper (a frequency study).
Risk of bias is reported only as a narrative and a figure; the GRADE table is supplementary and was not read, so we cannot see why each outcome was rated moderate.
Subgroups split studies, not people; the authors call them observational and possibly underpowered.
Most trials lasted 4-14 weeks and used one or a few exercises; no trial in endurance athletes is identified, and no injury or adherence outcome was pooled.
Funding: the Chongqing Talents Program; the authors declare no competing interests.
Revision 2026-11-10 (headline): the headline now says that the trained and untrained results are subgroups of studies that were not tested against each other, and gives them as "significant" and "not significant" where it said "clear" and "not".
Revision 2026-11-11: the Practical Application line on training status no longer ranks the two subgroups ("clearest in people who already lift"); it says the edge was significant in one subgroup of studies and that the subgroups were not tested against each other, as the headline does.
How this entry was checked and changed
Why this level
a meta-analysis of RCTs whose risk of bias and GRADE rating (moderate for every outcome) we read in the full text, held below A because the trials are small (556 people over 20 trials, about 28 per trial by our arithmetic), the risk-of-bias result is given only in words (allocation concealment mostly unclear, no counts of studies per rating), and several included comparisons are not clean failure-versus-non-failure contrasts; no marker step, since strength, muscle size, endurance and power are outcomes