REF-1483 · Robinson et al., 2024 · Sports Medicine · Paper · PubMed

How close to failure should I lift?

For strength, estimated reps in reserve made no clear difference; for size, closer gained more. Exploratory: reps in reserve were estimated, not measured, and fit was modest. Not endurance athletes.

Main caveat: The predictor is an estimate. The authors (preprint): "the accuracy of these estimations is unknown".
moderate evidenceMeasured on non-athletes to trained athletes

Robinson and colleagues gathered trials from earlier meta-analyses of failure, velocity loss and set structures, and estimated for each group the reps in reserve at which sets ended: from training to failure, reported velocity, reported RIR, or load and reps through prediction equations (method details from the preprint).

They then regressed strength and muscle size on estimated RIR, adjusted for load, method of volume equating, duration and training status.

The abstract gives the finding: in all best-fit models for strength, the confidence intervals of the marginal slopes contained a null point estimate; in all best-fit models for hypertrophy, the slopes were negative with intervals that did not contain the null, so size rose as sets ended closer to failure. Fit was "modest".

The authors conclude that the optimal proximity may differ between strength and hypertrophy, that the exact relationship remains unclear, and that caution is warranted given the exploratory nature. In the preprint, 243 strength effects came from 55 studies and 140 hypertrophy effects from 26.

There too, fixed effects explained 12.29% (strength, SMD) and 19.2% (hypertrophy, SMD) of the variance, and the prediction intervals span zero widely. Our reading: the direction for strength supports stopping short for endurance athletes. The slope for size is small and rests on a modelled predictor. It does not reach to a "how many reps in reserve" number.

Protocol

Design (preprint): "performed without a systematic search and was not preregistered"; studies from existing relevant meta-analyses plus others known to the authors. Eligible: English; no known conditions; set- and/or rep-volume-equated; load-equated within ±5% of 1RM; at least two proximities to failure; maximal strength and/or direct hypertrophy.
Who (preprint): n = 55 studies, participants not totalled; interventions 8.28 ± 2.35 weeks; age 27.83 ± 12.84 years; typical training 9.58 ± 4.48 sets a week at 72.06 ± 13.27% of 1RM, 2.23 ± 0.48 sessions a week (strength models).
How RIR was estimated (preprint): failure groups coded 0 (or −1 with a clear failure definition, in the spline); subtraction of reps from a failure group at the same load; velocity-loss equations; reported RIR taken verbatim; load and reps through published prediction equations; each cluster or rest-redistribution bout counted as a set. Range extracted 0 to 23 RIR. Special rules applied when estimates looked implausible.
Analysis (preprint): multilevel mixed-effects models with cluster-robust variance; effects as standardised mean change and response ratio; six functional forms compared by approximate Bayes factors; 95% confidence and prediction intervals; no null-hypothesis testing. Sensitivity: estimated RIR under 10.
Funding and interests (preprint): Robinson, Pelland, Remmert and Zourdos "are all coaches and writers in the fitness industry"; Refalo, Jukic and Steele declare no relevant conflicts. Funding is not stated in the abstract record.

Key Results

From the preprint (PREPRINT numbers, not checked against the journal version). β is the marginal slope per estimated RIR.

Outcome and metricBest fitEffects, studiesMarginal / conditional R²β (95% CI; 95% PI)
Strength, SMDlinear-log243, 5512.29% / 78.66%0.003 (−0.012, 0.018; −0.675, 0.682)
Strength, response ratiolinear243, 5529.04% / 59.41%−0.059 (−0.304, 0.186; −12.944, 14.732)
Hypertrophy, SMDlinear140, 2619.2% / 72.09%−0.019 (−0.035, −0.004; −0.551, 0.513)
Hypertrophy, response ratiolinear140, 2626.38% / 63.24%−0.48 (−0.78, −0.179; −9.811, 9.817)
Sensitivity, RIR under 10 (preprint): strength SMD 0.004 (−0.014, 0.022), 232 effects from 54 studies; hypertrophy SMD −0.023 (−0.042, −0.004); hypertrophy response ratio −0.544 (−0.929, −0.158). Same conclusions.
Moderator with few data (preprint): hypertrophy slopes were positive in programmes that did not use progressive overload (n=5 effects) and negative in those that did (n=135).
From the published abstract: best-fit models "demonstrated modest quality of overall fit"; strength intervals contain a null point estimate; hypertrophy intervals do not; "caution is warranted".

