zai-org

/

GLM-5

like 2.09k

Follow

Z.ai 13.5k

Hi Z.AI Team, thank you so much for open-sourcing such impressive models and sharing your research!

I have 2 clarifying questions about the context management strategies described in Section 4.2.4 of the tech report, in order to help ensure community reproducibility of the BrowseComp/HLE results.

Question 1) What exactly is discarded in Discard-all?

In Section 4.2.4, you state:

“we discard the entire tool-call history and restart with a fresh context”

Could you clarify precisely what is removed when Discard-all is triggered?

Given a trajectory (as described in the paper):

(q, r_1, a_1, o_1, r_2, a_2, o_2, ..., r_n, a_n, o_n)

Where:

• q = question
• r_i = reasoning at round i
• a_i = action
• o_i = observation

Does Discard-all reset the context to one of the following?

Scenario A — Keep only system prompt + initial question

(system_prompt, q)

Meaning everything else is discarded:
r_i, a_i, o_i for all i.

Scenario B — Remove only observations

(system_prompt, q, r_1, a_1, r_2, a_2, ..., r_n, a_n)

Meaning:
o_i discarded for all i.

Scenario C — Remove all tool-related content (actions + observations)

(system_prompt, q, r_1, r_2, ..., r_n)

Meaning:
a_i and o_i discarded for all i.

Or is there another variant intended?

Question 2) Interaction between Discard-all and keep-recent-k in HCM

In Section 4.2.4, "Hierarchical Context Management" is described as:

“During inference with keep-recent, if the total context length exceeds a threshold T, we discard the entire tool-call history and restart with a fresh context, while continuing to apply the keep-recent strategy.”

Would you be able to provide clarification on the control flow logic? I see 3 plausible interpretations:

Interpretation 1 — Discard-all triggered once, keep-recent always active

• keep-recent-k is active from the start.
• If len(context) > T:
  • Trigger Discard-all once.
• After restart:
  • keep-recent-k continues to apply.
  • Discard-all is never triggered again.

Pseudocode:

discarded_once = False

while not done:
    if len(context) > T and not discarded_once:
        context = discard_all(context)
        discarded_once = True

    context = apply_keep_recent_k(context, k)

Interpretation 2 — Discard-all triggered every time T is exceeded

• keep-recent-k always active.
• Whenever len(context) > T:
  • Trigger Discard-all.
• This can happen multiple times during inference.

Pseudocode:

while not done:
    if len(context) > T:
        context = discard_all(context)

    context = apply_keep_recent_k(context, k)

Interpretation 3 — keep-recent-k activated only after first Discard-all

• Initially, no keep-recent-k.
• If len(context) > T:
  • Trigger Discard-all.
  • Activate keep-recent-k.
• After that:
  • keep-recent-k remains active.
  • Discard-all is never triggered again.

Pseudocode:

keep_recent_active = False
discarded_once = False

while not done:
    if len(context) > T and not discarded_once:
        context = discard_all(context)
        discarded_once = True
        keep_recent_active = True

    if keep_recent_active:
        context = apply_keep_recent_k(context, k)

Or is there another interpretation intended?

Thanks again for any clarifications you are able to provide, and for your amazing work! 🙏