No — you don't need to take every set to literal momentary failure, the point where the bar stops moving no matter how hard you push. But the set still has to end within 1-2 reps of failure, and that's the part people skip: two near-failure sets beat four junk sets. Refalo and colleagues' 2023 meta-analysis found training to failure beat stopping short by an effect size of 0.19 (CI 0.00 to 0.37, p = 0.045) — trivial — and when the analysis isolated momentary failure specifically, the difference fell to 0.12 (CI −0.13 to 0.37, p = 0.343), which is nothing. The harder problem isn't whether to go to failure. It's that you almost certainly can't tell how far from failure you actually are. The honest limit, up front: that first confidence interval's lower bound sits at exactly 0.00. This is "failure isn't required," not "failure does nothing."
What does the research actually say about training to failure?
It's close to a wash. Two separate meta-analyses land in the same place — training to failure isn't meaningfully better than stopping a rep or two short.
Refalo 2023, in Sports Medicine, pooled the proximity-to-failure literature and found sets taken to failure produced a standardized effect of 0.19 on hypertrophy versus sets stopped short (CI 0.00 to 0.37, p = 0.045). By any conventional reading that's trivial, and the lower bound of the interval is sitting on zero — significant by a hair, small in practice.
The subgroup is the more interesting number. When the authors restricted the comparison to studies that used momentary muscular failure — the true "the rep does not move" definition, rather than a coach calling the set — the effect dropped to 0.12 with a confidence interval of −0.13 to 0.37 (p = 0.343). Not significant. The strictest version of failure has the weakest case behind it.
Grgic and colleagues' 2022 review in the Journal of Sport and Health Science ran the question across 15 studies and reported no significant difference between failure and non-failure training for either strength or hypertrophy.
So: two reviews, a whisper of an effect in one and nothing in the other. The argument you see online — failure zealots against RIR purists — is being fought over a gap the data can barely detect. The decision that actually matters is somewhere else.
How close to failure do you need to get?
Close. The last rep isn't where the growth is, but proximity is — the effect lives in the neighborhood of failure, not at the doorstep.
Robinson and colleagues' 2024 meta-regressions modeled reps-in-reserve as a continuous variable instead of a yes/no, and the two outcomes split apart. For hypertrophy, training closer to failure was associated with more growth, with confidence intervals that excluded the null — a real, directional relationship. For strength, the relationship was negligible and those intervals contained the null. The authors themselves caution that the analysis is exploratory, so take the direction, not a decimal.
Read that next to the failure numbers and the two stop contradicting each other. Failure itself buys you 0.19. Proximity buys you something real. In practice that means the last rep is optional and the last three are not — a set abandoned at 4 or 5 RIR is where the stimulus leaks out, not a set stopped at 1.
Which is why the baseline I run and teach hasn't moved: 3 sets · 8-12 reps · 1-2 RIR. One or two reps in reserve is close enough to collect essentially all of the growth, and far enough back that the set doesn't eat the next one.
Can you actually tell when you're two reps from failure?
No — and that's the real problem, not the failure debate.
Halperin and colleagues' 2022 review in Sports Medicine gathered 12 studies and 414 participants on how accurately lifters predict their reps-to-failure. The pooled result: people underpredicted by an average of 0.95 reps (CI 0.17 to 1.73). They believed they were done roughly one rep before they actually were. And the finding that should end the "I've lifted for years, I know my body" defense — training status did not improve accuracy. Experience didn't fix it.
Translate that into your session. Your "2 RIR" is, on average, 3 RIR. Your 3 is a 4. The set you'd describe as hard is one rep further from the edge than your own report of it, and the bias points the same direction for a beginner and a ten-year lifter.
This is what reconciles two pieces of advice that sound like they cancel out. "Most people aren't training hard enough" and "leave 1-2 reps in reserve" are the same instruction once you account for the measurement error. The prescription is 1-2 RIR; what you'll actually execute is 2-3. Aiming at 1-2 is how you land near 2-3. Aim at 3-4 and you'll be lifting at 5, which is where the stimulus genuinely starts falling off.
Two things follow from that. First, when you're deciding whether to push a set, the default answer is one more rep than you want, because your gauge reads low — that's not a motivational line, it's a correction for a documented bias. Second, verify occasionally. Take an isolation set — a machine curl, a leg extension — all the way to the point where a rep genuinely fails, and count how many reps past your predicted stopping point you got. That number is your personal calibration error, and it tells you whether you run a one-rep bias or a three-rep one. The review didn't test whether self-checking improves accuracy over time, so treat it as a sanity check rather than a fix.
What does training to failure cost you?
Your next two sets, not your next two days. The cost is real and measurable, and far shorter-lived than the "CNS fatigue" talk implies.
Refalo and colleagues' 2023 paper in Sports Medicine – Open tracked exactly this. Four minutes after a set, mean velocity was down 25% when sets went to failure, 13% at 1 RIR, and 8% at 3 RIR. Across the session it compounds: set-to-set velocity declined 22% in the failure condition versus 9% at 1 RIR and 6% at 3 RIR, and reps fell from the first set to the last by 7.58 at failure, 5.13 at 1 RIR, and 2.79 at 3 RIR. Take the first set to failure and you are, measurably, doing less work for the rest of that exercise. Which raises the obvious follow-up that most lifters never ask: if you're still 25% down four minutes later, how long should you actually be resting between sets?
Now the counterweight, because this is where failure gets oversold as dangerous: by 48 hours every condition had recovered. A day out, the failure and 1-RIR sets were still about 3% down on velocity — a real residue, but a rounding error next to the 25% hole four minutes after the set. No fried nervous system. Failure is expensive inside the session and largely paid off by the next one.
Soreness isn't the readout for any of this either — you don't have to be sore to build muscle, and being wrecked the next day tells you about novelty, not effort.
Which sets should you take to failure?
The last one, on the exercises where failing is cheap. That's the whole rule.
It follows straight from the fatigue numbers: failure costs you the sets that come after it, so put it where nothing comes after. In practice that means the final set of an exercise, on isolation and machine work — curls, lateral raises, leg extensions, cable rows, anything where the muscle gives out and the machine catches the weight.
The heavy compounds are where you don't. A squat or bench set taken to genuine failure buys you a 0.12 effect size and costs you technique breakdown under a loaded bar, a spotter's judgment call, and a measurable hole in every set that follows. Bad trade. Those sets end at 1-2 RIR, and that isn't a compromise — it's what the data supports anyway.
Rep range interacts with this. Machine and isolation work naturally lives at the higher end, which is exactly where a failed rep is safest, while your heaviest lift belongs lower, where control matters most. That's the case for heavier 5-8 rep sets versus 12-15 and which exercise earns which.
So how should you structure your sets?
2-3 hard sets per exercise, 8-12 reps, 1-2 RIR, and the number in the log goes up every session. That's the spec, and it's what I run four days a week on a Push/Pull/Legs/Upper split around a full-time job as a software engineer.
Two sets, both genuinely near failure, gets you most of the result — two sets are enough to build muscle as long as neither one is a warm-up in disguise. Add the third when the week gives you room. Intensity over volume: two hard sets beat four lazy ones, and most guys' problem isn't that they need more sets, it's that they aren't training hard enough on the ones they already do.
The logbook is the audit, and it's the only honest instrument you've got. Your RIR estimate is biased low, your soreness is noise, but reps at a given weight is a number that can't flatter you. If the log has been flat for a month at the same load, the missing variable is effort, not sets — adding a fourth set to sets you're bailing out of early just buys you more junk volume. The split, the set-and-rep spec, and the progression rule as one system is The Momentum Method.