Mechanisms

Measured: strength and size change against an estimated predictor; no direct RIR for most groups, no mechanism data.
Authors' interpretation (preprint): load predicts strength better than proximity to failure; force falls as a set approaches failure but rises with load. Possibly opposed advantages (fatigue management far from failure; specificity and more size close to failure) cancel out for strength. For size, the size principle fits, and heavier loads may not need as low a RIR.
Inference: the preprint's response-ratio slope of −0.48 reads as about half a percentage point of muscle size per RIR across a study of about 8 weeks (both preprint figures, not in the published abstract), with the prediction interval well over zero on both sides. That is small enough for fatigue, or the rest of the week, to matter more for a person with a race calendar.

Practical Application

How close to failure should I lift? For strength, estimated reps in reserve made no clear difference; for size, closer gained more. Exploratory: reps in reserve were estimated, not measured, and fit was modest. Not endurance athletes.
Strength goal: nothing here argues against stopping a few reps short. In line with SYNTH-43, SYNTH-44 and REF-1476.
Size goal: if muscle size is really the aim, closer to failure may give a little more. Not from this paper; the base's own practice for endurance athletes: stay short of failure as routine, since the gym is a supplement and SYNTH-31 sets longer gaps after sessions taken to failure.
No RIR number from this paper: the authors (preprint) point only to "potential RIR thresholds of interest for future research". REF-1476 also says targets can't be quantified.
The log: the app logs RPE, not RIR, and neither is measured here. A reps-in-reserve line in the guide is a label for a practice, not a finding of this paper.
Not quite in line with the base: REF-678 (A-) reports momentary failure is not superior for size (ES 0.12, p = 0.343) and a trivial advantage for set failure under any definition (ES 0.19, p = 0.045), and this paper reports size rising toward failure. These answer different questions: failure vs not, against a whole-range slope that reaches up to 23 estimated RIR (range from the preprint). Neither licenses grinding.

Limitations

The predictor is an estimate. The authors (preprint): "the accuracy of these estimations is unknown". Group-average RIR hides individuals, later sets, and strength gained during the study; errors could run either way.
Exploratory, by its own label. No systematic search and no preregistration (preprint), so study selection is open to bias.
Modest fit, wide intervals. The published abstract says "modest"; in the preprint, fixed effects explained 12.29% to 29.04% of the variance in the primary models, 95% prediction intervals span both sides of zero, and many dose-response shapes remain compatible.
Volume-equated only. Training farther from failure might allow more sets, which these models can't see (authors' point, preprint).
Numbers. The Key Results are from a preprint. The published article may differ; it could not be read here.
Moderators (preprint). Thin subgroups, with n=5 in the no-progressive-overload group, must be read cautiously.
Interests. Four of seven authors are coaches and writers in the fitness industry, as printed in the preprint.
How this entry was checked and changed
Why this level

a meta-regression of volume-equated comparisons would sit at A-; held at B+ because the predictor itself is an estimate, the fit is modest, the authors call it exploratory, and the preprint says it was done without a systematic search or preregistration. The level grades the published paper; its detailed numbers are the preprint's and were not checked against the journal version

Full text (checked 2026-10-30)
This exploratory meta-analysis was performed without a systematic search and was not preregistered (preprint; full text)
Interventions lasted 8.28 ± 2.35 weeks; participants 27.83 ± 12.84 years old; 55 studies (preprint; full text)
Strength: 243 total effects from 55 studies; hypertrophy: 140 total effects from 26 studies (preprint; full text)
Strength SMD: marginal 12.29%, conditional 78.66%; β= 0.003 [95% CI: -0.012, 0.018; 95% PI: -0.675, 0.682] (preprint; full text)
Hypertrophy SMD: marginal 19.2%, conditional 72.09%; β= -0.019 [95% CI: -0.035, -0.004; 95% PI: -0.551, 0.513] (preprint; full text)
Hypertrophy response ratio: β= -0.48 [95% CI: -0.78, -0.179; 95% PI: -9.811, 9.817] (preprint; full text)
Sensitivity: strength 0.004 [95% CI: -0.014, 0.022]; hypertrophy -0.023 [95% CI: -0.042, -0.004] (preprint; full text)
Programs without progressive overload (n=5) a positive slope; with it (n=135) a negative slope (preprint; full text)
The accuracy of these estimations is unknown; the quality of model fit is modest (preprint; full text)