NTC thermistor temperature reading jumping ±4°C — ADC noise or something in the circuit?

±4°C is not normal for an NTC thermistor. Let's work through why you're seeing it.

At 25°C with a 10k/10k divider, the midpoint voltage is around 1.65 V and the divider sensitivity is roughly 13–15 mV/°C. For ±4°C of reading error you'd need ±55–60 mV of noise on that ADC input. On a 3.3 V / 12-bit ADC, 1 LSB = 0.8 mV — so 60 mV is about 75 LSBs of noise. That's a lot. You have multiple error sources stacking up.

1. AVDD decoupling on the ESP32

This is usually the biggest culprit, especially if you're running Wi-Fi. The ESP32 ADC and its internal reference share a supply pin (often labelled VDDA or VDD3P3_RTC depending on the module). Every time the Wi-Fi PA fires — even if you're not in an active TX burst — there are transient load spikes on the 3.3 V rail that couple into the ADC reference. If you don't have a 100 nF ceramic capacitor placed right at that VDDA pin (not 10 mm away, right at the pin), those spikes read as noise.

Add a 100 nF C0G or X7R ceramic directly at the ADC supply pin. Add a 10 µF bulk cap nearby on the same rail. Then check your 3.3 V supply decoupling more broadly — if your regulator output cap is far from the ESP32 or is a large electrolytic without a parallel ceramic, you'll see this problem regardless of Wi-Fi state.

2. No filter capacitor across the thermistor

Right now your thermistor divider is presenting a Thevenin source impedance of roughly 5 kΩ (the parallel combination of 10 kΩ and 10 kΩ) at the ADC input. That 5 kΩ source impedance, combined with the ADC's sample-and-hold capacitor (~1 pF for the ESP32), forms a charging time constant that can cause problems at higher sampling rates, but more importantly, 5 kΩ × 80 mm of trace acts as an antenna and a noise integrator.

Add a 100 nF ceramic capacitor across the thermistor (between the divider output and GND). This does two things: it forms a low-pass RC filter with the Thevenin source resistance (f_c = 1 / (2π × 5 kΩ × 100 nF) ≈ 320 Hz), and it reduces the impedance that the 80 mm trace is driving at high frequency. For temperature sensing — a signal that changes over seconds or minutes — a 320 Hz cutoff is no constraint at all.

3. The 80 mm analog trace

An 80 mm unshielded trace on a 2-layer board with no ground pour is an efficient pickup antenna for anything switching near it. At 12-bit resolution and 5 kΩ source impedance, that trace will couple mains-frequency interference, digital edge transients, and Wi-Fi harmonics into the ADC reading.

The cap in point 2 helps a lot here. Additionally: if you can add a ground pour on the top layer adjacent to the analog trace, do it. Or use a bottom-layer GND pour directly below it. Either creates a low-impedance return path that reduces the effective loop area for noise pickup.

4. ESP32 ADC calibration and oversampling

The ESP32 ADC has a known non-linearity (see the ESP-IDF ADC calibration documentation). For 1°C accuracy you should be using the calibrated ADC API — esp_adc_cal_get_voltage() rather than raw adc1_get_raw(). The calibrated path compensates for the internal reference offset and gain error, which can be several mV by themselves.

For oversampling: the ESP-IDF ADC multisampling API or a simple firmware loop averaging 64 or 256 readings (with conversion after averaging, not per-sample) is more effective than averaging 100 raw integers. More samples, combined with the 100 nF cap already doing pre-filtering, will get you to ±0.2–0.3°C real noise in a normal environment.

In short: add 100 nF to VDDA, add 100 nF across the thermistor, route carefully, and use the calibrated ADC API. The sensor signal conditioning page covers the full picture of ADC input filtering and reference decoupling. The thermistor page has the divider circuit and self-heating considerations if you haven't read it already.