qwen3.6:35b always fails on this, unless very high resolution

Posted by qfghclvx@reddit | LocalLLaMA | View on Reddit | 22 comments

This is an exercise every child (I guess) can solve correctly. qwen says solution B is right, or D.

What would you say? Try it.

This was the thinking process around point Q. I can’t follow how it is so wrong:

But much much later (many thousand thinking tokens later):

And again later:

I don’t understand how it can misinterpret the slope so wrongly. And then correct again.

Gemma4:26b got this right most of the times, but sometimes says solution A is correct. Gemini 3.1 flash lite is always wrong and says solution A. But Gemini 3.1 pro preview is always correct.

And very interestingly: Opus 4.7 and Opus 4.6 always say solution A (mostly) or D is correct. Oh my god.

Although this looks like an easy exercise, this seems to be very difficult visual input. A good benchmark.

All other “difficult” visual physics exercises were solved correctly by qwen3.6:35b, where even Opus 4.7 failed and gemma failed at 26b but got it right at 31b. Do you want to see them?

The worst thing of gemma:26b was, that it produced so many hallucinated words in longer solutions and therefore made also markdown/latex errors. gemma:31b didn’t have that problem. And qwen3.6 never has.

[-]

Euphoric_Emotion5397@reddit

I'm using qwen 3.6 q4km kv cahce q8.
I tried with a larger copy of the image and it worked. The original was a webp format, but I just snip to save as a png.

Thought for 47.26 seconds

Let me analyze this step by step using first principles of wave mechanics.

Causal Analysis

Step 1: Establish the Wave-Particle Relationship

For a transverse wave, there's a fundamental relationship between the wave shape at a snapshot in time and the particle velocity:

Where:

v_particle = vertical velocity of the rope particle (positive = upward)
v_wave = speed of wave propagation (positive = left to right, negative = right to left)
slope = dy/dx of the wave profile at that point

Step 2: Analyze Point P (The Given Clue)

From the diagram:

P is on the equilibrium line (dashed line)
At P, the wave has a negative slope — it's coming from a crest and heading toward a trough as you move rightward along the rope

Given: P's subsequent motion is upward → v_particle(P) > 0

Applying the formula:

v_particle(P) = −v_wave × (negative slope)
Upward = −v_wave × (−|slope|)
Upward = v_wave × |slope|

For this to yield a positive (upward) result, v_wave must be positive → the wave travels left to right.

Step 3: Analyze Point Q

From the diagram:

Q is below equilibrium, on the downward-sloping portion of the wave (after a crest, before reaching the next trough)
The slope at Q is negative (wave going from above toward below as x increases)

Applying the formula with v_wave > 0 (established above):

v_particle(Q) = −(positive) × (negative slope)
v_particle(Q) = positive → **Upward**

Verification via Wave-Shift Method

If wave travels left to right: In a small time Δt, the entire wave pattern shifts slightly to the right. At Q's position, the portion of the wave immediately to its LEFT (which is closer to equilibrium from above) moves into Q's location. This means Q's displacement becomes less negative → Q moves upward. ✓

If wave traveled right to left: The portion to the RIGHT of Q would shift into Q's position. That portion is further below equilibrium, so Q would move downward — which contradicts what we'd expect if P moved upward (verified above). ✗

Answer: A

Direction of motion of Q	Direction of travel of the wave
Upwards	Left to right

Why Other Options Fail:

B & D (Downwards for Q): Would require the slope at Q to be positive, but the diagram shows Q on a downward-sloping portion.
C & D (Right to left for wave): Contradicts the given that P moves upward — if the wave traveled right to left, P would move downward.

[-]

qfghclvx@reddit (OP)

that is what the title of this thread says: higher resolution works.

[-]

Euphoric_Emotion5397@reddit

LOL. So it's a human error. My apologies.

FinalCap2680@reddit

For such comparisons to be useful and make sense. at least say do you use full model or what quant.

Other settings as tomp, top_p, and so on will be useful to know too...

it was the 4 quant. and the defaults of llama.cpp which are temp=0.8 etc (don’t know right now).

Thank you

Direction of motion of Q	Direction of travel of the wave
Upwards	Left to right

Ok-Internal9317@reddit

I think it’s Right to left and Q going upwards? C. Correct me if I’m wrong.

correct. good luck with your exams in a few days. :-D

RandomTrollface@reddit

I think you are right. If the wave was moving left to right, P would move downward, so the wave must move right to left given that P is moving upward.

the only reason i know this is that I'm an IB physics hl student and this question just so happens to be an IB physics paper 1 question which i grinded A LOT over the past two years, nice knowing I got it right cuz my finals is in like 10 days lol

I tried this question with multiple LLMs and they all fumble, the main reason for this from what I can think of is that it's pretty hard to reason this with words. Like this phrase i used to justify my answer: "because we know that P is moving upward, we know wave energy travel is *right to left*. And so Q must go *up*" you see that the answer are like obvious (if you know) and unlogical (if you dont know) and that its pretty hard to justify things that comes straight out of actual human reasoning.

since the LLM gets trained on predicting the correct next word, it probably just got trained on similar phrases and sine curves images (maybe a flipped one) and one that used to say *down*, which kinda proofs the point that they infact cannot actually reason like humans but just predict the next word, this to me is kinda sad.

Lorian0x7@reddit

I was thinking the same, I was about to comment then I realized. You would be right if the 2 points where fixed and disconnected from the wave. But actually, the 2 points are on the wave(the rope).

I must say, the problem is not very clear, it doesn't say if the rope and the wave are the same thing.

because we know that P is moving upward, we know wave energy travel is right to left. And so Q must go up, I am talking when the points are fixed on the rope.

What do you reckon the answer is?

If wave and rope are the same thing: A wave moves in time and space so if P goes upwards and it's attached to the wave then the wave is moving from left to right. Answer A

If the point is fixed on the X axes not attached to the wave then the wave goes from right to left. Answer C

yes www.youtube.com/watch?v=WGu3JoyBVjw

moahmo88@reddit

unsloth/Qwen3.6-35B-A3B-GGUF UD-Q5_K_M**Conclusion:**

* **Direction of motion of Q:** Upwards

* **Direction of travel of the wave:** Left to right

This corresponds to **Option A**.

Acemang_Jedi@reddit

desculpa incomodar. que hardware tens para correr essa versao? estou a tentar fazer uma rig para correr localmente uma IA, nao sei se compensa a nivel de consumo de eletricidade. tenho 2 3080Ti, 2 RTX 6800 2 3070 ti 2 RTX 6700. todas paradas.

Thank you for posting it!

Strange. I run it on unsloths UD-Q8_K-XL and got:

Conclusion:
Direction of motion of Q: Downwards
Direction of travel of the wave: Right to left

ilintar@reddit

Tried on Bartowski's Q5_K_S. Its answer:

Conclusion:

Direction of motion of Q: Upwards
Direction of travel of the wave: Left to right

This corresponds to option A.

Correct Answer: A

It arrived at the solution very quickly too, only took 15k tokens :D

DinoAmino@reddit

15K tokens - so how long did that take?

FWIW: Gemma 4 31B FP8 nails it with 448 output tokens in 5 secs. Temp 0.3, top_k 10, top_p 0.75. No system prompt.

Not long, the generation is 110 t/s. But still kinda funny 😃

soyalemujica@reddit

Did it solve it after a second try though ?

"only" XD