Do you need to train to failure?

Momentary muscular failure is the strict version: you cannot complete the lifting part of another rep despite maximal effort. Studies also use volitional failure (you decided to stop), velocity-loss thresholds, or self-reported reps in reserve, and many never defined it. Pelland and colleagues' methods review documents this, and it is the main reason the headline findings look muddier than the headlines.

Reps in reserve (RIR) is the other currency: 2 RIR means you think you had two more reps. The RPE scale most lifting apps use is anchored to it, with RPE 10 as 0 RIR and RPE 9 as 1 RIR. Zourdos and colleagues introduced the scale and found bar velocity tracked reported RPE closely (r = −0.88 in experienced squatters). That supports the scale's logic. It does not make any single rating accurate. Our guide to RPE and RIR covers how to rate a set.

"Stop when your form breaks down" is a coaching convention, not a research definition.

Almost every disagreement turns on whether the volumes were equal. Grgic's team traced it. Across studies that did not equate volume, non-failure looked better for strength (ES −0.32, 95% CI −0.57 to −0.07). In most of those studies the non-failure groups simply did more sets. Equate volume and the effect is gone (ES 0.01, 95% CI −0.12 to 0.15). Vieira AF's meta-analysis shows the same for growth: failure looked better for hypertrophy overall (SMD 0.75), with no difference once volumes were equal.

Proximity is not simply neutral, and it does not act the same on both goals. Robinson's meta-regressions found that in every best-fit hypertrophy model, muscle size rose as sets ended closer to failure, with confidence intervals that excluded null. In every best-fit strength model the intervals contained null. The authors call the fit modest and the analysis exploratory. Pareja-Blanco's squat trial shows the split inside one study: stopping each set at 20% velocity loss matched 40% velocity loss for 1RM gains with 40% fewer reps and a bigger jump gain, while the closer-to-failure group grew more quadriceps. Velocity loss is a proxy for proximity, and that trial did not match volume, so read it as a pattern, not a proof.

Returns also flatten near the top. Refalo and colleagues' 2023 meta-analysis found that going past 25% velocity loss, against 20 to 25%, was worth ES 0.08 (95% CI −0.16 to 0.32).

Before you believe any comparison — Ask whether the two groups did the same amount of work. Most headline findings on both sides of this debate disappear when they did.

This is the live dispute, and it is unresolved.

For: Grgic's resistance-trained subgroup showed a small hypertrophy advantage for failure, ES 0.15 (95% CI 0.03 to 0.26), built on only two studies. Refalo's 2023 meta-analysis (Sports Medicine 53:649) found a pooled set-failure effect of ES 0.19 (95% CI 0.00 to 0.37, p = 0.045). That interval touches zero, and by the authors' own sensitivity analysis it loses significance if the assumed within-subject correlation falls below 0.73 (they used 0.75). For momentary muscular failure specifically it is ES 0.12 (−0.13 to 0.37), and the review concludes there is no evidence that momentary failure is superior for hypertrophy. Hermann and colleagues (2025) trained 42 resistance-trained men and women for eight weeks on one set of nine exercises, failure against 2 RIR. Several hypertrophy measures tended to favour failure, with modest absolute differences, and strength and endurance were similar.

Against: Refalo's 2024 within-person trial in trained lifters, which is a different paper from the 2023 meta-analysis, compared each person's failure leg with their own 1 to 2 RIR leg over eight weeks. Quadriceps thickness rose 0.181 cm and 0.182 cm. Ruple's group found no condition-by-time interaction for 1RM or vastus lateralis size at 0 to 1 RIR against 4 to 6 RIR.

All of these groups are small, from 18 to 42 people over five to eight weeks. That cannot settle an effect this size either way, and "no detectable difference" is not "proven identical".

A meta-analysis of acute fatigue (Vieira JG and colleagues, a different paper from the Vieira AF review above) found sets taken to failure left more disruption than stopping short. Neuromuscular function fell more (SMD −0.96, 95% CI −1.43 to −0.49), and metabolic response, muscle damage and perceived exertion rose. Its abstract flags the exertion gap in studies that did not equalise volume, so the effort cost of the extra reps is easiest to see when failure also means more work.

Refalo's crossover trial (Sports Medicine – Open, 2023) puts a dose on it. After six sets of bench press at 75% 1RM, lifting velocity four minutes later was down 25% in the failure condition, 13% at 1 RIR and 8% at 3 RIR. The loss was mostly short-lived. At 24 hours failure and 1 RIR were both down 3% and 3 RIR was up 2%, and the differences were gone by 48 hours. That was one exercise and six sets, not a full session of heavy compounds.

Over eight weeks, failure also felt worse in a separate Refalo paper (2025): discomfort 5.1 against 4.1, session RPE 5.4 against 4.3, and less positive general feelings. The authors read that as a possible adherence cost. Adherence itself was not tracked long enough to show that worse feelings make people quit.

Load is the clearest moderator anyone has shown. In Lasevicius' trial, with volume equated, failure added nothing at 80% 1RM (quadriceps CSA +8.1% to failure, +7.7% stopping short). At 30% 1RM, failure was the difference between growth and none (+7.8% against +2.8%, and the light non-failure leg did not grow significantly). When the load is light, effort is what makes the set count. Those men were untrained, so applying it to trained lifters is an extrapolation.

From there, the familiar advice is a reasonable inference, not a finding. We label it as convention: about 1 to 3 RIR on heavy compounds, and 0 to 1 RIR on isolation lifts and light work. It follows from the load result, the fatigue data, and the fact that a heavy multi-joint set is harder to judge. The direct evidence is thin and mixed. Davies' compound-exercise subgroup favoured non-failure for strength (ES 0.37 to 0.38, p = 0.03), though the authors put the difference at 0.6 to 1.3% and doubted it mattered. Grgic's team ran a similar subgroup and found no exercise-selection moderation. No meta-analysis has tested exercise type for hypertrophy.

The related rule, taking only the last set of each exercise to failure, is practitioner convention. In Robinson's 2025 four-arm trial (38 men, eight weeks), bench 1RM rose 9.05 kg at 4 to 6 RIR and 9.72 kg at 1 to 3 RIR. It rose 5.07 kg where only the last set went to failure and 0.71 kg where every set did. The failure squat arms were too hard to sustain and their data are hard to interpret, and the arms are small, so nobody should lean on this. It is the nearest test, not a verdict.

There is a gap between the RIR you report and the RIR you have. Halperin's review of 262 effect sizes from 12 studies found people underpredict their remaining reps by about one on average (0.95, 95% CI 0.17 to 1.73). They had more left than they thought, so a reported 2 RIR is often nearer 3. Between-person variation was small (SD 1.45 reps), but the studies disagreed widely (I² = 97.9%), so treat it as a direction, not a figure. Accuracy improved slightly when the judgement was made closer to failure and on sets of about 12 reps or fewer. Hermann's group found it improved over eight weeks of practice and was better on the bench press than the squat.

Training experience does not reliably help. Halperin found no meaningful moderation by training status (β = −0.006, 95% CI −0.02 to 0.007), and Droguett found none in the back squat, with 16 people. Droguett's authors suggest familiarity with high effort may matter beyond training age. Zourdos found experienced squatters rated a true 1RM higher than novices did (p = 0.023), the one point where the groups differed. Inferring effort from load percentages is worse: Qin found a percentage-based table overestimated intended RIR by about 2.9 reps at 60% 1RM (a generalised velocity table was off by 5.8), while errors at 80% were within a rep.

• "You need failure to recruit high-threshold motor units, so you need it to grow." That is a mechanism, not an outcome. Ruple's group measured motor-unit firing-rate changes that differed between 0 to 1 RIR and 4 to 6 RIR, with no difference in strength or muscle size.
• "Everything before the last few reps is junk volume." Volume-equated comparisons contradict it, and so does Pareja-Blanco, where stopping at 20% velocity loss matched 40% for squat strength with 40% fewer reps.
• "Training to failure is dangerous." We found no evidence either way. Long-term injury outcomes of habitual failure training have not been studied, and adverse-event reporting in this literature is limited. That is an absence of evidence, not a clean bill of health.

When you log an RPE, Shojin's next-load suggestion turns on one number: the highest RPE you logged among the sets at your top weight last session. Shojin checks for a slip first. If that RPE was 9 or 10 and you did fewer reps than the session before at the same weight, it suggests about 5% less. Otherwise RPE 10 holds the weight, and RPE 9 or lower adds one step (2.5 kg on a barbell with standard plates). If you leave RPE blank, your reps decide, and the default is to hold. Load goes up when your reps stay the same at the same weight two sessions running, or when you reach the top of the rep range.

A worked example: you bench 80 kg for 8, 8 and 7 reps at RPE 8, 9 and 10. Shojin reads the top weight's highest RPE, which is 10, and holds 80 kg. Log that last set as a 9 and it offers 82.5 kg. The rating does real work, so rate each set as best you can. Under-report and the app pushes load onto you. Over-report and it freezes you at a weight you could have beaten. The design reads the one set where effort is easiest to judge, but it is still your estimate, and the literature says that estimate is off by about a rep.

It cannot tell you how many reps to leave in reserve, because no trial has established a number. Grgic's team say the literature cannot show whether stopping five reps short differs from stopping two, and Refalo's 2023 review says the proximity that would maximise hypertrophy is unclear. Every specific target in circulation, including the ones we find reasonable, is convention. It cannot tell you whether failure gives a trained lifter a small hypertrophy edge, because the studies on both sides run 18 to 42 people over five to eight weeks. Studies that size cannot detect effects this small, so a null result is weak evidence of no difference and a barely significant one is fragile. The pooled finding most often quoted as proof that closer to failure grows more muscle stops being significant if one analytic assumption moves by 0.02. It cannot tell you about long-term injury risk, because that has not been studied. This is training practice, not medical advice. If you are returning from injury, managing a health condition, pregnant, or new to lifting, ask a clinician or a qualified coach.

Shojin is one subscription with no cap on sets. Shojin Coach is part of it: it writes your program from your own lifts and adjusts it after every session you log.

New subscribers get one month free, then A$4.99 a month or A$34.99 a year, billed by the App Store in your local currency. The subscription renews until you cancel. Cancel any time in Settings. You subscribe in the iPhone app. If you stop, your history stays, and Coach stops adjusting your program.

Every number in this article traces to one of these. Where the evidence is contested or thin, the article says so rather than picking a side.